• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2016 Bas Nieuwenhuizen
3  *
4  * SPDX-License-Identifier: MIT
5  */
6 
7 #include "ac_nir_to_llvm.h"
8 #include "ac_gpu_info.h"
9 #include "ac_binary.h"
10 #include "ac_llvm_build.h"
11 #include "ac_llvm_util.h"
12 #include "ac_shader_abi.h"
13 #include "ac_shader_util.h"
14 #include "ac_nir.h"
15 #include "nir/nir.h"
16 #include "nir/nir_deref.h"
17 #include "sid.h"
18 #include "util/bitscan.h"
19 #include "util/u_math.h"
20 #include <llvm/Config/llvm-config.h>
21 
22 struct ac_nir_context {
23    struct ac_llvm_context ac;
24    struct ac_shader_abi *abi;
25    const struct ac_shader_args *args;
26 
27    gl_shader_stage stage;
28    shader_info *info;
29 
30    LLVMValueRef *ssa_defs;
31 
32    struct ac_llvm_pointer scratch;
33    struct ac_llvm_pointer constant_data;
34 
35    struct hash_table *defs;
36    struct hash_table *phis;
37    struct hash_table *verified_interp;
38 
39    LLVMValueRef main_function;
40    LLVMBasicBlockRef continue_block;
41    LLVMBasicBlockRef break_block;
42 };
43 
get_def_type(struct ac_nir_context * ctx,const nir_def * def)44 static LLVMTypeRef get_def_type(struct ac_nir_context *ctx, const nir_def *def)
45 {
46    LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, def->bit_size);
47    if (def->num_components > 1) {
48       type = LLVMVectorType(type, def->num_components);
49    }
50    return type;
51 }
52 
get_src(struct ac_nir_context * nir,nir_src src)53 static LLVMValueRef get_src(struct ac_nir_context *nir, nir_src src)
54 {
55    return nir->ssa_defs[src.ssa->index];
56 }
57 
get_memory_ptr(struct ac_nir_context * ctx,nir_src src,unsigned c_off)58 static LLVMValueRef get_memory_ptr(struct ac_nir_context *ctx, nir_src src, unsigned c_off)
59 {
60    LLVMValueRef ptr = get_src(ctx, src);
61    ptr = LLVMBuildAdd(ctx->ac.builder, ptr, LLVMConstInt(ctx->ac.i32, c_off, 0), "");
62    /* LDS is used here as a i8 pointer. */
63    return LLVMBuildGEP2(ctx->ac.builder, ctx->ac.i8, ctx->ac.lds.value, &ptr, 1, "");
64 }
65 
get_block(struct ac_nir_context * nir,const struct nir_block * b)66 static LLVMBasicBlockRef get_block(struct ac_nir_context *nir, const struct nir_block *b)
67 {
68    struct hash_entry *entry = _mesa_hash_table_search(nir->defs, b);
69    return (LLVMBasicBlockRef)entry->data;
70 }
71 
get_alu_src(struct ac_nir_context * ctx,nir_alu_src src,unsigned num_components)72 static LLVMValueRef get_alu_src(struct ac_nir_context *ctx, nir_alu_src src,
73                                 unsigned num_components)
74 {
75    LLVMValueRef value = get_src(ctx, src.src);
76    bool need_swizzle = false;
77 
78    assert(value);
79    unsigned src_components = ac_get_llvm_num_components(value);
80    for (unsigned i = 0; i < num_components; ++i) {
81       assert(src.swizzle[i] < src_components);
82       if (src.swizzle[i] != i)
83          need_swizzle = true;
84    }
85 
86    if (need_swizzle || num_components != src_components) {
87       LLVMValueRef masks[] = {LLVMConstInt(ctx->ac.i32, src.swizzle[0], false),
88                               LLVMConstInt(ctx->ac.i32, src.swizzle[1], false),
89                               LLVMConstInt(ctx->ac.i32, src.swizzle[2], false),
90                               LLVMConstInt(ctx->ac.i32, src.swizzle[3], false)};
91 
92       if (src_components > 1 && num_components == 1) {
93          value = LLVMBuildExtractElement(ctx->ac.builder, value, masks[0], "");
94       } else if (src_components == 1 && num_components > 1) {
95          LLVMValueRef values[] = {value, value, value, value};
96          value = ac_build_gather_values(&ctx->ac, values, num_components);
97       } else {
98          LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
99          value = LLVMBuildShuffleVector(ctx->ac.builder, value, value, swizzle, "");
100       }
101    }
102    return value;
103 }
104 
emit_int_cmp(struct ac_llvm_context * ctx,LLVMIntPredicate pred,LLVMValueRef src0,LLVMValueRef src1)105 static LLVMValueRef emit_int_cmp(struct ac_llvm_context *ctx, LLVMIntPredicate pred,
106                                  LLVMValueRef src0, LLVMValueRef src1)
107 {
108    src0 = ac_to_integer(ctx, src0);
109    src1 = ac_to_integer(ctx, src1);
110    return LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
111 }
112 
emit_float_cmp(struct ac_llvm_context * ctx,LLVMRealPredicate pred,LLVMValueRef src0,LLVMValueRef src1)113 static LLVMValueRef emit_float_cmp(struct ac_llvm_context *ctx, LLVMRealPredicate pred,
114                                    LLVMValueRef src0, LLVMValueRef src1)
115 {
116    src0 = ac_to_float(ctx, src0);
117    src1 = ac_to_float(ctx, src1);
118    return LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
119 }
120 
emit_intrin_1f_param(struct ac_llvm_context * ctx,const char * intrin,LLVMTypeRef result_type,LLVMValueRef src0)121 static LLVMValueRef emit_intrin_1f_param(struct ac_llvm_context *ctx, const char *intrin,
122                                          LLVMTypeRef result_type, LLVMValueRef src0)
123 {
124    char name[64], type[64];
125    LLVMValueRef params[] = {
126       ac_to_float(ctx, src0),
127    };
128 
129    ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
130    ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
131    assert(length < sizeof(name));
132    return ac_build_intrinsic(ctx, name, result_type, params, 1, 0);
133 }
134 
emit_intrin_1f_param_scalar(struct ac_llvm_context * ctx,const char * intrin,LLVMTypeRef result_type,LLVMValueRef src0)135 static LLVMValueRef emit_intrin_1f_param_scalar(struct ac_llvm_context *ctx, const char *intrin,
136                                                 LLVMTypeRef result_type, LLVMValueRef src0)
137 {
138    if (LLVMGetTypeKind(result_type) != LLVMVectorTypeKind)
139       return emit_intrin_1f_param(ctx, intrin, result_type, src0);
140 
141    LLVMTypeRef elem_type = LLVMGetElementType(result_type);
142    LLVMValueRef ret = LLVMGetUndef(result_type);
143 
144    /* Scalarize the intrinsic, because vectors are not supported. */
145    for (unsigned i = 0; i < LLVMGetVectorSize(result_type); i++) {
146       char name[64], type[64];
147       LLVMValueRef params[] = {
148          ac_to_float(ctx, ac_llvm_extract_elem(ctx, src0, i)),
149       };
150 
151       ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
152       ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
153       assert(length < sizeof(name));
154       ret = LLVMBuildInsertElement(
155          ctx->builder, ret,
156          ac_build_intrinsic(ctx, name, elem_type, params, 1, 0),
157          LLVMConstInt(ctx->i32, i, 0), "");
158    }
159    return ret;
160 }
161 
emit_intrin_2f_param(struct ac_llvm_context * ctx,const char * intrin,LLVMTypeRef result_type,LLVMValueRef src0,LLVMValueRef src1)162 static LLVMValueRef emit_intrin_2f_param(struct ac_llvm_context *ctx, const char *intrin,
163                                          LLVMTypeRef result_type, LLVMValueRef src0,
164                                          LLVMValueRef src1)
165 {
166    char name[64], type[64];
167    LLVMValueRef params[] = {
168       ac_to_float(ctx, src0),
169       ac_to_float(ctx, src1),
170    };
171 
172    ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
173    ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
174    assert(length < sizeof(name));
175    return ac_build_intrinsic(ctx, name, result_type, params, 2, 0);
176 }
177 
emit_intrin_3f_param(struct ac_llvm_context * ctx,const char * intrin,LLVMTypeRef result_type,LLVMValueRef src0,LLVMValueRef src1,LLVMValueRef src2)178 static LLVMValueRef emit_intrin_3f_param(struct ac_llvm_context *ctx, const char *intrin,
179                                          LLVMTypeRef result_type, LLVMValueRef src0,
180                                          LLVMValueRef src1, LLVMValueRef src2)
181 {
182    char name[64], type[64];
183    LLVMValueRef params[] = {
184       ac_to_float(ctx, src0),
185       ac_to_float(ctx, src1),
186       ac_to_float(ctx, src2),
187    };
188 
189    ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
190    ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
191    assert(length < sizeof(name));
192    return ac_build_intrinsic(ctx, name, result_type, params, 3, 0);
193 }
194 
emit_bcsel(struct ac_llvm_context * ctx,LLVMValueRef src0,LLVMValueRef src1,LLVMValueRef src2)195 static LLVMValueRef emit_bcsel(struct ac_llvm_context *ctx, LLVMValueRef src0, LLVMValueRef src1,
196                                LLVMValueRef src2)
197 {
198    LLVMTypeRef src1_type = LLVMTypeOf(src1);
199    LLVMTypeRef src2_type = LLVMTypeOf(src2);
200 
201    if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
202        LLVMGetTypeKind(src2_type) != LLVMPointerTypeKind) {
203       src2 = LLVMBuildIntToPtr(ctx->builder, src2, src1_type, "");
204    } else if (LLVMGetTypeKind(src2_type) == LLVMPointerTypeKind &&
205               LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
206       src1 = LLVMBuildIntToPtr(ctx->builder, src1, src2_type, "");
207    }
208 
209    return LLVMBuildSelect(ctx->builder, src0, ac_to_integer_or_pointer(ctx, src1),
210                           ac_to_integer_or_pointer(ctx, src2), "");
211 }
212 
emit_iabs(struct ac_llvm_context * ctx,LLVMValueRef src0)213 static LLVMValueRef emit_iabs(struct ac_llvm_context *ctx, LLVMValueRef src0)
214 {
215    return ac_build_imax(ctx, src0, LLVMBuildNeg(ctx->builder, src0, ""));
216 }
217 
emit_uint_carry(struct ac_llvm_context * ctx,const char * intrin,LLVMValueRef src0,LLVMValueRef src1)218 static LLVMValueRef emit_uint_carry(struct ac_llvm_context *ctx, const char *intrin,
219                                     LLVMValueRef src0, LLVMValueRef src1)
220 {
221    LLVMTypeRef ret_type;
222    LLVMTypeRef types[] = {ctx->i32, ctx->i1};
223    LLVMValueRef res;
224    LLVMValueRef params[] = {src0, src1};
225    ret_type = LLVMStructTypeInContext(ctx->context, types, 2, false);
226 
227    res = ac_build_intrinsic(ctx, intrin, ret_type, params, 2, 0);
228 
229    res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
230    res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
231    return res;
232 }
233 
emit_b2f(struct ac_llvm_context * ctx,LLVMValueRef src0,unsigned bitsize)234 static LLVMValueRef emit_b2f(struct ac_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize)
235 {
236    assert(ac_get_elem_bits(ctx, LLVMTypeOf(src0)) == 1);
237 
238    switch (bitsize) {
239    case 16:
240       if (LLVMGetTypeKind(LLVMTypeOf(src0)) == LLVMVectorTypeKind) {
241          assert(LLVMGetVectorSize(LLVMTypeOf(src0)) == 2);
242          LLVMValueRef f[] = {
243             LLVMBuildSelect(ctx->builder, ac_llvm_extract_elem(ctx, src0, 0),
244                             ctx->f16_1, ctx->f16_0, ""),
245             LLVMBuildSelect(ctx->builder, ac_llvm_extract_elem(ctx, src0, 1),
246                             ctx->f16_1, ctx->f16_0, ""),
247          };
248          return ac_build_gather_values(ctx, f, 2);
249       }
250       return LLVMBuildSelect(ctx->builder, src0, ctx->f16_1, ctx->f16_0, "");
251    case 32:
252       return LLVMBuildSelect(ctx->builder, src0, ctx->f32_1, ctx->f32_0, "");
253    case 64:
254       return LLVMBuildSelect(ctx->builder, src0, ctx->f64_1, ctx->f64_0, "");
255    default:
256       unreachable("Unsupported bit size.");
257    }
258 }
259 
emit_b2i(struct ac_llvm_context * ctx,LLVMValueRef src0,unsigned bitsize)260 static LLVMValueRef emit_b2i(struct ac_llvm_context *ctx, LLVMValueRef src0, unsigned bitsize)
261 {
262    switch (bitsize) {
263    case 8:
264       return LLVMBuildSelect(ctx->builder, src0, ctx->i8_1, ctx->i8_0, "");
265    case 16:
266       if (LLVMGetTypeKind(LLVMTypeOf(src0)) == LLVMVectorTypeKind) {
267          assert(LLVMGetVectorSize(LLVMTypeOf(src0)) == 2);
268          LLVMValueRef i[] = {
269             LLVMBuildSelect(ctx->builder, ac_llvm_extract_elem(ctx, src0, 0),
270                             ctx->i16_1, ctx->i16_0, ""),
271             LLVMBuildSelect(ctx->builder, ac_llvm_extract_elem(ctx, src0, 1),
272                             ctx->i16_1, ctx->i16_0, ""),
273          };
274          return ac_build_gather_values(ctx, i, 2);
275       }
276       return LLVMBuildSelect(ctx->builder, src0, ctx->i16_1, ctx->i16_0, "");
277    case 32:
278       return LLVMBuildSelect(ctx->builder, src0, ctx->i32_1, ctx->i32_0, "");
279    case 64:
280       return LLVMBuildSelect(ctx->builder, src0, ctx->i64_1, ctx->i64_0, "");
281    default:
282       unreachable("Unsupported bit size.");
283    }
284 }
285 
emit_i2b(struct ac_llvm_context * ctx,LLVMValueRef src0)286 static LLVMValueRef emit_i2b(struct ac_llvm_context *ctx, LLVMValueRef src0)
287 {
288    LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
289    return LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, zero, "");
290 }
291 
emit_f2f16(struct ac_llvm_context * ctx,LLVMValueRef src0)292 static LLVMValueRef emit_f2f16(struct ac_llvm_context *ctx, LLVMValueRef src0)
293 {
294    LLVMValueRef result;
295    LLVMValueRef cond = NULL;
296 
297    src0 = ac_to_float(ctx, src0);
298    result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
299 
300    if (ctx->gfx_level >= GFX8) {
301       LLVMValueRef args[2];
302       /* Check if the result is a denormal - and flush to 0 if so. */
303       args[0] = result;
304       args[1] = LLVMConstInt(ctx->i32, N_SUBNORMAL | P_SUBNORMAL, false);
305       cond =
306          ac_build_intrinsic(ctx, "llvm.amdgcn.class.f16", ctx->i1, args, 2, 0);
307    }
308 
309    /* need to convert back up to f32 */
310    result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
311 
312    if (ctx->gfx_level >= GFX8)
313       result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
314    else {
315       /* for GFX6-GFX7 */
316       /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
317        * so compare the result and flush to 0 if it's smaller.
318        */
319       LLVMValueRef temp, cond2;
320       temp = emit_intrin_1f_param(ctx, "llvm.fabs", ctx->f32, result);
321       cond = LLVMBuildFCmp(
322          ctx->builder, LLVMRealOGT,
323          LLVMBuildBitCast(ctx->builder, LLVMConstInt(ctx->i32, 0x38800000, false), ctx->f32, ""),
324          temp, "");
325       cond2 = LLVMBuildFCmp(ctx->builder, LLVMRealONE, temp, ctx->f32_0, "");
326       cond = LLVMBuildAnd(ctx->builder, cond, cond2, "");
327       result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
328    }
329    return result;
330 }
331 
emit_umul_high(struct ac_llvm_context * ctx,LLVMValueRef src0,LLVMValueRef src1)332 static LLVMValueRef emit_umul_high(struct ac_llvm_context *ctx, LLVMValueRef src0,
333                                    LLVMValueRef src1)
334 {
335    LLVMValueRef dst64, result;
336    src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
337    src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");
338 
339    dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
340    dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
341    result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
342    return result;
343 }
344 
emit_imul_high(struct ac_llvm_context * ctx,LLVMValueRef src0,LLVMValueRef src1)345 static LLVMValueRef emit_imul_high(struct ac_llvm_context *ctx, LLVMValueRef src0,
346                                    LLVMValueRef src1)
347 {
348    LLVMValueRef dst64, result;
349    src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
350    src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");
351 
352    dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
353    dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
354    result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
355    return result;
356 }
357 
emit_bfm(struct ac_llvm_context * ctx,LLVMValueRef bits,LLVMValueRef offset)358 static LLVMValueRef emit_bfm(struct ac_llvm_context *ctx, LLVMValueRef bits, LLVMValueRef offset)
359 {
360    /* mask = ((1 << bits) - 1) << offset */
361    return LLVMBuildShl(
362       ctx->builder,
363       LLVMBuildSub(ctx->builder, LLVMBuildShl(ctx->builder, ctx->i32_1, bits, ""), ctx->i32_1, ""),
364       offset, "");
365 }
366 
emit_bitfield_select(struct ac_llvm_context * ctx,LLVMValueRef mask,LLVMValueRef insert,LLVMValueRef base)367 static LLVMValueRef emit_bitfield_select(struct ac_llvm_context *ctx, LLVMValueRef mask,
368                                          LLVMValueRef insert, LLVMValueRef base)
369 {
370    /* Calculate:
371     *   (mask & insert) | (~mask & base) = base ^ (mask & (insert ^ base))
372     * Use the right-hand side, which the LLVM backend can convert to V_BFI.
373     */
374    return LLVMBuildXor(
375       ctx->builder, base,
376       LLVMBuildAnd(ctx->builder, mask, LLVMBuildXor(ctx->builder, insert, base, ""), ""), "");
377 }
378 
emit_pack_2x16(struct ac_llvm_context * ctx,LLVMValueRef src0,LLVMValueRef (* pack)(struct ac_llvm_context * ctx,LLVMValueRef args[2]))379 static LLVMValueRef emit_pack_2x16(struct ac_llvm_context *ctx, LLVMValueRef src0,
380                                    LLVMValueRef (*pack)(struct ac_llvm_context *ctx,
381                                                         LLVMValueRef args[2]))
382 {
383    LLVMValueRef comp[2];
384 
385    src0 = ac_to_float(ctx, src0);
386    comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
387    comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");
388 
389    return LLVMBuildBitCast(ctx->builder, pack(ctx, comp), ctx->i32, "");
390 }
391 
emit_unpack_half_2x16(struct ac_llvm_context * ctx,LLVMValueRef src0)392 static LLVMValueRef emit_unpack_half_2x16(struct ac_llvm_context *ctx, LLVMValueRef src0)
393 {
394    LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
395    LLVMValueRef temps[2], val;
396    int i;
397 
398    for (i = 0; i < 2; i++) {
399       val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
400       val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
401       val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
402       temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
403    }
404    return ac_build_gather_values(ctx, temps, 2);
405 }
406 
emit_ddxy(struct ac_nir_context * ctx,nir_op op,LLVMValueRef src0)407 static LLVMValueRef emit_ddxy(struct ac_nir_context *ctx, nir_op op, LLVMValueRef src0)
408 {
409    unsigned mask;
410    int idx;
411    LLVMValueRef result;
412 
413    if (op == nir_op_fddx_fine)
414       mask = AC_TID_MASK_LEFT;
415    else if (op == nir_op_fddy_fine)
416       mask = AC_TID_MASK_TOP;
417    else
418       mask = AC_TID_MASK_TOP_LEFT;
419 
420    /* for DDX we want to next X pixel, DDY next Y pixel. */
421    if (op == nir_op_fddx_fine || op == nir_op_fddx_coarse || op == nir_op_fddx)
422       idx = 1;
423    else
424       idx = 2;
425 
426    result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
427    return result;
428 }
429 
430 struct waterfall_context {
431    LLVMBasicBlockRef phi_bb[2];
432    bool use_waterfall;
433 };
434 
435 /* To deal with divergent descriptors we can create a loop that handles all
436  * lanes with the same descriptor on a given iteration (henceforth a
437  * waterfall loop).
438  *
439  * These helper create the begin and end of the loop leaving the caller
440  * to implement the body.
441  *
442  * params:
443  *  - ctx is the usual nir context
444  *  - wctx is a temporary struct containing some loop info. Can be left uninitialized.
445  *  - value is the possibly divergent value for which we built the loop
446  *  - divergent is whether value is actually divergent. If false we just pass
447  *     things through.
448  */
enter_waterfall(struct ac_nir_context * ctx,struct waterfall_context * wctx,LLVMValueRef value,bool divergent)449 static LLVMValueRef enter_waterfall(struct ac_nir_context *ctx, struct waterfall_context *wctx,
450                                     LLVMValueRef value, bool divergent)
451 {
452    /* If the app claims the value is divergent but it is constant we can
453     * end up with a dynamic index of NULL. */
454    if (!value)
455       divergent = false;
456 
457    wctx->use_waterfall = divergent;
458    if (!divergent)
459       return value;
460 
461    ac_build_bgnloop(&ctx->ac, 6000);
462 
463    LLVMValueRef active = ctx->ac.i1true;
464    LLVMValueRef scalar_value[NIR_MAX_VEC_COMPONENTS];
465 
466    for (unsigned i = 0; i < ac_get_llvm_num_components(value); i++) {
467       LLVMValueRef comp = ac_llvm_extract_elem(&ctx->ac, value, i);
468       scalar_value[i] = ac_build_readlane(&ctx->ac, comp, NULL);
469       active = LLVMBuildAnd(ctx->ac.builder, active,
470                             LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, comp, scalar_value[i], ""), "");
471    }
472 
473    wctx->phi_bb[0] = LLVMGetInsertBlock(ctx->ac.builder);
474    ac_build_ifcc(&ctx->ac, active, 6001);
475 
476    return ac_build_gather_values(&ctx->ac, scalar_value, ac_get_llvm_num_components(value));
477 }
478 
exit_waterfall(struct ac_nir_context * ctx,struct waterfall_context * wctx,LLVMValueRef value)479 static LLVMValueRef exit_waterfall(struct ac_nir_context *ctx, struct waterfall_context *wctx,
480                                    LLVMValueRef value)
481 {
482    LLVMValueRef ret = NULL;
483    LLVMValueRef phi_src[2];
484    LLVMValueRef cc_phi_src[2] = {
485       ctx->ac.i32_0,
486       LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
487    };
488 
489    if (!wctx->use_waterfall)
490       return value;
491 
492    wctx->phi_bb[1] = LLVMGetInsertBlock(ctx->ac.builder);
493 
494    ac_build_endif(&ctx->ac, 6001);
495 
496    if (value) {
497       phi_src[0] = LLVMGetUndef(LLVMTypeOf(value));
498       phi_src[1] = value;
499 
500       ret = ac_build_phi(&ctx->ac, LLVMTypeOf(value), 2, phi_src, wctx->phi_bb);
501    }
502 
503    /*
504     * By using the optimization barrier on the exit decision, we decouple
505     * the operations from the break, and hence avoid LLVM hoisting the
506     * opteration into the break block.
507     */
508    LLVMValueRef cc = ac_build_phi(&ctx->ac, ctx->ac.i32, 2, cc_phi_src, wctx->phi_bb);
509    ac_build_optimization_barrier(&ctx->ac, &cc, false);
510 
511    LLVMValueRef active =
512       LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, cc, ctx->ac.i32_0, "uniform_active2");
513    ac_build_ifcc(&ctx->ac, active, 6002);
514    ac_build_break(&ctx->ac);
515    ac_build_endif(&ctx->ac, 6002);
516 
517    ac_build_endloop(&ctx->ac, 6000);
518    return ret;
519 }
520 
521 static LLVMValueRef
ac_build_const_int_vec(struct ac_llvm_context * ctx,LLVMTypeRef type,long long val,bool sign_extend)522 ac_build_const_int_vec(struct ac_llvm_context *ctx, LLVMTypeRef type, long long val, bool sign_extend)
523 {
524    unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind ? LLVMGetVectorSize(type) : 1;
525 
526    if (num_components == 1)
527       return LLVMConstInt(type, val, sign_extend);
528 
529    assert(num_components == 2);
530    assert(ac_get_elem_bits(ctx, type) == 16);
531 
532    LLVMTypeRef elem_type = LLVMGetElementType(type);
533 
534    LLVMValueRef elems[2];
535    for (unsigned i = 0; i < 2; ++i)
536       elems[i] = LLVMConstInt(elem_type, val, sign_extend);
537 
538    return LLVMConstVector(elems, 2);
539 }
540 
visit_alu(struct ac_nir_context * ctx,const nir_alu_instr * instr)541 static bool visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
542 {
543    LLVMValueRef src[16], result = NULL;
544    unsigned num_components = instr->def.num_components;
545    unsigned src_components;
546    LLVMTypeRef def_type = get_def_type(ctx, &instr->def);
547 
548    assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
549    switch (instr->op) {
550    case nir_op_vec2:
551    case nir_op_vec3:
552    case nir_op_vec4:
553    case nir_op_vec5:
554    case nir_op_vec8:
555    case nir_op_vec16:
556    case nir_op_unpack_32_4x8:
557    case nir_op_unpack_32_2x16:
558    case nir_op_unpack_64_2x32:
559    case nir_op_unpack_64_4x16:
560       src_components = 1;
561       break;
562    case nir_op_pack_snorm_2x16:
563    case nir_op_pack_unorm_2x16:
564    case nir_op_pack_uint_2x16:
565    case nir_op_pack_sint_2x16:
566       src_components = 2;
567       break;
568    case nir_op_cube_amd:
569       src_components = 3;
570       break;
571    case nir_op_pack_32_4x8:
572       src_components = 4;
573       break;
574    default:
575       src_components = num_components;
576       break;
577    }
578    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
579       src[i] = get_alu_src(ctx, instr->src[i], src_components);
580 
581    switch (instr->op) {
582    case nir_op_mov:
583       result = src[0];
584       break;
585    case nir_op_fneg:
586       src[0] = ac_to_float(&ctx->ac, src[0]);
587       result = LLVMBuildFNeg(ctx->ac.builder, src[0], "");
588       if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
589          /* fneg will be optimized by backend compiler with sign
590           * bit removed via XOR. This is probably a LLVM bug.
591           */
592          result = ac_build_canonicalize(&ctx->ac, result, instr->def.bit_size);
593       }
594       break;
595    case nir_op_inot:
596       result = LLVMBuildNot(ctx->ac.builder, src[0], "");
597       break;
598    case nir_op_iadd:
599       if (instr->no_unsigned_wrap)
600          result = LLVMBuildNUWAdd(ctx->ac.builder, src[0], src[1], "");
601       else if (instr->no_signed_wrap)
602          result = LLVMBuildNSWAdd(ctx->ac.builder, src[0], src[1], "");
603       else
604          result = LLVMBuildAdd(ctx->ac.builder, src[0], src[1], "");
605       break;
606    case nir_op_uadd_sat:
607    case nir_op_iadd_sat: {
608       char name[64], type[64];
609       ac_build_type_name_for_intr(def_type, type, sizeof(type));
610       snprintf(name, sizeof(name), "llvm.%cadd.sat.%s",
611                instr->op == nir_op_uadd_sat ? 'u' : 's', type);
612       result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 2, 0);
613       break;
614    }
615    case nir_op_usub_sat:
616    case nir_op_isub_sat: {
617       char name[64], type[64];
618       ac_build_type_name_for_intr(def_type, type, sizeof(type));
619       snprintf(name, sizeof(name), "llvm.%csub.sat.%s",
620                instr->op == nir_op_usub_sat ? 'u' : 's', type);
621       result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 2, 0);
622       break;
623    }
624    case nir_op_fadd:
625       src[0] = ac_to_float(&ctx->ac, src[0]);
626       src[1] = ac_to_float(&ctx->ac, src[1]);
627       result = LLVMBuildFAdd(ctx->ac.builder, src[0], src[1], "");
628       break;
629    case nir_op_fsub:
630       src[0] = ac_to_float(&ctx->ac, src[0]);
631       src[1] = ac_to_float(&ctx->ac, src[1]);
632       result = LLVMBuildFSub(ctx->ac.builder, src[0], src[1], "");
633       break;
634    case nir_op_isub:
635       if (instr->no_unsigned_wrap)
636          result = LLVMBuildNUWSub(ctx->ac.builder, src[0], src[1], "");
637       else if (instr->no_signed_wrap)
638          result = LLVMBuildNSWSub(ctx->ac.builder, src[0], src[1], "");
639       else
640          result = LLVMBuildSub(ctx->ac.builder, src[0], src[1], "");
641       break;
642    case nir_op_imul:
643       if (instr->no_unsigned_wrap)
644          result = LLVMBuildNUWMul(ctx->ac.builder, src[0], src[1], "");
645       else if (instr->no_signed_wrap)
646          result = LLVMBuildNSWMul(ctx->ac.builder, src[0], src[1], "");
647       else
648          result = LLVMBuildMul(ctx->ac.builder, src[0], src[1], "");
649       break;
650    case nir_op_fmul:
651       src[0] = ac_to_float(&ctx->ac, src[0]);
652       src[1] = ac_to_float(&ctx->ac, src[1]);
653       result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
654       break;
655    case nir_op_fmulz:
656       src[0] = ac_to_float(&ctx->ac, src[0]);
657       src[1] = ac_to_float(&ctx->ac, src[1]);
658       result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.fmul.legacy", ctx->ac.f32,
659                                   src, 2, 0);
660       break;
661    case nir_op_frcp:
662       result = emit_intrin_1f_param_scalar(&ctx->ac, "llvm.amdgcn.rcp",
663                                            ac_to_float_type(&ctx->ac, def_type), src[0]);
664       if (ctx->abi->clamp_div_by_zero)
665          result = ac_build_fmin(&ctx->ac, result,
666                                 LLVMConstReal(ac_to_float_type(&ctx->ac, def_type), FLT_MAX));
667       break;
668    case nir_op_iand:
669       result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
670       break;
671    case nir_op_ior:
672       result = LLVMBuildOr(ctx->ac.builder, src[0], src[1], "");
673       break;
674    case nir_op_ixor:
675       result = LLVMBuildXor(ctx->ac.builder, src[0], src[1], "");
676       break;
677    case nir_op_ishl:
678    case nir_op_ishr:
679    case nir_op_ushr: {
680       if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) <
681           ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
682          src[1] = LLVMBuildZExt(ctx->ac.builder, src[1], LLVMTypeOf(src[0]), "");
683       else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) >
684                ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
685          src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1], LLVMTypeOf(src[0]), "");
686       LLVMTypeRef type = LLVMTypeOf(src[1]);
687       src[1] = LLVMBuildAnd(ctx->ac.builder, src[1],
688                             ac_build_const_int_vec(&ctx->ac, type, ac_get_elem_bits(&ctx->ac, type) - 1, false), "");
689       switch (instr->op) {
690       case nir_op_ishl:
691          result = LLVMBuildShl(ctx->ac.builder, src[0], src[1], "");
692          break;
693       case nir_op_ishr:
694          result = LLVMBuildAShr(ctx->ac.builder, src[0], src[1], "");
695          break;
696       case nir_op_ushr:
697          result = LLVMBuildLShr(ctx->ac.builder, src[0], src[1], "");
698          break;
699       default:
700          break;
701       }
702       break;
703    }
704    case nir_op_ilt:
705       result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
706       break;
707    case nir_op_ine:
708       result = emit_int_cmp(&ctx->ac, LLVMIntNE, src[0], src[1]);
709       break;
710    case nir_op_ieq:
711       result = emit_int_cmp(&ctx->ac, LLVMIntEQ, src[0], src[1]);
712       break;
713    case nir_op_ige:
714       result = emit_int_cmp(&ctx->ac, LLVMIntSGE, src[0], src[1]);
715       break;
716    case nir_op_ult:
717       result = emit_int_cmp(&ctx->ac, LLVMIntULT, src[0], src[1]);
718       break;
719    case nir_op_uge:
720       result = emit_int_cmp(&ctx->ac, LLVMIntUGE, src[0], src[1]);
721       break;
722    case nir_op_feq:
723       result = emit_float_cmp(&ctx->ac, LLVMRealOEQ, src[0], src[1]);
724       break;
725    case nir_op_fneu:
726       result = emit_float_cmp(&ctx->ac, LLVMRealUNE, src[0], src[1]);
727       break;
728    case nir_op_flt:
729       result = emit_float_cmp(&ctx->ac, LLVMRealOLT, src[0], src[1]);
730       break;
731    case nir_op_fge:
732       result = emit_float_cmp(&ctx->ac, LLVMRealOGE, src[0], src[1]);
733       break;
734    case nir_op_fabs:
735       result =
736          emit_intrin_1f_param(&ctx->ac, "llvm.fabs", ac_to_float_type(&ctx->ac, def_type), src[0]);
737       if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
738          /* fabs will be optimized by backend compiler with sign
739           * bit removed via AND.
740           */
741          result = ac_build_canonicalize(&ctx->ac, result, instr->def.bit_size);
742       }
743       break;
744    case nir_op_fsat:
745       src[0] = ac_to_float(&ctx->ac, src[0]);
746       result = ac_build_fsat(&ctx->ac, src[0],
747                              ac_to_float_type(&ctx->ac, def_type));
748       break;
749    case nir_op_iabs:
750       result = emit_iabs(&ctx->ac, src[0]);
751       break;
752    case nir_op_imax:
753       result = ac_build_imax(&ctx->ac, src[0], src[1]);
754       break;
755    case nir_op_imin:
756       result = ac_build_imin(&ctx->ac, src[0], src[1]);
757       break;
758    case nir_op_umax:
759       result = ac_build_umax(&ctx->ac, src[0], src[1]);
760       break;
761    case nir_op_umin:
762       result = ac_build_umin(&ctx->ac, src[0], src[1]);
763       break;
764    case nir_op_isign:
765       result = ac_build_isign(&ctx->ac, src[0]);
766       break;
767    case nir_op_fsign:
768       src[0] = ac_to_float(&ctx->ac, src[0]);
769       result = ac_build_fsign(&ctx->ac, src[0]);
770       break;
771    case nir_op_ffloor:
772       result =
773          emit_intrin_1f_param(&ctx->ac, "llvm.floor", ac_to_float_type(&ctx->ac, def_type), src[0]);
774       break;
775    case nir_op_ftrunc:
776       result =
777          emit_intrin_1f_param(&ctx->ac, "llvm.trunc", ac_to_float_type(&ctx->ac, def_type), src[0]);
778       break;
779    case nir_op_fceil:
780       result =
781          emit_intrin_1f_param(&ctx->ac, "llvm.ceil", ac_to_float_type(&ctx->ac, def_type), src[0]);
782       break;
783    case nir_op_fround_even:
784       result =
785          emit_intrin_1f_param(&ctx->ac, "llvm.rint", ac_to_float_type(&ctx->ac, def_type), src[0]);
786       break;
787    case nir_op_ffract:
788       result = emit_intrin_1f_param_scalar(&ctx->ac, "llvm.amdgcn.fract",
789                                            ac_to_float_type(&ctx->ac, def_type), src[0]);
790       break;
791    case nir_op_fsin_amd:
792    case nir_op_fcos_amd:
793       /* before GFX9, v_sin_f32 and v_cos_f32 had a valid input domain of [-256, +256] */
794       if (ctx->ac.gfx_level < GFX9)
795          src[0] = emit_intrin_1f_param_scalar(&ctx->ac, "llvm.amdgcn.fract",
796                                               ac_to_float_type(&ctx->ac, def_type), src[0]);
797       result =
798          emit_intrin_1f_param(&ctx->ac, instr->op == nir_op_fsin_amd ? "llvm.amdgcn.sin" : "llvm.amdgcn.cos",
799                               ac_to_float_type(&ctx->ac, def_type), src[0]);
800       break;
801    case nir_op_fsqrt:
802       result =
803          emit_intrin_1f_param(&ctx->ac, "llvm.sqrt", ac_to_float_type(&ctx->ac, def_type), src[0]);
804       LLVMSetMetadata(result, ctx->ac.fpmath_md_kind, ctx->ac.three_md);
805       break;
806    case nir_op_fexp2:
807       result =
808          emit_intrin_1f_param(&ctx->ac, "llvm.exp2", ac_to_float_type(&ctx->ac, def_type), src[0]);
809       break;
810    case nir_op_flog2:
811       result =
812          emit_intrin_1f_param(&ctx->ac, "llvm.log2", ac_to_float_type(&ctx->ac, def_type), src[0]);
813       break;
814    case nir_op_frsq:
815       result = emit_intrin_1f_param_scalar(&ctx->ac, "llvm.amdgcn.rsq",
816                                            ac_to_float_type(&ctx->ac, def_type), src[0]);
817       if (ctx->abi->clamp_div_by_zero)
818          result = ac_build_fmin(&ctx->ac, result,
819                                 LLVMConstReal(ac_to_float_type(&ctx->ac, def_type), FLT_MAX));
820       break;
821    case nir_op_frexp_exp:
822       src[0] = ac_to_float(&ctx->ac, src[0]);
823       result = ac_build_frexp_exp(&ctx->ac, src[0], ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])));
824       if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) == 16)
825          result = LLVMBuildSExt(ctx->ac.builder, result, ctx->ac.i32, "");
826       break;
827    case nir_op_frexp_sig:
828       src[0] = ac_to_float(&ctx->ac, src[0]);
829       result = ac_build_frexp_mant(&ctx->ac, src[0], instr->def.bit_size);
830       break;
831    case nir_op_fmax:
832       result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum", ac_to_float_type(&ctx->ac, def_type),
833                                     src[0], src[1]);
834       if (ctx->ac.gfx_level < GFX9 && instr->def.bit_size == 32) {
835          /* Only pre-GFX9 chips do not flush denorms. */
836          result = ac_build_canonicalize(&ctx->ac, result, instr->def.bit_size);
837       }
838       break;
839    case nir_op_fmin:
840       result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum", ac_to_float_type(&ctx->ac, def_type),
841                                     src[0], src[1]);
842       if (ctx->ac.gfx_level < GFX9 && instr->def.bit_size == 32) {
843          /* Only pre-GFX9 chips do not flush denorms. */
844          result = ac_build_canonicalize(&ctx->ac, result, instr->def.bit_size);
845       }
846       break;
847    case nir_op_ffma:
848       /* FMA is slow on gfx6-8, so it shouldn't be used. */
849       assert(instr->def.bit_size != 32 || ctx->ac.gfx_level >= GFX9);
850       result = emit_intrin_3f_param(&ctx->ac, "llvm.fma", ac_to_float_type(&ctx->ac, def_type),
851                                     src[0], src[1], src[2]);
852       break;
853    case nir_op_ffmaz:
854       assert(ctx->ac.gfx_level >= GFX10_3);
855       src[0] = ac_to_float(&ctx->ac, src[0]);
856       src[1] = ac_to_float(&ctx->ac, src[1]);
857       src[2] = ac_to_float(&ctx->ac, src[2]);
858       result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.fma.legacy", ctx->ac.f32,
859                                   src, 3, 0);
860       break;
861    case nir_op_ldexp:
862       src[0] = ac_to_float(&ctx->ac, src[0]);
863       if (ac_get_elem_bits(&ctx->ac, def_type) == 32)
864          result = ac_build_intrinsic(&ctx->ac,
865                                      LLVM_VERSION_MAJOR >= 18 ? "llvm.ldexp.f32.i32"
866                                                               : "llvm.amdgcn.ldexp.f32",
867                                      ctx->ac.f32, src, 2, 0);
868       else if (ac_get_elem_bits(&ctx->ac, def_type) == 16)
869          result = ac_build_intrinsic(&ctx->ac,
870                                      LLVM_VERSION_MAJOR >= 18 ? "llvm.ldexp.f16.i32"
871                                                               : "llvm.amdgcn.ldexp.f16",
872                                      ctx->ac.f16, src, 2, 0);
873       else
874          result = ac_build_intrinsic(&ctx->ac,
875                                      LLVM_VERSION_MAJOR >= 18 ? "llvm.ldexp.f64.i32"
876                                                               : "llvm.amdgcn.ldexp.f64",
877                                      ctx->ac.f64, src, 2, 0);
878       break;
879    case nir_op_bfm:
880       result = emit_bfm(&ctx->ac, src[0], src[1]);
881       break;
882    case nir_op_bitfield_select:
883       result = emit_bitfield_select(&ctx->ac, src[0], src[1], src[2]);
884       break;
885    case nir_op_ubfe:
886       result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], false);
887       break;
888    case nir_op_ibfe:
889       result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], true);
890       break;
891    case nir_op_bitfield_reverse:
892       result = ac_build_bitfield_reverse(&ctx->ac, src[0]);
893       break;
894    case nir_op_bit_count:
895       result = ac_build_bit_count(&ctx->ac, src[0]);
896       break;
897    case nir_op_vec2:
898    case nir_op_vec3:
899    case nir_op_vec4:
900    case nir_op_vec5:
901    case nir_op_vec8:
902    case nir_op_vec16:
903       for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
904          src[i] = ac_to_integer(&ctx->ac, src[i]);
905       result = ac_build_gather_values(&ctx->ac, src, num_components);
906       break;
907    case nir_op_f2i8:
908    case nir_op_f2i16:
909    case nir_op_f2i32:
910    case nir_op_f2i64:
911       src[0] = ac_to_float(&ctx->ac, src[0]);
912       result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
913       break;
914    case nir_op_f2u8:
915    case nir_op_f2u16:
916    case nir_op_f2u32:
917    case nir_op_f2u64:
918       src[0] = ac_to_float(&ctx->ac, src[0]);
919       result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
920       break;
921    case nir_op_i2f16:
922    case nir_op_i2f32:
923    case nir_op_i2f64:
924       result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
925       break;
926    case nir_op_u2f16:
927    case nir_op_u2f32:
928    case nir_op_u2f64:
929       result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
930       break;
931    case nir_op_f2f16_rtz: {
932       src[0] = ac_to_float(&ctx->ac, src[0]);
933 
934       if (LLVMTypeOf(src[0]) == ctx->ac.f64)
935          src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
936 
937       /* Fast path conversion. This only works if NIR is vectorized
938        * to vec2 16.
939        */
940       if (LLVMTypeOf(src[0]) == ctx->ac.v2f32) {
941          LLVMValueRef args[] = {
942             ac_llvm_extract_elem(&ctx->ac, src[0], 0),
943             ac_llvm_extract_elem(&ctx->ac, src[0], 1),
944          };
945          result = ac_build_cvt_pkrtz_f16(&ctx->ac, args);
946          break;
947       }
948 
949       assert(ac_get_llvm_num_components(src[0]) == 1);
950       LLVMValueRef param[2] = {src[0], LLVMGetUndef(ctx->ac.f32)};
951       result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
952       result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
953       break;
954    }
955    case nir_op_f2f16:
956    case nir_op_f2f16_rtne:
957    case nir_op_f2f32:
958    case nir_op_f2f64:
959       src[0] = ac_to_float(&ctx->ac, src[0]);
960       if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
961          result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
962       else
963          result =
964             LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
965       break;
966    case nir_op_u2u8:
967    case nir_op_u2u16:
968    case nir_op_u2u32:
969    case nir_op_u2u64:
970       if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
971          result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
972       else
973          result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
974       break;
975    case nir_op_i2i8:
976    case nir_op_i2i16:
977    case nir_op_i2i32:
978    case nir_op_i2i64:
979       if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
980          result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
981       else
982          result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
983       break;
984    case nir_op_bcsel:
985       result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
986       break;
987    case nir_op_find_lsb:
988       result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
989       break;
990    case nir_op_ufind_msb:
991       result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32, false);
992       break;
993    case nir_op_ifind_msb:
994       result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
995       break;
996    case nir_op_ufind_msb_rev:
997       result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32, true);
998       break;
999    case nir_op_ifind_msb_rev:
1000       result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.sffbh.i32", ctx->ac.i32, &src[0], 1,
1001                                   0);
1002       break;
1003    case nir_op_uclz: {
1004       LLVMValueRef params[2] = {
1005          src[0],
1006          ctx->ac.i1false,
1007       };
1008       result = ac_build_intrinsic(&ctx->ac, "llvm.ctlz.i32", ctx->ac.i32, params, 2, 0);
1009       break;
1010    }
1011    case nir_op_uadd_carry:
1012       result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
1013       break;
1014    case nir_op_usub_borrow:
1015       result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
1016       break;
1017    case nir_op_b2f16:
1018    case nir_op_b2f32:
1019    case nir_op_b2f64:
1020       result = emit_b2f(&ctx->ac, src[0], instr->def.bit_size);
1021       break;
1022    case nir_op_b2i8:
1023    case nir_op_b2i16:
1024    case nir_op_b2i32:
1025    case nir_op_b2i64:
1026       result = emit_b2i(&ctx->ac, src[0], instr->def.bit_size);
1027       break;
1028    case nir_op_b2b1: /* after loads */
1029       result = emit_i2b(&ctx->ac, src[0]);
1030       break;
1031    case nir_op_b2b16: /* before stores */
1032       result = LLVMBuildZExt(ctx->ac.builder, src[0], ctx->ac.i16, "");
1033       break;
1034    case nir_op_b2b32: /* before stores */
1035       result = LLVMBuildZExt(ctx->ac.builder, src[0], ctx->ac.i32, "");
1036       break;
1037    case nir_op_fquantize2f16:
1038       result = emit_f2f16(&ctx->ac, src[0]);
1039       break;
1040    case nir_op_umul_high:
1041       result = emit_umul_high(&ctx->ac, src[0], src[1]);
1042       break;
1043    case nir_op_imul_high:
1044       result = emit_imul_high(&ctx->ac, src[0], src[1]);
1045       break;
1046    case nir_op_pack_half_2x16_rtz_split:
1047    case nir_op_pack_half_2x16_split:
1048       src[0] = ac_to_float(&ctx->ac, src[0]);
1049       src[1] = ac_to_float(&ctx->ac, src[1]);
1050       result = LLVMBuildBitCast(ctx->ac.builder,
1051                                 ac_build_cvt_pkrtz_f16(&ctx->ac, src),
1052                                 ctx->ac.i32, "");
1053       break;
1054    case nir_op_pack_snorm_2x16:
1055    case nir_op_pack_unorm_2x16: {
1056       unsigned bit_size = instr->src[0].src.ssa->bit_size;
1057       /* Only support 16 and 32bit. */
1058       assert(bit_size == 16 || bit_size == 32);
1059 
1060       LLVMValueRef data = src[0];
1061       /* Work around for pre-GFX9 GPU which don't have fp16 pknorm instruction. */
1062       if (bit_size == 16 && ctx->ac.gfx_level < GFX9) {
1063          data = LLVMBuildFPExt(ctx->ac.builder, data, ctx->ac.v2f32, "");
1064          bit_size = 32;
1065       }
1066 
1067       LLVMValueRef (*pack)(struct ac_llvm_context *ctx, LLVMValueRef args[2]);
1068       if (bit_size == 32) {
1069          pack = instr->op == nir_op_pack_snorm_2x16 ?
1070             ac_build_cvt_pknorm_i16 : ac_build_cvt_pknorm_u16;
1071       } else {
1072          pack = instr->op == nir_op_pack_snorm_2x16 ?
1073             ac_build_cvt_pknorm_i16_f16 : ac_build_cvt_pknorm_u16_f16;
1074       }
1075       result = emit_pack_2x16(&ctx->ac, data, pack);
1076       break;
1077    }
1078    case nir_op_pack_uint_2x16: {
1079       LLVMValueRef comp[2];
1080 
1081       comp[0] = LLVMBuildExtractElement(ctx->ac.builder, src[0], ctx->ac.i32_0, "");
1082       comp[1] = LLVMBuildExtractElement(ctx->ac.builder, src[0], ctx->ac.i32_1, "");
1083 
1084       result = ac_build_cvt_pk_u16(&ctx->ac, comp, 16, false);
1085       break;
1086    }
1087    case nir_op_pack_sint_2x16: {
1088       LLVMValueRef comp[2];
1089 
1090       comp[0] = LLVMBuildExtractElement(ctx->ac.builder, src[0], ctx->ac.i32_0, "");
1091       comp[1] = LLVMBuildExtractElement(ctx->ac.builder, src[0], ctx->ac.i32_1, "");
1092 
1093       result = ac_build_cvt_pk_i16(&ctx->ac, comp, 16, false);
1094       break;
1095    }
1096    case nir_op_unpack_half_2x16_split_x: {
1097       assert(ac_get_llvm_num_components(src[0]) == 1);
1098       LLVMValueRef tmp = emit_unpack_half_2x16(&ctx->ac, src[0]);
1099       result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_0, "");
1100       break;
1101    }
1102    case nir_op_unpack_half_2x16_split_y: {
1103       assert(ac_get_llvm_num_components(src[0]) == 1);
1104       LLVMValueRef tmp = emit_unpack_half_2x16(&ctx->ac, src[0]);
1105       result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_1, "");
1106       break;
1107    }
1108    case nir_op_fddx:
1109    case nir_op_fddy:
1110    case nir_op_fddx_fine:
1111    case nir_op_fddy_fine:
1112    case nir_op_fddx_coarse:
1113    case nir_op_fddy_coarse:
1114       result = emit_ddxy(ctx, instr->op, src[0]);
1115       break;
1116 
1117    case nir_op_unpack_64_4x16: {
1118       result = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v4i16, "");
1119       break;
1120    }
1121 
1122    case nir_op_unpack_64_2x32: {
1123       result = LLVMBuildBitCast(ctx->ac.builder, src[0],
1124             ctx->ac.v2i32, "");
1125       break;
1126    }
1127    case nir_op_unpack_64_2x32_split_x: {
1128       assert(ac_get_llvm_num_components(src[0]) == 1);
1129       LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i32, "");
1130       result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_0, "");
1131       break;
1132    }
1133    case nir_op_unpack_64_2x32_split_y: {
1134       assert(ac_get_llvm_num_components(src[0]) == 1);
1135       LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i32, "");
1136       result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_1, "");
1137       break;
1138    }
1139 
1140    case nir_op_pack_64_2x32_split: {
1141       LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1142       result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
1143       break;
1144    }
1145 
1146    case nir_op_pack_32_4x8: {
1147       result = LLVMBuildBitCast(ctx->ac.builder, src[0],
1148             ctx->ac.i32, "");
1149       break;
1150    }
1151    case nir_op_pack_32_2x16_split: {
1152       LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1153       result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i32, "");
1154       break;
1155    }
1156 
1157    case nir_op_unpack_32_4x8:
1158       result = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v4i8, "");
1159       break;
1160    case nir_op_unpack_32_2x16: {
1161       result = LLVMBuildBitCast(ctx->ac.builder, src[0],
1162             ctx->ac.v2i16, "");
1163       break;
1164    }
1165    case nir_op_unpack_32_2x16_split_x: {
1166       LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i16, "");
1167       result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_0, "");
1168       break;
1169    }
1170    case nir_op_unpack_32_2x16_split_y: {
1171       LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i16, "");
1172       result = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_1, "");
1173       break;
1174    }
1175 
1176    case nir_op_cube_amd: {
1177       src[0] = ac_to_float(&ctx->ac, src[0]);
1178       LLVMValueRef results[4];
1179       LLVMValueRef in[3];
1180       for (unsigned chan = 0; chan < 3; chan++)
1181          in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1182       results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc", ctx->ac.f32, in, 3, 0);
1183       results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc", ctx->ac.f32, in, 3, 0);
1184       results[2] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema", ctx->ac.f32, in, 3, 0);
1185       results[3] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubeid", ctx->ac.f32, in, 3, 0);
1186       result = ac_build_gather_values(&ctx->ac, results, 4);
1187       break;
1188    }
1189 
1190    case nir_op_extract_u8:
1191    case nir_op_extract_i8:
1192    case nir_op_extract_u16:
1193    case nir_op_extract_i16: {
1194       bool is_signed = instr->op == nir_op_extract_i16 || instr->op == nir_op_extract_i8;
1195       unsigned size = instr->op == nir_op_extract_u8 || instr->op == nir_op_extract_i8 ? 8 : 16;
1196       LLVMValueRef offset = LLVMConstInt(LLVMTypeOf(src[0]), nir_src_as_uint(instr->src[1].src) * size, false);
1197       result = LLVMBuildLShr(ctx->ac.builder, src[0], offset, "");
1198       result = LLVMBuildTrunc(ctx->ac.builder, result, LLVMIntTypeInContext(ctx->ac.context, size), "");
1199       if (is_signed)
1200          result = LLVMBuildSExt(ctx->ac.builder, result, LLVMTypeOf(src[0]), "");
1201       else
1202          result = LLVMBuildZExt(ctx->ac.builder, result, LLVMTypeOf(src[0]), "");
1203       break;
1204    }
1205 
1206    case nir_op_insert_u8:
1207    case nir_op_insert_u16: {
1208       unsigned size = instr->op == nir_op_insert_u8 ? 8 : 16;
1209       LLVMValueRef offset = LLVMConstInt(LLVMTypeOf(src[0]), nir_src_as_uint(instr->src[1].src) * size, false);
1210       LLVMValueRef mask = LLVMConstInt(LLVMTypeOf(src[0]), u_bit_consecutive(0, size), false);
1211       result = LLVMBuildShl(ctx->ac.builder, LLVMBuildAnd(ctx->ac.builder, src[0], mask, ""), offset, "");
1212       break;
1213    }
1214 
1215    case nir_op_sdot_4x8_iadd:
1216    case nir_op_sdot_4x8_iadd_sat: {
1217       if (ctx->ac.gfx_level >= GFX11) {
1218          result = ac_build_sudot_4x8(&ctx->ac, src[0], src[1], src[2],
1219                                      instr->op == nir_op_sdot_4x8_iadd_sat, 0x3);
1220       } else {
1221          const char *name = "llvm.amdgcn.sdot4";
1222          src[3] = LLVMConstInt(ctx->ac.i1, instr->op == nir_op_sdot_4x8_iadd_sat, false);
1223          result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 4, 0);
1224       }
1225       break;
1226    }
1227    case nir_op_sudot_4x8_iadd:
1228    case nir_op_sudot_4x8_iadd_sat: {
1229       result = ac_build_sudot_4x8(&ctx->ac, src[0], src[1], src[2],
1230                                   instr->op == nir_op_sudot_4x8_iadd_sat, 0x1);
1231       break;
1232    }
1233    case nir_op_udot_4x8_uadd:
1234    case nir_op_udot_4x8_uadd_sat: {
1235       const char *name = "llvm.amdgcn.udot4";
1236       src[3] = LLVMConstInt(ctx->ac.i1, instr->op == nir_op_udot_4x8_uadd_sat, false);
1237       result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 4, 0);
1238       break;
1239    }
1240 
1241    case nir_op_sdot_2x16_iadd:
1242    case nir_op_udot_2x16_uadd:
1243    case nir_op_sdot_2x16_iadd_sat:
1244    case nir_op_udot_2x16_uadd_sat: {
1245       const char *name = instr->op == nir_op_sdot_2x16_iadd ||
1246                          instr->op == nir_op_sdot_2x16_iadd_sat
1247                          ? "llvm.amdgcn.sdot2" : "llvm.amdgcn.udot2";
1248       src[0] = LLVMBuildBitCast(ctx->ac.builder, src[0], ctx->ac.v2i16, "");
1249       src[1] = LLVMBuildBitCast(ctx->ac.builder, src[1], ctx->ac.v2i16, "");
1250       src[3] = LLVMConstInt(ctx->ac.i1, instr->op == nir_op_sdot_2x16_iadd_sat ||
1251                                         instr->op == nir_op_udot_2x16_uadd_sat, false);
1252       result = ac_build_intrinsic(&ctx->ac, name, def_type, src, 4, 0);
1253       break;
1254    }
1255 
1256    case nir_op_msad_4x8:
1257       result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.msad.u8", ctx->ac.i32,
1258                                   (LLVMValueRef[]){src[1], src[0], src[2]}, 3, 0);
1259       break;
1260 
1261    default:
1262       fprintf(stderr, "Unknown NIR alu instr: ");
1263       nir_print_instr(&instr->instr, stderr);
1264       fprintf(stderr, "\n");
1265       return false;
1266    }
1267 
1268    if (result) {
1269       result = ac_to_integer_or_pointer(&ctx->ac, result);
1270       ctx->ssa_defs[instr->def.index] = result;
1271    }
1272    return true;
1273 }
1274 
visit_load_const(struct ac_nir_context * ctx,const nir_load_const_instr * instr)1275 static bool visit_load_const(struct ac_nir_context *ctx, const nir_load_const_instr *instr)
1276 {
1277    LLVMValueRef values[16], value = NULL;
1278    LLVMTypeRef element_type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
1279 
1280    for (unsigned i = 0; i < instr->def.num_components; ++i) {
1281       switch (instr->def.bit_size) {
1282       case 1:
1283          values[i] = LLVMConstInt(element_type, instr->value[i].b, false);
1284          break;
1285       case 8:
1286          values[i] = LLVMConstInt(element_type, instr->value[i].u8, false);
1287          break;
1288       case 16:
1289          values[i] = LLVMConstInt(element_type, instr->value[i].u16, false);
1290          break;
1291       case 32:
1292          values[i] = LLVMConstInt(element_type, instr->value[i].u32, false);
1293          break;
1294       case 64:
1295          values[i] = LLVMConstInt(element_type, instr->value[i].u64, false);
1296          break;
1297       default:
1298          fprintf(stderr, "unsupported nir load_const bit_size: %d\n", instr->def.bit_size);
1299          return false;
1300       }
1301    }
1302    if (instr->def.num_components > 1) {
1303       value = LLVMConstVector(values, instr->def.num_components);
1304    } else
1305       value = values[0];
1306 
1307    ctx->ssa_defs[instr->def.index] = value;
1308    return true;
1309 }
1310 
1311 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
1312  * incorrectly forces nearest filtering if the texture format is integer.
1313  * The only effect it has on Gather4, which always returns 4 texels for
1314  * bilinear filtering, is that the final coordinates are off by 0.5 of
1315  * the texel size.
1316  *
1317  * The workaround is to subtract 0.5 from the unnormalized coordinates,
1318  * or (0.5 / size) from the normalized coordinates.
1319  *
1320  * However, cube textures with 8_8_8_8 data formats require a different
1321  * workaround of overriding the num format to USCALED/SSCALED. This would lose
1322  * precision in 32-bit data formats, so it needs to be applied dynamically at
1323  * runtime. In this case, return an i1 value that indicates whether the
1324  * descriptor was overridden (and hence a fixup of the sampler result is needed).
1325  */
lower_gather4_integer(struct ac_llvm_context * ctx,struct ac_image_args * args,const nir_tex_instr * instr)1326 static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx, struct ac_image_args *args,
1327                                           const nir_tex_instr *instr)
1328 {
1329    nir_alu_type stype = nir_alu_type_get_base_type(instr->dest_type);
1330    LLVMValueRef wa_8888 = NULL;
1331    LLVMValueRef half_texel[2];
1332    LLVMValueRef result;
1333 
1334    assert(stype == nir_type_int || stype == nir_type_uint);
1335 
1336    if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1337       LLVMValueRef formats;
1338       LLVMValueRef data_format;
1339       LLVMValueRef wa_formats;
1340 
1341       formats = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
1342 
1343       data_format = LLVMBuildLShr(ctx->builder, formats, LLVMConstInt(ctx->i32, 20, false), "");
1344       data_format =
1345          LLVMBuildAnd(ctx->builder, data_format, LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
1346       wa_8888 = LLVMBuildICmp(ctx->builder, LLVMIntEQ, data_format,
1347                               LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false), "");
1348 
1349       uint32_t wa_num_format = stype == nir_type_uint
1350                                   ? S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED)
1351                                   : S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
1352       wa_formats = LLVMBuildAnd(ctx->builder, formats,
1353                                 LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false), "");
1354       wa_formats =
1355          LLVMBuildOr(ctx->builder, wa_formats, LLVMConstInt(ctx->i32, wa_num_format, false), "");
1356 
1357       formats = LLVMBuildSelect(ctx->builder, wa_8888, wa_formats, formats, "");
1358       args->resource =
1359          LLVMBuildInsertElement(ctx->builder, args->resource, formats, ctx->i32_1, "");
1360    }
1361 
1362    if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
1363       assert(!wa_8888);
1364       half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
1365    } else {
1366       struct ac_image_args resinfo = {0};
1367       LLVMBasicBlockRef bbs[2];
1368 
1369       LLVMValueRef unnorm = NULL;
1370       LLVMValueRef default_offset = ctx->f32_0;
1371       if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D && !instr->is_array) {
1372          /* In vulkan, whether the sampler uses unnormalized
1373           * coordinates or not is a dynamic property of the
1374           * sampler. Hence, to figure out whether or not we
1375           * need to divide by the texture size, we need to test
1376           * the sampler at runtime. This tests the bit set by
1377           * radv_init_sampler().
1378           */
1379          LLVMValueRef sampler0 =
1380             LLVMBuildExtractElement(ctx->builder, args->sampler, ctx->i32_0, "");
1381          sampler0 = LLVMBuildLShr(ctx->builder, sampler0, LLVMConstInt(ctx->i32, 15, false), "");
1382          sampler0 = LLVMBuildAnd(ctx->builder, sampler0, ctx->i32_1, "");
1383          unnorm = LLVMBuildICmp(ctx->builder, LLVMIntEQ, sampler0, ctx->i32_1, "");
1384          default_offset = LLVMConstReal(ctx->f32, -0.5);
1385       }
1386 
1387       bbs[0] = LLVMGetInsertBlock(ctx->builder);
1388       if (wa_8888 || unnorm) {
1389          assert(!(wa_8888 && unnorm));
1390          LLVMValueRef not_needed = wa_8888 ? wa_8888 : unnorm;
1391          /* Skip the texture size query entirely if we don't need it. */
1392          ac_build_ifcc(ctx, LLVMBuildNot(ctx->builder, not_needed, ""), 2000);
1393          bbs[1] = LLVMGetInsertBlock(ctx->builder);
1394       }
1395 
1396       /* Query the texture size. */
1397       resinfo.dim = ac_get_sampler_dim(ctx->gfx_level, instr->sampler_dim, instr->is_array);
1398       resinfo.opcode = ac_image_get_resinfo;
1399       resinfo.dmask = 0xf;
1400       resinfo.lod = ctx->i32_0;
1401       resinfo.resource = args->resource;
1402       resinfo.attributes = AC_ATTR_INVARIANT_LOAD;
1403       LLVMValueRef size = ac_build_image_opcode(ctx, &resinfo);
1404 
1405       /* Compute -0.5 / size. */
1406       for (unsigned c = 0; c < 2; c++) {
1407          half_texel[c] =
1408             LLVMBuildExtractElement(ctx->builder, size, LLVMConstInt(ctx->i32, c, 0), "");
1409          half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
1410          half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
1411          half_texel[c] =
1412             LLVMBuildFMul(ctx->builder, half_texel[c], LLVMConstReal(ctx->f32, -0.5), "");
1413       }
1414 
1415       if (wa_8888 || unnorm) {
1416          ac_build_endif(ctx, 2000);
1417 
1418          for (unsigned c = 0; c < 2; c++) {
1419             LLVMValueRef values[2] = {default_offset, half_texel[c]};
1420             half_texel[c] = ac_build_phi(ctx, ctx->f32, 2, values, bbs);
1421          }
1422       }
1423    }
1424 
1425    for (unsigned c = 0; c < 2; c++) {
1426       LLVMValueRef tmp;
1427       tmp = LLVMBuildBitCast(ctx->builder, args->coords[c], ctx->f32, "");
1428       args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
1429    }
1430 
1431    args->attributes = AC_ATTR_INVARIANT_LOAD;
1432    result = ac_build_image_opcode(ctx, args);
1433 
1434    if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1435       LLVMValueRef tmp, tmp2;
1436 
1437       /* if the cube workaround is in place, f2i the result. */
1438       for (unsigned c = 0; c < 4; c++) {
1439          tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
1440          if (stype == nir_type_uint)
1441             tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
1442          else
1443             tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
1444          tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
1445          tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
1446          tmp = LLVMBuildSelect(ctx->builder, wa_8888, tmp2, tmp, "");
1447          tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
1448          result =
1449             LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
1450       }
1451    }
1452    return result;
1453 }
1454 
build_tex_intrinsic(struct ac_nir_context * ctx,const nir_tex_instr * instr,struct ac_image_args * args)1455 static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx, const nir_tex_instr *instr,
1456                                         struct ac_image_args *args)
1457 {
1458    assert((!args->tfe || !args->d16) && "unsupported");
1459 
1460    if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
1461       unsigned mask = nir_def_components_read(&instr->def);
1462 
1463       /* Buffers don't support A16. */
1464       if (args->a16)
1465          args->coords[0] = LLVMBuildZExt(ctx->ac.builder, args->coords[0], ctx->ac.i32, "");
1466 
1467       return ac_build_buffer_load_format(&ctx->ac, args->resource, args->coords[0], ctx->ac.i32_0,
1468                                          util_last_bit(mask), 0, true,
1469                                          instr->def.bit_size == 16,
1470                                          args->tfe);
1471    }
1472 
1473    args->opcode = ac_image_sample;
1474 
1475    switch (instr->op) {
1476    case nir_texop_txf:
1477    case nir_texop_txf_ms:
1478       args->opcode = args->level_zero || instr->sampler_dim == GLSL_SAMPLER_DIM_MS
1479                         ? ac_image_load
1480                         : ac_image_load_mip;
1481       args->level_zero = false;
1482       break;
1483    case nir_texop_txs:
1484    case nir_texop_query_levels:
1485    case nir_texop_texture_samples:
1486       assert(!"should have been lowered");
1487       break;
1488    case nir_texop_tex:
1489       if (ctx->stage != MESA_SHADER_FRAGMENT &&
1490           (!gl_shader_stage_is_compute(ctx->stage) ||
1491            ctx->info->cs.derivative_group == DERIVATIVE_GROUP_NONE)) {
1492          assert(!args->lod);
1493          args->level_zero = true;
1494       }
1495       break;
1496    case nir_texop_tg4:
1497       args->opcode = ac_image_gather4;
1498       if (!args->lod && !instr->is_gather_implicit_lod)
1499          args->level_zero = true;
1500       /* GFX11 supports implicit LOD, but the extension is unsupported. */
1501       assert(args->level_zero || ctx->ac.gfx_level < GFX11);
1502       break;
1503    case nir_texop_lod:
1504       args->opcode = ac_image_get_lod;
1505       break;
1506    case nir_texop_fragment_fetch_amd:
1507    case nir_texop_fragment_mask_fetch_amd:
1508       args->opcode = ac_image_load;
1509       args->level_zero = false;
1510       break;
1511    default:
1512       break;
1513    }
1514 
1515    /* MI200 doesn't have image_sample_lz, but image_sample behaves like lz. */
1516    if (!ctx->ac.info->has_3d_cube_border_color_mipmap)
1517       args->level_zero = false;
1518 
1519    if (instr->op == nir_texop_tg4 && ctx->ac.gfx_level <= GFX8 &&
1520        (instr->dest_type & (nir_type_int | nir_type_uint))) {
1521       return lower_gather4_integer(&ctx->ac, args, instr);
1522    }
1523 
1524    args->attributes = AC_ATTR_INVARIANT_LOAD;
1525    bool cs_derivs =
1526       gl_shader_stage_is_compute(ctx->stage) && ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
1527    if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
1528       /* Prevent texture instructions with implicit derivatives from being
1529        * sinked into branches. */
1530       switch (instr->op) {
1531       case nir_texop_tex:
1532       case nir_texop_txb:
1533       case nir_texop_lod:
1534          args->attributes |= AC_ATTR_CONVERGENT;
1535          break;
1536       default:
1537          break;
1538       }
1539    }
1540 
1541    return ac_build_image_opcode(&ctx->ac, args);
1542 }
1543 
visit_load_push_constant(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)1544 static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
1545 {
1546    LLVMValueRef ptr, addr;
1547    LLVMValueRef src0 = get_src(ctx, instr->src[0]);
1548    unsigned index = nir_intrinsic_base(instr);
1549 
1550    addr = LLVMConstInt(ctx->ac.i32, index, 0);
1551    addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
1552 
1553    /* Load constant values from user SGPRS when possible, otherwise
1554     * fallback to the default path that loads directly from memory.
1555     */
1556    if (LLVMIsConstant(src0) && instr->def.bit_size >= 32) {
1557       unsigned count = instr->def.num_components;
1558       unsigned offset = index;
1559 
1560       if (instr->def.bit_size == 64)
1561          count *= 2;
1562 
1563       offset += LLVMConstIntGetZExtValue(src0);
1564       offset /= 4;
1565 
1566       uint64_t mask = BITFIELD64_MASK(count) << offset;
1567       if ((ctx->args->inline_push_const_mask | mask) == ctx->args->inline_push_const_mask &&
1568           offset + count <= (sizeof(ctx->args->inline_push_const_mask) * 8u)) {
1569          LLVMValueRef *const push_constants = alloca(count * sizeof(LLVMValueRef));
1570          unsigned arg_index =
1571             util_bitcount64(ctx->args->inline_push_const_mask & BITFIELD64_MASK(offset));
1572          for (unsigned i = 0; i < count; i++)
1573             push_constants[i] = ac_get_arg(&ctx->ac, ctx->args->inline_push_consts[arg_index++]);
1574          LLVMValueRef res = ac_build_gather_values(&ctx->ac, push_constants, count);
1575          return instr->def.bit_size == 64
1576                    ? LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->def), "")
1577                    : res;
1578       }
1579    }
1580 
1581    struct ac_llvm_pointer pc = ac_get_ptr_arg(&ctx->ac, ctx->args, ctx->args->push_constants);
1582    ptr = LLVMBuildGEP2(ctx->ac.builder, pc.t, pc.v, &addr, 1, "");
1583 
1584    if (instr->def.bit_size == 8) {
1585       unsigned load_dwords = instr->def.num_components > 1 ? 2 : 1;
1586       LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i8, 4 * load_dwords);
1587       ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1588       LLVMValueRef res = LLVMBuildLoad2(ctx->ac.builder, vec_type, ptr, "");
1589 
1590       LLVMValueRef params[3];
1591       if (load_dwords > 1) {
1592          LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.v2i32, "");
1593          params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec,
1594                                              ctx->ac.i32_1, "");
1595          params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec,
1596                                              ctx->ac.i32_0, "");
1597       } else {
1598          res = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.i32, "");
1599          params[0] = ctx->ac.i32_0;
1600          params[1] = res;
1601       }
1602       params[2] = addr;
1603       res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.alignbyte", ctx->ac.i32, params, 3, 0);
1604 
1605       res = LLVMBuildTrunc(
1606          ctx->ac.builder, res,
1607          LLVMIntTypeInContext(ctx->ac.context, instr->def.num_components * 8), "");
1608       if (instr->def.num_components > 1)
1609          res = LLVMBuildBitCast(ctx->ac.builder, res,
1610                                 LLVMVectorType(ctx->ac.i8, instr->def.num_components), "");
1611       return res;
1612    } else if (instr->def.bit_size == 16) {
1613       unsigned load_dwords = instr->def.num_components / 2 + 1;
1614       LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i16, 2 * load_dwords);
1615       ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1616       LLVMValueRef res = LLVMBuildLoad2(ctx->ac.builder, vec_type, ptr, "");
1617       res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1618       LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1619       cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1620       LLVMValueRef mask[] = {
1621          ctx->ac.i32_0, ctx->ac.i32_1,
1622          LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1623          LLVMConstInt(ctx->ac.i32, 4, false)};
1624       LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->def.num_components);
1625       LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->def.num_components);
1626       LLVMValueRef shuffle_aligned =
1627          LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1628       LLVMValueRef shuffle_unaligned =
1629          LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1630       res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1631       return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->def), "");
1632    }
1633 
1634    LLVMTypeRef ptr_type = get_def_type(ctx, &instr->def);
1635    ptr = ac_cast_ptr(&ctx->ac, ptr, ptr_type);
1636 
1637    return LLVMBuildLoad2(ctx->ac.builder, ptr_type, ptr, "");
1638 }
1639 
visit_get_ssbo_size(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)1640 static LLVMValueRef visit_get_ssbo_size(struct ac_nir_context *ctx,
1641                                         const nir_intrinsic_instr *instr)
1642 {
1643    bool non_uniform = nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM;
1644 
1645    LLVMValueRef rsrc = get_src(ctx, instr->src[0]);
1646    if (ctx->abi->load_ssbo)
1647       rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc, false, non_uniform);
1648 
1649    return LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, false), "");
1650 }
1651 
extract_vector_range(struct ac_llvm_context * ctx,LLVMValueRef src,unsigned start,unsigned count)1652 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1653                                          unsigned start, unsigned count)
1654 {
1655    LLVMValueRef mask[] = {ctx->i32_0, ctx->i32_1, LLVMConstInt(ctx->i32, 2, false),
1656                           LLVMConstInt(ctx->i32, 3, false)};
1657 
1658    unsigned src_elements = ac_get_llvm_num_components(src);
1659 
1660    if (count == src_elements) {
1661       assert(start == 0);
1662       return src;
1663    } else if (count == 1) {
1664       assert(start < src_elements);
1665       return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1666    } else {
1667       assert(start + count <= src_elements);
1668       assert(count <= 4);
1669       LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1670       return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1671    }
1672 }
1673 
enter_waterfall_ssbo(struct ac_nir_context * ctx,struct waterfall_context * wctx,const nir_intrinsic_instr * instr,nir_src src)1674 static LLVMValueRef enter_waterfall_ssbo(struct ac_nir_context *ctx, struct waterfall_context *wctx,
1675                                          const nir_intrinsic_instr *instr, nir_src src)
1676 {
1677    return enter_waterfall(ctx, wctx, get_src(ctx, src),
1678                           nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
1679 }
1680 
visit_store_ssbo(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)1681 static void visit_store_ssbo(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
1682 {
1683    LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1684    int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1685    unsigned writemask = nir_intrinsic_write_mask(instr);
1686    enum gl_access_qualifier access = ac_get_mem_access_flags(instr);
1687 
1688    struct waterfall_context wctx;
1689    LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[1]);
1690 
1691    LLVMValueRef rsrc = ctx->abi->load_ssbo ?
1692       ctx->abi->load_ssbo(ctx->abi, rsrc_base, true, false) : rsrc_base;
1693 
1694    LLVMValueRef base_data = src_data;
1695    base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1696    LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1697 
1698    while (writemask) {
1699       int start, count;
1700       LLVMValueRef data, offset;
1701       LLVMTypeRef data_type;
1702 
1703       u_bit_scan_consecutive_range(&writemask, &start, &count);
1704 
1705       if (count == 3 && elem_size_bytes != 4) {
1706          writemask |= 1 << (start + 2);
1707          count = 2;
1708       }
1709       int num_bytes = count * elem_size_bytes; /* count in bytes */
1710 
1711       /* we can only store 4 DWords at the same time.
1712        * can only happen for 64 Bit vectors. */
1713       if (num_bytes > 16) {
1714          writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1715          count = 2;
1716          num_bytes = 16;
1717       }
1718 
1719       /* check alignment of 16 Bit stores */
1720       if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1721          writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1722          count = 1;
1723          num_bytes = 2;
1724       }
1725 
1726       /* Due to alignment issues, split stores of 8-bit/16-bit
1727        * vectors.
1728        */
1729       if (ctx->ac.gfx_level == GFX6 && count > 1 && elem_size_bytes < 4) {
1730          writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1731          count = 1;
1732          num_bytes = elem_size_bytes;
1733       }
1734 
1735       data = extract_vector_range(&ctx->ac, base_data, start, count);
1736 
1737       offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1738                             LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1739 
1740       if (num_bytes == 1) {
1741          ac_build_buffer_store_byte(&ctx->ac, rsrc, data, offset, ctx->ac.i32_0, access);
1742       } else if (num_bytes == 2) {
1743          ac_build_buffer_store_short(&ctx->ac, rsrc, data, offset, ctx->ac.i32_0, access);
1744       } else {
1745          switch (num_bytes) {
1746          case 16: /* v4f32 */
1747             data_type = ctx->ac.v4f32;
1748             break;
1749          case 12: /* v3f32 */
1750             data_type = ctx->ac.v3f32;
1751             break;
1752          case 8: /* v2f32 */
1753             data_type = ctx->ac.v2f32;
1754             break;
1755          case 4: /* f32 */
1756             data_type = ctx->ac.f32;
1757             break;
1758          default:
1759             unreachable("Malformed vector store.");
1760          }
1761          data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1762 
1763          ac_build_buffer_store_dword(&ctx->ac, rsrc, data, NULL, offset,
1764                                      ctx->ac.i32_0, access);
1765       }
1766    }
1767 
1768    exit_waterfall(ctx, &wctx, NULL);
1769 }
1770 
emit_ssbo_comp_swap_64(struct ac_nir_context * ctx,LLVMValueRef descriptor,LLVMValueRef offset,LLVMValueRef compare,LLVMValueRef exchange,bool image)1771 static LLVMValueRef emit_ssbo_comp_swap_64(struct ac_nir_context *ctx, LLVMValueRef descriptor,
1772                                            LLVMValueRef offset, LLVMValueRef compare,
1773                                            LLVMValueRef exchange, bool image)
1774 {
1775    LLVMBasicBlockRef start_block = NULL, then_block = NULL;
1776    if (ctx->abi->robust_buffer_access || image) {
1777       LLVMValueRef size = ac_llvm_extract_elem(&ctx->ac, descriptor, 2);
1778 
1779       LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
1780       start_block = LLVMGetInsertBlock(ctx->ac.builder);
1781 
1782       ac_build_ifcc(&ctx->ac, cond, -1);
1783 
1784       then_block = LLVMGetInsertBlock(ctx->ac.builder);
1785    }
1786 
1787    if (image)
1788       offset = LLVMBuildMul(ctx->ac.builder, offset, LLVMConstInt(ctx->ac.i32, 8, false), "");
1789 
1790    LLVMValueRef ptr_parts[2] = {
1791       ac_llvm_extract_elem(&ctx->ac, descriptor, 0),
1792       LLVMBuildAnd(ctx->ac.builder, ac_llvm_extract_elem(&ctx->ac, descriptor, 1),
1793                    LLVMConstInt(ctx->ac.i32, 65535, 0), "")};
1794 
1795    ptr_parts[1] = LLVMBuildTrunc(ctx->ac.builder, ptr_parts[1], ctx->ac.i16, "");
1796    ptr_parts[1] = LLVMBuildSExt(ctx->ac.builder, ptr_parts[1], ctx->ac.i32, "");
1797 
1798    offset = LLVMBuildZExt(ctx->ac.builder, offset, ctx->ac.i64, "");
1799 
1800    LLVMValueRef ptr = ac_build_gather_values(&ctx->ac, ptr_parts, 2);
1801    ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ctx->ac.i64, "");
1802    ptr = LLVMBuildAdd(ctx->ac.builder, ptr, offset, "");
1803    ptr = LLVMBuildIntToPtr(ctx->ac.builder, ptr, LLVMPointerType(ctx->ac.i64, AC_ADDR_SPACE_GLOBAL),
1804                            "");
1805 
1806    LLVMValueRef result =
1807       ac_build_atomic_cmp_xchg(&ctx->ac, ptr, compare, exchange, "singlethread-one-as");
1808    result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
1809 
1810    if (ctx->abi->robust_buffer_access || image) {
1811       ac_build_endif(&ctx->ac, -1);
1812 
1813       LLVMBasicBlockRef incoming_blocks[2] = {
1814          start_block,
1815          then_block,
1816       };
1817 
1818       LLVMValueRef incoming_values[2] = {
1819          ctx->ac.i64_0,
1820          result,
1821       };
1822       LLVMValueRef ret = LLVMBuildPhi(ctx->ac.builder, ctx->ac.i64, "");
1823       LLVMAddIncoming(ret, incoming_values, incoming_blocks, 2);
1824       return ret;
1825    } else {
1826       return result;
1827    }
1828 }
1829 
1830 static const char *
translate_atomic_op_str(nir_atomic_op op)1831 translate_atomic_op_str(nir_atomic_op op)
1832 {
1833    switch (op) {
1834    case nir_atomic_op_iadd:     return "add";
1835    case nir_atomic_op_imin:     return "smin";
1836    case nir_atomic_op_umin:     return "umin";
1837    case nir_atomic_op_imax:     return "smax";
1838    case nir_atomic_op_umax:     return "umax";
1839    case nir_atomic_op_iand:     return "and";
1840    case nir_atomic_op_ior:      return "or";
1841    case nir_atomic_op_ixor:     return "xor";
1842    case nir_atomic_op_fadd:     return "fadd";
1843    case nir_atomic_op_fmin:     return "fmin";
1844    case nir_atomic_op_fmax:     return "fmax";
1845    case nir_atomic_op_xchg:     return "swap";
1846    case nir_atomic_op_cmpxchg:  return "cmpswap";
1847    case nir_atomic_op_inc_wrap: return "inc";
1848    case nir_atomic_op_dec_wrap: return "dec";
1849    default: abort();
1850    }
1851 }
1852 
1853 static LLVMAtomicRMWBinOp
translate_atomic_op(nir_atomic_op op)1854 translate_atomic_op(nir_atomic_op op)
1855 {
1856    switch (op) {
1857    case nir_atomic_op_iadd: return LLVMAtomicRMWBinOpAdd;
1858    case nir_atomic_op_xchg: return LLVMAtomicRMWBinOpXchg;
1859    case nir_atomic_op_iand: return LLVMAtomicRMWBinOpAnd;
1860    case nir_atomic_op_ior:  return LLVMAtomicRMWBinOpOr;
1861    case nir_atomic_op_ixor: return LLVMAtomicRMWBinOpXor;
1862    case nir_atomic_op_umin: return LLVMAtomicRMWBinOpUMin;
1863    case nir_atomic_op_umax: return LLVMAtomicRMWBinOpUMax;
1864    case nir_atomic_op_imin: return LLVMAtomicRMWBinOpMin;
1865    case nir_atomic_op_imax: return LLVMAtomicRMWBinOpMax;
1866    case nir_atomic_op_fadd: return LLVMAtomicRMWBinOpFAdd;
1867    default: unreachable("Unexpected atomic");
1868    }
1869 }
1870 
visit_atomic_ssbo(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)1871 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
1872 {
1873    nir_atomic_op nir_op = nir_intrinsic_atomic_op(instr);
1874    const char *op = translate_atomic_op_str(nir_op);
1875    bool is_float = nir_atomic_op_type(nir_op) == nir_type_float;
1876 
1877    LLVMTypeRef return_type = LLVMTypeOf(get_src(ctx, instr->src[2]));
1878    char name[64], type[8];
1879    LLVMValueRef params[6], descriptor;
1880    LLVMValueRef result;
1881    int arg_count = 0;
1882 
1883    struct waterfall_context wctx;
1884    LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1885 
1886    descriptor = ctx->abi->load_ssbo ?
1887       ctx->abi->load_ssbo(ctx->abi, rsrc_base, true, false) : rsrc_base;
1888 
1889    if (instr->intrinsic == nir_intrinsic_ssbo_atomic_swap && return_type == ctx->ac.i64) {
1890       result = emit_ssbo_comp_swap_64(ctx, descriptor, get_src(ctx, instr->src[1]),
1891                                       get_src(ctx, instr->src[2]), get_src(ctx, instr->src[3]), false);
1892    } else {
1893       LLVMValueRef data = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
1894 
1895       if (instr->intrinsic == nir_intrinsic_ssbo_atomic_swap) {
1896          params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
1897       }
1898       if (is_float) {
1899          data = ac_to_float(&ctx->ac, data);
1900          return_type = LLVMTypeOf(data);
1901       }
1902 
1903       unsigned cache_flags =
1904          ac_get_hw_cache_flags(ctx->ac.info,
1905 			       ac_get_mem_access_flags(instr) | ACCESS_TYPE_ATOMIC).value;
1906 
1907       params[arg_count++] = data;
1908       params[arg_count++] = descriptor;
1909       params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1910       params[arg_count++] = ctx->ac.i32_0;               /* soffset */
1911       params[arg_count++] = LLVMConstInt(ctx->ac.i32, cache_flags, 0);
1912 
1913       ac_build_type_name_for_intr(return_type, type, sizeof(type));
1914       snprintf(name, sizeof(name), "llvm.amdgcn.raw.buffer.atomic.%s.%s", op, type);
1915 
1916       result = ac_build_intrinsic(&ctx->ac, name, return_type, params, arg_count, 0);
1917 
1918       if (is_float) {
1919          result = ac_to_integer(&ctx->ac, result);
1920       }
1921    }
1922 
1923    return exit_waterfall(ctx, &wctx, result);
1924 }
1925 
visit_load_buffer(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)1926 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
1927 {
1928    struct waterfall_context wctx;
1929    LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1930 
1931    int elem_size_bytes = instr->def.bit_size / 8;
1932    int num_components = instr->num_components;
1933    enum gl_access_qualifier access = ac_get_mem_access_flags(instr);
1934 
1935    LLVMValueRef offset = get_src(ctx, instr->src[1]);
1936    LLVMValueRef rsrc = ctx->abi->load_ssbo ?
1937       ctx->abi->load_ssbo(ctx->abi, rsrc_base, false, false) : rsrc_base;
1938    LLVMValueRef vindex = ctx->ac.i32_0;
1939 
1940    LLVMTypeRef def_type = get_def_type(ctx, &instr->def);
1941    LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
1942 
1943    LLVMValueRef results[4];
1944    for (int i = 0; i < num_components;) {
1945       int num_elems = num_components - i;
1946       if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
1947          num_elems = 1;
1948       if (num_elems * elem_size_bytes > 16)
1949          num_elems = 16 / elem_size_bytes;
1950       int load_bytes = num_elems * elem_size_bytes;
1951 
1952       LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
1953       LLVMValueRef voffset = LLVMBuildAdd(ctx->ac.builder, offset, immoffset, "");
1954 
1955       LLVMValueRef ret;
1956 
1957       if (load_bytes == 1) {
1958          ret = ac_build_buffer_load_byte(&ctx->ac, rsrc, voffset, ctx->ac.i32_0,
1959                                           access);
1960       } else if (load_bytes == 2) {
1961          ret = ac_build_buffer_load_short(&ctx->ac, rsrc, voffset, ctx->ac.i32_0,
1962                                            access);
1963       } else {
1964          int num_channels = util_next_power_of_two(load_bytes) / 4;
1965          bool can_speculate = access & ACCESS_CAN_REORDER;
1966 
1967          ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels, vindex, voffset, ctx->ac.i32_0,
1968                                     ctx->ac.f32, access, can_speculate, false);
1969       }
1970 
1971       LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
1972       ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
1973       ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
1974 
1975       LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
1976       ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
1977 
1978       for (unsigned j = 0; j < num_elems; j++) {
1979          results[i + j] =
1980             LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
1981       }
1982       i += num_elems;
1983    }
1984 
1985    LLVMValueRef ret = ac_build_gather_values(&ctx->ac, results, num_components);
1986    return exit_waterfall(ctx, &wctx, ret);
1987 }
1988 
enter_waterfall_ubo(struct ac_nir_context * ctx,struct waterfall_context * wctx,const nir_intrinsic_instr * instr)1989 static LLVMValueRef enter_waterfall_ubo(struct ac_nir_context *ctx, struct waterfall_context *wctx,
1990                                         const nir_intrinsic_instr *instr)
1991 {
1992    return enter_waterfall(ctx, wctx, get_src(ctx, instr->src[0]),
1993                           nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
1994 }
1995 
get_global_address(struct ac_nir_context * ctx,nir_intrinsic_instr * instr,LLVMTypeRef type)1996 static LLVMValueRef get_global_address(struct ac_nir_context *ctx,
1997                                        nir_intrinsic_instr *instr,
1998                                        LLVMTypeRef type)
1999 {
2000    bool is_store = instr->intrinsic == nir_intrinsic_store_global ||
2001                    instr->intrinsic == nir_intrinsic_store_global_amd;
2002    LLVMValueRef addr = get_src(ctx, instr->src[is_store ? 1 : 0]);
2003 
2004    LLVMTypeRef ptr_type = LLVMPointerType(type, AC_ADDR_SPACE_GLOBAL);
2005 
2006    if (nir_intrinsic_has_base(instr)) {
2007       /* _amd variants */
2008       uint32_t base = nir_intrinsic_base(instr);
2009       unsigned num_src = nir_intrinsic_infos[instr->intrinsic].num_srcs;
2010       LLVMValueRef offset = get_src(ctx, instr->src[num_src - 1]);
2011       offset = LLVMBuildAdd(ctx->ac.builder, offset, LLVMConstInt(ctx->ac.i32, base, false), "");
2012 
2013       LLVMTypeRef i8_ptr_type = LLVMPointerType(ctx->ac.i8, AC_ADDR_SPACE_GLOBAL);
2014       addr = LLVMBuildIntToPtr(ctx->ac.builder, addr, i8_ptr_type, "");
2015       addr = LLVMBuildGEP2(ctx->ac.builder, ctx->ac.i8, addr, &offset, 1, "");
2016       return LLVMBuildPointerCast(ctx->ac.builder, addr, ptr_type, "");
2017    } else {
2018       return LLVMBuildIntToPtr(ctx->ac.builder, addr, ptr_type, "");
2019    }
2020 }
2021 
visit_load_global(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)2022 static LLVMValueRef visit_load_global(struct ac_nir_context *ctx,
2023                                       nir_intrinsic_instr *instr)
2024 {
2025    LLVMTypeRef result_type = get_def_type(ctx, &instr->def);
2026    LLVMValueRef val;
2027    LLVMValueRef addr = get_global_address(ctx, instr, result_type);
2028 
2029    val = LLVMBuildLoad2(ctx->ac.builder, result_type, addr, "");
2030 
2031    if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE)) {
2032       LLVMSetOrdering(val, LLVMAtomicOrderingMonotonic);
2033       LLVMSetAlignment(val, ac_get_type_size(result_type));
2034    }
2035 
2036    return val;
2037 }
2038 
visit_store_global(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)2039 static void visit_store_global(struct ac_nir_context *ctx,
2040 				     nir_intrinsic_instr *instr)
2041 {
2042    LLVMValueRef data = get_src(ctx, instr->src[0]);
2043    LLVMTypeRef type = LLVMTypeOf(data);
2044    LLVMValueRef addr = get_global_address(ctx, instr, type);
2045    LLVMValueRef val;
2046 
2047    val = LLVMBuildStore(ctx->ac.builder, data, addr);
2048 
2049    if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE)) {
2050       LLVMSetOrdering(val, LLVMAtomicOrderingMonotonic);
2051       LLVMSetAlignment(val, ac_get_type_size(type));
2052    }
2053 }
2054 
visit_global_atomic(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)2055 static LLVMValueRef visit_global_atomic(struct ac_nir_context *ctx,
2056 					nir_intrinsic_instr *instr)
2057 {
2058    LLVMValueRef data = get_src(ctx, instr->src[1]);
2059    LLVMAtomicRMWBinOp op;
2060    LLVMValueRef result;
2061 
2062    /* use "singlethread" sync scope to implement relaxed ordering */
2063    const char *sync_scope = "singlethread-one-as";
2064 
2065    nir_atomic_op nir_op = nir_intrinsic_atomic_op(instr);
2066    bool is_float = nir_atomic_op_type(nir_op) == nir_type_float;
2067 
2068    LLVMTypeRef data_type = LLVMTypeOf(data);
2069 
2070    assert(instr->src[1].ssa->num_components == 1);
2071    if (is_float) {
2072       switch (instr->src[1].ssa->bit_size) {
2073       case 32:
2074          data_type = ctx->ac.f32;
2075          break;
2076       case 64:
2077          data_type = ctx->ac.f64;
2078          break;
2079       default:
2080          unreachable("Unsupported float bit size");
2081       }
2082 
2083       data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
2084    }
2085 
2086    LLVMValueRef addr = get_global_address(ctx, instr, data_type);
2087 
2088    if (instr->intrinsic == nir_intrinsic_global_atomic_swap ||
2089        instr->intrinsic == nir_intrinsic_global_atomic_swap_amd) {
2090       LLVMValueRef data1 = get_src(ctx, instr->src[2]);
2091       result = ac_build_atomic_cmp_xchg(&ctx->ac, addr, data, data1, sync_scope);
2092       result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
2093    } else if (is_float) {
2094       const char *op = translate_atomic_op_str(nir_op);
2095       char name[64], type[8];
2096       LLVMValueRef params[2];
2097       int arg_count = 0;
2098 
2099       params[arg_count++] = addr;
2100       params[arg_count++] = data;
2101 
2102       ac_build_type_name_for_intr(data_type, type, sizeof(type));
2103       snprintf(name, sizeof(name), "llvm.amdgcn.global.atomic.%s.%s.p1.%s", op, type, type);
2104 
2105       result = ac_build_intrinsic(&ctx->ac, name, data_type, params, arg_count, 0);
2106    } else {
2107       op = translate_atomic_op(nir_op);
2108       result = ac_build_atomic_rmw(&ctx->ac, op, addr, ac_to_integer(&ctx->ac, data), sync_scope);
2109    }
2110 
2111    result = ac_to_integer(&ctx->ac, result);
2112 
2113    return result;
2114 }
2115 
visit_load_ubo_buffer(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)2116 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
2117 {
2118    struct waterfall_context wctx;
2119    LLVMValueRef rsrc_base = enter_waterfall_ubo(ctx, &wctx, instr);
2120 
2121    LLVMValueRef ret;
2122    LLVMValueRef rsrc = rsrc_base;
2123    LLVMValueRef offset = get_src(ctx, instr->src[1]);
2124    int num_components = instr->num_components;
2125 
2126    assert(instr->def.bit_size >= 32 && instr->def.bit_size % 32 == 0);
2127 
2128    if (ctx->abi->load_ubo)
2129       rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
2130 
2131    /* Convert to a 32-bit load. */
2132    if (instr->def.bit_size == 64)
2133       num_components *= 2;
2134 
2135    ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset, NULL,
2136                               ctx->ac.f32, 0, true, true);
2137    ret = LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->def), "");
2138 
2139    return exit_waterfall(ctx, &wctx, ret);
2140 }
2141 
visit_store_output(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)2142 static void visit_store_output(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
2143 {
2144    unsigned base = nir_intrinsic_base(instr);
2145    unsigned writemask = nir_intrinsic_write_mask(instr);
2146    unsigned component = nir_intrinsic_component(instr);
2147    LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
2148    ASSERTED nir_src offset = *nir_get_io_offset_src(instr);
2149 
2150    /* No indirect indexing is allowed here. */
2151    assert(nir_src_is_const(offset) && nir_src_as_uint(offset) == 0);
2152 
2153    switch (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src))) {
2154    case 16:
2155    case 32:
2156       break;
2157    case 64:
2158       unreachable("64-bit IO should have been lowered to 32 bits");
2159       return;
2160    default:
2161       unreachable("unhandled store_output bit size");
2162       return;
2163    }
2164 
2165    writemask <<= component;
2166 
2167    for (unsigned chan = 0; chan < 8; chan++) {
2168       if (!(writemask & (1 << chan)))
2169          continue;
2170 
2171       LLVMValueRef value = ac_llvm_extract_elem(&ctx->ac, src, chan - component);
2172       LLVMValueRef output_addr = ctx->abi->outputs[base * 4 + chan];
2173 
2174       if (!ctx->abi->is_16bit[base * 4 + chan] &&
2175           LLVMTypeOf(value) == ctx->ac.f16) {
2176          LLVMValueRef output, index;
2177 
2178          /* Insert the 16-bit value into the low or high bits of the 32-bit output
2179           * using read-modify-write.
2180           */
2181          index = LLVMConstInt(ctx->ac.i32, nir_intrinsic_io_semantics(instr).high_16bits, 0);
2182 
2183          output = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.v2f16, output_addr, "");
2184          output = LLVMBuildInsertElement(ctx->ac.builder, output, value, index, "");
2185          value = LLVMBuildBitCast(ctx->ac.builder, output, ctx->ac.f32, "");
2186       }
2187       LLVMBuildStore(ctx->ac.builder, value, output_addr);
2188    }
2189 }
2190 
image_type_to_components_count(enum glsl_sampler_dim dim,bool array)2191 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2192 {
2193    switch (dim) {
2194    case GLSL_SAMPLER_DIM_BUF:
2195       return 1;
2196    case GLSL_SAMPLER_DIM_1D:
2197       return array ? 2 : 1;
2198    case GLSL_SAMPLER_DIM_2D:
2199       return array ? 3 : 2;
2200    case GLSL_SAMPLER_DIM_MS:
2201       return array ? 4 : 3;
2202    case GLSL_SAMPLER_DIM_3D:
2203    case GLSL_SAMPLER_DIM_CUBE:
2204       return 3;
2205    case GLSL_SAMPLER_DIM_RECT:
2206    case GLSL_SAMPLER_DIM_SUBPASS:
2207       return 2;
2208    case GLSL_SAMPLER_DIM_SUBPASS_MS:
2209       return 3;
2210    default:
2211       break;
2212    }
2213    return 0;
2214 }
2215 
get_image_coords(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr,LLVMValueRef dynamic_desc_index,struct ac_image_args * args,enum glsl_sampler_dim dim,bool is_array)2216 static void get_image_coords(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr,
2217                              LLVMValueRef dynamic_desc_index, struct ac_image_args *args,
2218                              enum glsl_sampler_dim dim, bool is_array)
2219 {
2220    LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2221    LLVMValueRef masks[] = {
2222       ctx->ac.i32_0,
2223       ctx->ac.i32_1,
2224       LLVMConstInt(ctx->ac.i32, 2, false),
2225       LLVMConstInt(ctx->ac.i32, 3, false),
2226    };
2227 
2228    int count;
2229    ASSERTED bool add_frag_pos =
2230       (dim == GLSL_SAMPLER_DIM_SUBPASS || dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2231    bool is_ms = (dim == GLSL_SAMPLER_DIM_MS || dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2232    bool gfx9_1d = ctx->ac.gfx_level == GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2233    assert(!add_frag_pos && "Input attachments should be lowered by this point.");
2234    count = image_type_to_components_count(dim, is_array);
2235 
2236    if (count == 1 && !gfx9_1d) {
2237       if (instr->src[1].ssa->num_components)
2238          args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2239       else
2240          args->coords[0] = src0;
2241    } else {
2242       int chan;
2243       if (is_ms)
2244          count--;
2245       for (chan = 0; chan < count; ++chan) {
2246          args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2247       }
2248 
2249       if (gfx9_1d) {
2250          if (is_array) {
2251             args->coords[2] = args->coords[1];
2252             args->coords[1] = ctx->ac.i32_0;
2253          } else
2254             args->coords[1] = ctx->ac.i32_0;
2255          count++;
2256       }
2257       if (ctx->ac.gfx_level == GFX9 && dim == GLSL_SAMPLER_DIM_2D && !is_array) {
2258          /* The hw can't bind a slice of a 3D image as a 2D
2259           * image, because it ignores BASE_ARRAY if the target
2260           * is 3D. The workaround is to read BASE_ARRAY and set
2261           * it as the 3rd address operand for all 2D images.
2262           */
2263          LLVMValueRef first_layer, const5, mask;
2264 
2265          const5 = LLVMConstInt(ctx->ac.i32, 5, 0);
2266          mask = LLVMConstInt(ctx->ac.i32, S_008F24_BASE_ARRAY(~0), 0);
2267          first_layer = LLVMBuildExtractElement(ctx->ac.builder, args->resource, const5, "");
2268          first_layer = LLVMBuildAnd(ctx->ac.builder, first_layer, mask, "");
2269 
2270          if (instr->intrinsic == nir_intrinsic_bindless_image_load ||
2271              instr->intrinsic == nir_intrinsic_bindless_image_sparse_load ||
2272              instr->intrinsic == nir_intrinsic_bindless_image_store) {
2273             int lod_index = instr->intrinsic == nir_intrinsic_bindless_image_store ? 4 : 3;
2274             bool has_lod = !nir_src_is_const(instr->src[lod_index]) ||
2275                            nir_src_as_uint(instr->src[lod_index]) != 0;
2276             if (has_lod) {
2277                /* If there's a lod parameter it matter if the image is 3d or 2d because
2278                 * the hw reads either the fourth or third component as lod. So detect
2279                 * 3d images and place the lod at the third component otherwise.
2280                 */
2281                LLVMValueRef const3, const28, const4, rword3, type3d, type, is_3d, lod;
2282                const3 = LLVMConstInt(ctx->ac.i32, 3, 0);
2283                const28 = LLVMConstInt(ctx->ac.i32, 28, 0);
2284                const4 = LLVMConstInt(ctx->ac.i32, 4, 0);
2285                type3d = LLVMConstInt(ctx->ac.i32, V_008F1C_SQ_RSRC_IMG_3D, 0);
2286                rword3 = LLVMBuildExtractElement(ctx->ac.builder, args->resource, const3, "");
2287                type = ac_build_bfe(&ctx->ac, rword3, const28, const4, false);
2288                is_3d = emit_int_cmp(&ctx->ac, LLVMIntEQ, type, type3d);
2289                lod = get_src(ctx, instr->src[lod_index]);
2290                first_layer = emit_bcsel(&ctx->ac, is_3d, first_layer, lod);
2291             }
2292          }
2293 
2294          args->coords[count] = first_layer;
2295          count++;
2296       }
2297 
2298       if (is_ms) {
2299          /* sample index */
2300          args->coords[count] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2301          count++;
2302       }
2303    }
2304 }
2305 
enter_waterfall_image(struct ac_nir_context * ctx,struct waterfall_context * wctx,const nir_intrinsic_instr * instr)2306 static LLVMValueRef enter_waterfall_image(struct ac_nir_context *ctx,
2307                                           struct waterfall_context *wctx,
2308                                           const nir_intrinsic_instr *instr)
2309 {
2310    /* src0 is desc when uniform, desc index when non uniform */
2311    LLVMValueRef value = get_src(ctx, instr->src[0]);
2312 
2313    return enter_waterfall(ctx, wctx, value, nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2314 }
2315 
visit_image_load(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2316 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2317 {
2318    LLVMValueRef res;
2319 
2320    enum glsl_sampler_dim dim = nir_intrinsic_image_dim(instr);
2321    enum gl_access_qualifier access = nir_intrinsic_access(instr);
2322    bool is_array = nir_intrinsic_image_array(instr);
2323 
2324    struct waterfall_context wctx;
2325    LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2326 
2327    struct ac_image_args args = {0};
2328 
2329    args.access = ac_get_mem_access_flags(instr);
2330    args.tfe = instr->intrinsic == nir_intrinsic_bindless_image_sparse_load;
2331 
2332    if (dim == GLSL_SAMPLER_DIM_BUF) {
2333       unsigned num_channels = util_last_bit(nir_def_components_read(&instr->def));
2334       if (instr->def.bit_size == 64)
2335          num_channels = num_channels < 4 ? 2 : 4;
2336       LLVMValueRef rsrc, vindex;
2337 
2338       rsrc = ctx->abi->load_sampler_desc(ctx->abi, dynamic_index, AC_DESC_BUFFER);
2339       vindex =
2340          LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]), ctx->ac.i32_0, "");
2341 
2342       bool can_speculate = access & ACCESS_CAN_REORDER;
2343       res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex, ctx->ac.i32_0, num_channels,
2344                                         args.access, can_speculate,
2345                                         instr->def.bit_size == 16,
2346                                         args.tfe);
2347       res = ac_build_expand(&ctx->ac, res, num_channels, args.tfe ? 5 : 4);
2348 
2349       res = ac_trim_vector(&ctx->ac, res, instr->def.num_components);
2350       res = ac_to_integer(&ctx->ac, res);
2351    } else if (instr->intrinsic == nir_intrinsic_bindless_image_fragment_mask_load_amd) {
2352       assert(ctx->ac.gfx_level < GFX11);
2353 
2354       args.opcode = ac_image_load;
2355       args.resource = ctx->abi->load_sampler_desc(ctx->abi, dynamic_index, AC_DESC_FMASK);
2356       get_image_coords(ctx, instr, dynamic_index, &args, GLSL_SAMPLER_DIM_2D, is_array);
2357       args.dmask = 0xf;
2358       args.dim = is_array ? ac_image_2darray : ac_image_2d;
2359       args.attributes = AC_ATTR_INVARIANT_LOAD;
2360       args.a16 = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.coords[0])) == 16;
2361 
2362       res = ac_build_image_opcode(&ctx->ac, &args);
2363    } else {
2364       bool level_zero = nir_src_is_const(instr->src[3]) && nir_src_as_uint(instr->src[3]) == 0;
2365 
2366       args.opcode = level_zero ? ac_image_load : ac_image_load_mip;
2367       args.resource = ctx->abi->load_sampler_desc(ctx->abi, dynamic_index, AC_DESC_IMAGE);
2368       get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2369       args.dim = ac_get_image_dim(ctx->ac.gfx_level, dim, is_array);
2370       if (!level_zero)
2371          args.lod = get_src(ctx, instr->src[3]);
2372       args.dmask = 15;
2373       args.attributes = access & ACCESS_CAN_REORDER ? AC_ATTR_INVARIANT_LOAD : 0;
2374 
2375       args.d16 = instr->def.bit_size == 16;
2376 
2377       res = ac_build_image_opcode(&ctx->ac, &args);
2378    }
2379 
2380    if (instr->def.bit_size == 64) {
2381       LLVMValueRef code = NULL;
2382       if (args.tfe) {
2383          code = ac_llvm_extract_elem(&ctx->ac, res, 4);
2384          res = ac_trim_vector(&ctx->ac, res, 4);
2385       }
2386 
2387       res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i64, 2), "");
2388       LLVMValueRef x = LLVMBuildExtractElement(ctx->ac.builder, res, ctx->ac.i32_0, "");
2389       LLVMValueRef w = LLVMBuildExtractElement(ctx->ac.builder, res, ctx->ac.i32_1, "");
2390 
2391       if (code)
2392          code = LLVMBuildZExt(ctx->ac.builder, code, ctx->ac.i64, "");
2393       LLVMValueRef values[5] = {x, ctx->ac.i64_0, ctx->ac.i64_0, w, code};
2394       res = ac_build_gather_values(&ctx->ac, values, 4 + args.tfe);
2395    }
2396 
2397    return exit_waterfall(ctx, &wctx, res);
2398 }
2399 
visit_image_store(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2400 static void visit_image_store(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2401 {
2402    enum glsl_sampler_dim dim = nir_intrinsic_image_dim(instr);
2403    bool is_array = nir_intrinsic_image_array(instr);
2404 
2405    struct waterfall_context wctx;
2406    LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2407 
2408    struct ac_image_args args = {0};
2409    args.access = ac_get_mem_access_flags(instr);
2410 
2411    LLVMValueRef src = get_src(ctx, instr->src[3]);
2412    if (instr->src[3].ssa->bit_size == 64) {
2413       /* only R64_UINT and R64_SINT supported */
2414       src = ac_llvm_extract_elem(&ctx->ac, src, 0);
2415       src = LLVMBuildBitCast(ctx->ac.builder, src, ctx->ac.v2f32, "");
2416    } else {
2417       src = ac_to_float(&ctx->ac, src);
2418    }
2419 
2420    if (dim == GLSL_SAMPLER_DIM_BUF) {
2421       LLVMValueRef rsrc = ctx->abi->load_sampler_desc(ctx->abi, dynamic_index, AC_DESC_BUFFER);
2422       unsigned src_channels = ac_get_llvm_num_components(src);
2423       LLVMValueRef vindex;
2424 
2425       if (src_channels == 3)
2426          src = ac_build_expand_to_vec4(&ctx->ac, src, 3);
2427 
2428       vindex =
2429          LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]), ctx->ac.i32_0, "");
2430 
2431       ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex, ctx->ac.i32_0, args.access);
2432    } else {
2433       bool level_zero = nir_src_is_const(instr->src[4]) && nir_src_as_uint(instr->src[4]) == 0;
2434 
2435       args.opcode = level_zero ? ac_image_store : ac_image_store_mip;
2436       args.data[0] = src;
2437       args.resource = ctx->abi->load_sampler_desc(ctx->abi, dynamic_index, AC_DESC_IMAGE);
2438       get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2439       args.dim = ac_get_image_dim(ctx->ac.gfx_level, dim, is_array);
2440       if (!level_zero)
2441          args.lod = get_src(ctx, instr->src[4]);
2442       args.dmask = 15;
2443       args.d16 = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.data[0])) == 16;
2444 
2445       ac_build_image_opcode(&ctx->ac, &args);
2446    }
2447 
2448    exit_waterfall(ctx, &wctx, NULL);
2449 }
2450 
visit_image_atomic(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2451 static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2452 {
2453    LLVMValueRef params[7];
2454    int param_count = 0;
2455 
2456    nir_atomic_op op = nir_intrinsic_atomic_op(instr);
2457    bool cmpswap = op == nir_atomic_op_cmpxchg;
2458    const char *atomic_name = translate_atomic_op_str(op);
2459    char intrinsic_name[64];
2460    enum ac_atomic_op atomic_subop;
2461    ASSERTED int length;
2462 
2463    enum glsl_sampler_dim dim = nir_intrinsic_image_dim(instr);
2464    bool is_array = nir_intrinsic_image_array(instr);
2465 
2466    struct waterfall_context wctx;
2467    LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2468 
2469    switch (op) {
2470    case nir_atomic_op_iadd:
2471       atomic_subop = ac_atomic_add;
2472       break;
2473    case nir_atomic_op_imin:
2474       atomic_subop = ac_atomic_smin;
2475       break;
2476    case nir_atomic_op_umin:
2477       atomic_subop = ac_atomic_umin;
2478       break;
2479    case nir_atomic_op_imax:
2480       atomic_subop = ac_atomic_smax;
2481       break;
2482    case nir_atomic_op_umax:
2483       atomic_subop = ac_atomic_umax;
2484       break;
2485    case nir_atomic_op_iand:
2486       atomic_subop = ac_atomic_and;
2487       break;
2488    case nir_atomic_op_ior:
2489       atomic_subop = ac_atomic_or;
2490       break;
2491    case nir_atomic_op_ixor:
2492       atomic_subop = ac_atomic_xor;
2493       break;
2494    case nir_atomic_op_xchg:
2495       atomic_subop = ac_atomic_swap;
2496       break;
2497    case nir_atomic_op_cmpxchg:
2498       atomic_subop = 0; /* not used */
2499       break;
2500    case nir_atomic_op_inc_wrap:
2501       atomic_subop = ac_atomic_inc_wrap;
2502       break;
2503    case nir_atomic_op_dec_wrap:
2504       atomic_subop = ac_atomic_dec_wrap;
2505       break;
2506    case nir_atomic_op_fadd:
2507       atomic_subop = ac_atomic_fmin; /* Non-buffer fadd atomics are not supported. */
2508       break;
2509    case nir_atomic_op_fmin:
2510       atomic_subop = ac_atomic_fmin;
2511       break;
2512    case nir_atomic_op_fmax:
2513       atomic_subop = ac_atomic_fmax;
2514       break;
2515    default:
2516       abort();
2517    }
2518 
2519    if (cmpswap)
2520       params[param_count++] = get_src(ctx, instr->src[4]);
2521    params[param_count++] = get_src(ctx, instr->src[3]);
2522 
2523    if (atomic_subop == ac_atomic_fmin || atomic_subop == ac_atomic_fmax)
2524       params[0] = ac_to_float(&ctx->ac, params[0]);
2525 
2526    LLVMValueRef result;
2527    if (dim == GLSL_SAMPLER_DIM_BUF) {
2528       params[param_count++] = ctx->abi->load_sampler_desc(ctx->abi, dynamic_index, AC_DESC_BUFFER);
2529       params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2530                                                       ctx->ac.i32_0, ""); /* vindex */
2531       params[param_count++] = ctx->ac.i32_0;                              /* voffset */
2532       if (cmpswap && instr->def.bit_size == 64) {
2533          result = emit_ssbo_comp_swap_64(ctx, params[2], params[3], params[1], params[0], true);
2534       } else {
2535          LLVMTypeRef data_type = LLVMTypeOf(params[0]);
2536          char type[8];
2537          unsigned cache_flags =
2538             ac_get_hw_cache_flags(ctx->ac.info,
2539 				  ac_get_mem_access_flags(instr) | ACCESS_TYPE_ATOMIC).value;
2540 
2541          params[param_count++] = ctx->ac.i32_0; /* soffset */
2542          params[param_count++] = LLVMConstInt(ctx->ac.i32, cache_flags, 0);
2543 
2544          ac_build_type_name_for_intr(data_type, type, sizeof(type));
2545          length = snprintf(intrinsic_name, sizeof(intrinsic_name),
2546                            "llvm.amdgcn.struct.buffer.atomic.%s.%s",
2547                            atomic_name, type);
2548 
2549          assert(length < sizeof(intrinsic_name));
2550          result = ac_build_intrinsic(&ctx->ac, intrinsic_name, LLVMTypeOf(params[0]), params, param_count, 0);
2551       }
2552    } else {
2553       struct ac_image_args args = {0};
2554       args.opcode = cmpswap ? ac_image_atomic_cmpswap : ac_image_atomic;
2555       args.atomic = atomic_subop;
2556       args.data[0] = params[0];
2557       if (cmpswap)
2558          args.data[1] = params[1];
2559       args.resource = ctx->abi->load_sampler_desc(ctx->abi, dynamic_index, AC_DESC_IMAGE);
2560       get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2561       args.dim = ac_get_image_dim(ctx->ac.gfx_level, dim, is_array);
2562       args.access = ac_get_mem_access_flags(instr);
2563 
2564       result = ac_build_image_opcode(&ctx->ac, &args);
2565    }
2566 
2567    return exit_waterfall(ctx, &wctx, result);
2568 }
2569 
emit_discard(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2570 static void emit_discard(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2571 {
2572    LLVMValueRef cond;
2573 
2574    if (instr->intrinsic == nir_intrinsic_discard_if ||
2575        instr->intrinsic == nir_intrinsic_terminate_if) {
2576       cond = LLVMBuildNot(ctx->ac.builder, get_src(ctx, instr->src[0]), "");
2577    } else {
2578       assert(instr->intrinsic == nir_intrinsic_discard ||
2579              instr->intrinsic == nir_intrinsic_terminate);
2580       cond = ctx->ac.i1false;
2581    }
2582 
2583    ac_build_kill_if_false(&ctx->ac, cond);
2584 }
2585 
emit_demote(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2586 static void emit_demote(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2587 {
2588    LLVMValueRef cond;
2589 
2590    if (instr->intrinsic == nir_intrinsic_demote_if) {
2591       cond = LLVMBuildNot(ctx->ac.builder, get_src(ctx, instr->src[0]), "");
2592    } else {
2593       assert(instr->intrinsic == nir_intrinsic_demote);
2594       cond = ctx->ac.i1false;
2595    }
2596 
2597    /* This demotes the pixel if the condition is false. */
2598    ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.wqm.demote", ctx->ac.voidt, &cond, 1, 0);
2599 }
2600 
visit_load_subgroup_id(struct ac_nir_context * ctx)2601 static LLVMValueRef visit_load_subgroup_id(struct ac_nir_context *ctx)
2602 {
2603    if (gl_shader_stage_is_compute(ctx->stage)) {
2604       if (ctx->ac.gfx_level >= GFX10_3)
2605          return ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->tg_size), 20, 5);
2606       else
2607          return ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->tg_size), 6, 6);
2608    } else if (ctx->args->tcs_wave_id.used) {
2609       return ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->tcs_wave_id), 0, 3);
2610    } else if (ctx->args->merged_wave_info.used) {
2611       return ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 24, 4);
2612    } else {
2613       return ctx->ac.i32_0;
2614    }
2615 }
2616 
visit_load_local_invocation_index(struct ac_nir_context * ctx)2617 static LLVMValueRef visit_load_local_invocation_index(struct ac_nir_context *ctx)
2618 {
2619    if (ctx->abi->vs_rel_patch_id)
2620       return ctx->abi->vs_rel_patch_id;
2621 
2622    return ac_build_imad(&ctx->ac, visit_load_subgroup_id(ctx),
2623                         LLVMConstInt(ctx->ac.i32, ctx->ac.wave_size, 0),
2624                         ac_get_thread_id(&ctx->ac));
2625 }
2626 
visit_first_invocation(struct ac_nir_context * ctx)2627 static LLVMValueRef visit_first_invocation(struct ac_nir_context *ctx)
2628 {
2629    LLVMValueRef active_set = ac_build_ballot(&ctx->ac, ctx->ac.i32_1);
2630    const char *intr = ctx->ac.wave_size == 32 ? "llvm.cttz.i32" : "llvm.cttz.i64";
2631 
2632    /* The second argument is whether cttz(0) should be defined, but we do not care. */
2633    LLVMValueRef args[] = {active_set, ctx->ac.i1false};
2634    LLVMValueRef result = ac_build_intrinsic(&ctx->ac, intr, ctx->ac.iN_wavemask, args, 2, 0);
2635 
2636    return LLVMBuildTrunc(ctx->ac.builder, result, ctx->ac.i32, "");
2637 }
2638 
visit_load_shared(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2639 static LLVMValueRef visit_load_shared(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2640 {
2641    LLVMValueRef values[16], derived_ptr, index, ret;
2642    unsigned const_off = nir_intrinsic_base(instr);
2643 
2644    LLVMTypeRef elem_type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
2645    LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0], const_off);
2646 
2647    for (int chan = 0; chan < instr->num_components; chan++) {
2648       index = LLVMConstInt(ctx->ac.i32, chan, 0);
2649       derived_ptr = LLVMBuildGEP2(ctx->ac.builder, elem_type, ptr, &index, 1, "");
2650       values[chan] = LLVMBuildLoad2(ctx->ac.builder, elem_type, derived_ptr, "");
2651    }
2652 
2653    ret = ac_build_gather_values(&ctx->ac, values, instr->num_components);
2654 
2655    return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->def), "");
2656 }
2657 
visit_store_shared(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2658 static void visit_store_shared(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2659 {
2660    LLVMValueRef derived_ptr, data, index;
2661    LLVMBuilderRef builder = ctx->ac.builder;
2662 
2663    unsigned const_off = nir_intrinsic_base(instr);
2664    LLVMTypeRef elem_type = LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
2665    LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[1], const_off);
2666    LLVMValueRef src = get_src(ctx, instr->src[0]);
2667 
2668    int writemask = nir_intrinsic_write_mask(instr);
2669    for (int chan = 0; chan < 16; chan++) {
2670       if (!(writemask & (1 << chan))) {
2671          continue;
2672       }
2673       data = ac_llvm_extract_elem(&ctx->ac, src, chan);
2674       index = LLVMConstInt(ctx->ac.i32, chan, 0);
2675       derived_ptr = LLVMBuildGEP2(builder, elem_type, ptr, &index, 1, "");
2676       LLVMBuildStore(builder, data, derived_ptr);
2677    }
2678 }
2679 
visit_load_shared2_amd(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2680 static LLVMValueRef visit_load_shared2_amd(struct ac_nir_context *ctx,
2681                                            const nir_intrinsic_instr *instr)
2682 {
2683    LLVMTypeRef pointee_type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
2684    LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0], 0);
2685 
2686    LLVMValueRef values[2];
2687    uint8_t offsets[] = {nir_intrinsic_offset0(instr), nir_intrinsic_offset1(instr)};
2688    unsigned stride = nir_intrinsic_st64(instr) ? 64 : 1;
2689    for (unsigned i = 0; i < 2; i++) {
2690       LLVMValueRef index = LLVMConstInt(ctx->ac.i32, offsets[i] * stride, 0);
2691       LLVMValueRef derived_ptr = LLVMBuildGEP2(ctx->ac.builder, pointee_type, ptr, &index, 1, "");
2692       values[i] = LLVMBuildLoad2(ctx->ac.builder, pointee_type, derived_ptr, "");
2693    }
2694 
2695    LLVMValueRef ret = ac_build_gather_values(&ctx->ac, values, 2);
2696    return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->def), "");
2697 }
2698 
visit_store_shared2_amd(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr)2699 static void visit_store_shared2_amd(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr)
2700 {
2701    LLVMTypeRef pointee_type = LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
2702    LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[1], 0);
2703    LLVMValueRef src = get_src(ctx, instr->src[0]);
2704 
2705    uint8_t offsets[] = {nir_intrinsic_offset0(instr), nir_intrinsic_offset1(instr)};
2706    unsigned stride = nir_intrinsic_st64(instr) ? 64 : 1;
2707    for (unsigned i = 0; i < 2; i++) {
2708       LLVMValueRef index = LLVMConstInt(ctx->ac.i32, offsets[i] * stride, 0);
2709       LLVMValueRef derived_ptr = LLVMBuildGEP2(ctx->ac.builder, pointee_type, ptr, &index, 1, "");
2710       LLVMBuildStore(ctx->ac.builder, ac_llvm_extract_elem(&ctx->ac, src, i), derived_ptr);
2711    }
2712 }
2713 
visit_var_atomic(struct ac_nir_context * ctx,const nir_intrinsic_instr * instr,LLVMValueRef ptr,int src_idx)2714 static LLVMValueRef visit_var_atomic(struct ac_nir_context *ctx, const nir_intrinsic_instr *instr,
2715                                      LLVMValueRef ptr, int src_idx)
2716 {
2717    LLVMValueRef result;
2718    LLVMValueRef src = get_src(ctx, instr->src[src_idx]);
2719    nir_atomic_op nir_op = nir_intrinsic_atomic_op(instr);
2720 
2721    const char *sync_scope = "workgroup-one-as";
2722 
2723    if (nir_op == nir_atomic_op_cmpxchg) {
2724       LLVMValueRef src1 = get_src(ctx, instr->src[src_idx + 1]);
2725       result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, src, src1, sync_scope);
2726       result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
2727    } else if (nir_op == nir_atomic_op_fmin || nir_op == nir_atomic_op_fmax) {
2728       const char *op = translate_atomic_op_str(nir_op);
2729       char name[64], type[8];
2730       LLVMValueRef params[5];
2731       LLVMTypeRef src_type;
2732       int arg_count = 0;
2733 
2734       src = ac_to_float(&ctx->ac, src);
2735       src_type = LLVMTypeOf(src);
2736 
2737       params[arg_count++] = ptr;
2738       params[arg_count++] = src;
2739       params[arg_count++] = ctx->ac.i32_0;
2740       params[arg_count++] = ctx->ac.i32_0;
2741       params[arg_count++] = ctx->ac.i1false;
2742 
2743       ac_build_type_name_for_intr(src_type, type, sizeof(type));
2744       snprintf(name, sizeof(name), "llvm.amdgcn.ds.%s.%s", op, type);
2745 
2746       result = ac_build_intrinsic(&ctx->ac, name, src_type, params, arg_count, 0);
2747       result = ac_to_integer(&ctx->ac, result);
2748    } else {
2749       LLVMAtomicRMWBinOp op = translate_atomic_op(nir_op);
2750       LLVMValueRef val;
2751 
2752       if (nir_op == nir_atomic_op_fadd) {
2753          val = ac_to_float(&ctx->ac, src);
2754       } else {
2755          val = ac_to_integer(&ctx->ac, src);
2756       }
2757 
2758       result = ac_build_atomic_rmw(&ctx->ac, op, ptr, val, sync_scope);
2759 
2760       if (nir_op == nir_atomic_op_fadd) {
2761          result = ac_to_integer(&ctx->ac, result);
2762       }
2763    }
2764 
2765    return result;
2766 }
2767 
load_sample_pos(struct ac_nir_context * ctx)2768 static LLVMValueRef load_sample_pos(struct ac_nir_context *ctx)
2769 {
2770    LLVMValueRef values[2];
2771    LLVMValueRef pos[2];
2772 
2773    pos[0] = ac_to_float(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]));
2774    pos[1] = ac_to_float(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]));
2775 
2776    values[0] = ac_build_fract(&ctx->ac, pos[0], 32);
2777    values[1] = ac_build_fract(&ctx->ac, pos[1], 32);
2778    return ac_build_gather_values(&ctx->ac, values, 2);
2779 }
2780 
lookup_interp_param(struct ac_nir_context * ctx,enum glsl_interp_mode interp,unsigned location)2781 static LLVMValueRef lookup_interp_param(struct ac_nir_context *ctx, enum glsl_interp_mode interp,
2782                                         unsigned location)
2783 {
2784    switch (interp) {
2785    case INTERP_MODE_FLAT:
2786    default:
2787       return NULL;
2788    case INTERP_MODE_SMOOTH:
2789    case INTERP_MODE_NONE:
2790       if (location == INTERP_CENTER)
2791          return ac_get_arg(&ctx->ac, ctx->args->persp_center);
2792       else if (location == INTERP_CENTROID)
2793          return ac_get_arg(&ctx->ac, ctx->args->persp_centroid);
2794       else if (location == INTERP_SAMPLE)
2795          return ac_get_arg(&ctx->ac, ctx->args->persp_sample);
2796       break;
2797    case INTERP_MODE_NOPERSPECTIVE:
2798       if (location == INTERP_CENTER)
2799          return ac_get_arg(&ctx->ac, ctx->args->linear_center);
2800       else if (location == INTERP_CENTROID)
2801          return ac_get_arg(&ctx->ac, ctx->args->linear_centroid);
2802       else if (location == INTERP_SAMPLE)
2803          return ac_get_arg(&ctx->ac, ctx->args->linear_sample);
2804       break;
2805    }
2806    return NULL;
2807 }
2808 
barycentric_center(struct ac_nir_context * ctx,unsigned mode)2809 static LLVMValueRef barycentric_center(struct ac_nir_context *ctx, unsigned mode)
2810 {
2811    LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTER);
2812    return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
2813 }
2814 
barycentric_offset(struct ac_nir_context * ctx,unsigned mode,LLVMValueRef offset)2815 static LLVMValueRef barycentric_offset(struct ac_nir_context *ctx, unsigned mode,
2816                                        LLVMValueRef offset)
2817 {
2818    LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTER);
2819    LLVMValueRef src_c0 =
2820       ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_0, ""));
2821    LLVMValueRef src_c1 =
2822       ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_1, ""));
2823 
2824    LLVMValueRef ij_out[2];
2825    LLVMValueRef ddxy_out = ac_build_ddxy_interp(&ctx->ac, interp_param);
2826 
2827    /*
2828     * take the I then J parameters, and the DDX/Y for it, and
2829     * calculate the IJ inputs for the interpolator.
2830     * temp1 = ddx * offset/sample.x + I;
2831     * interp_param.I = ddy * offset/sample.y + temp1;
2832     * temp1 = ddx * offset/sample.x + J;
2833     * interp_param.J = ddy * offset/sample.y + temp1;
2834     */
2835    for (unsigned i = 0; i < 2; i++) {
2836       LLVMValueRef ix_ll = LLVMConstInt(ctx->ac.i32, i, false);
2837       LLVMValueRef iy_ll = LLVMConstInt(ctx->ac.i32, i + 2, false);
2838       LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder, ddxy_out, ix_ll, "");
2839       LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder, ddxy_out, iy_ll, "");
2840       LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder, interp_param, ix_ll, "");
2841       LLVMValueRef temp1, temp2;
2842 
2843       interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el, ctx->ac.f32, "");
2844 
2845       temp1 = ac_build_fmad(&ctx->ac, ddx_el, src_c0, interp_el);
2846       temp2 = ac_build_fmad(&ctx->ac, ddy_el, src_c1, temp1);
2847 
2848       ij_out[i] = LLVMBuildBitCast(ctx->ac.builder, temp2, ctx->ac.i32, "");
2849    }
2850    interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
2851    return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
2852 }
2853 
barycentric_centroid(struct ac_nir_context * ctx,unsigned mode)2854 static LLVMValueRef barycentric_centroid(struct ac_nir_context *ctx, unsigned mode)
2855 {
2856    LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTROID);
2857    return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
2858 }
2859 
barycentric_sample(struct ac_nir_context * ctx,unsigned mode)2860 static LLVMValueRef barycentric_sample(struct ac_nir_context *ctx, unsigned mode)
2861 {
2862    LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_SAMPLE);
2863    return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
2864 }
2865 
barycentric_model(struct ac_nir_context * ctx)2866 static LLVMValueRef barycentric_model(struct ac_nir_context *ctx)
2867 {
2868    return LLVMBuildBitCast(ctx->ac.builder, ac_get_arg(&ctx->ac, ctx->args->pull_model),
2869                            ctx->ac.v3i32, "");
2870 }
2871 
load_interpolated_input(struct ac_nir_context * ctx,LLVMValueRef interp_param,unsigned index,unsigned comp_start,unsigned num_components,unsigned bitsize,bool high_16bits)2872 static LLVMValueRef load_interpolated_input(struct ac_nir_context *ctx, LLVMValueRef interp_param,
2873                                             unsigned index, unsigned comp_start,
2874                                             unsigned num_components, unsigned bitsize,
2875                                             bool high_16bits)
2876 {
2877    LLVMValueRef attr_number = LLVMConstInt(ctx->ac.i32, index, false);
2878    LLVMValueRef interp_param_f;
2879 
2880    interp_param_f = LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2f32, "");
2881    LLVMValueRef i = LLVMBuildExtractElement(ctx->ac.builder, interp_param_f, ctx->ac.i32_0, "");
2882    LLVMValueRef j = LLVMBuildExtractElement(ctx->ac.builder, interp_param_f, ctx->ac.i32_1, "");
2883 
2884    /* Workaround for issue 2647: kill threads with infinite interpolation coeffs */
2885    if (ctx->verified_interp && !_mesa_hash_table_search(ctx->verified_interp, interp_param)) {
2886       LLVMValueRef cond = ac_build_is_inf_or_nan(&ctx->ac, i);
2887       ac_build_kill_if_false(&ctx->ac, LLVMBuildNot(ctx->ac.builder, cond, ""));
2888       _mesa_hash_table_insert(ctx->verified_interp, interp_param, interp_param);
2889    }
2890 
2891    LLVMValueRef values[4];
2892    assert(bitsize == 16 || bitsize == 32);
2893    for (unsigned comp = 0; comp < num_components; comp++) {
2894       LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, comp_start + comp, false);
2895       if (bitsize == 16) {
2896          values[comp] = ac_build_fs_interp_f16(&ctx->ac, llvm_chan, attr_number,
2897                                                ac_get_arg(&ctx->ac, ctx->args->prim_mask), i, j,
2898                                                high_16bits);
2899       } else {
2900          values[comp] = ac_build_fs_interp(&ctx->ac, llvm_chan, attr_number,
2901                                            ac_get_arg(&ctx->ac, ctx->args->prim_mask), i, j);
2902       }
2903    }
2904 
2905    return ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, num_components));
2906 }
2907 
visit_load(struct ac_nir_context * ctx,nir_intrinsic_instr * instr,bool is_output)2908 static LLVMValueRef visit_load(struct ac_nir_context *ctx, nir_intrinsic_instr *instr,
2909                                bool is_output)
2910 {
2911    LLVMValueRef values[8];
2912    LLVMTypeRef dest_type = get_def_type(ctx, &instr->def);
2913    LLVMTypeRef component_type;
2914    unsigned base = nir_intrinsic_base(instr);
2915    unsigned component = nir_intrinsic_component(instr);
2916    unsigned count = instr->def.num_components;
2917    nir_src *vertex_index_src = nir_get_io_arrayed_index_src(instr);
2918    LLVMValueRef vertex_index = vertex_index_src ? get_src(ctx, *vertex_index_src) : NULL;
2919    nir_src offset = *nir_get_io_offset_src(instr);
2920    LLVMValueRef indir_index = NULL;
2921 
2922    switch (instr->def.bit_size) {
2923    case 16:
2924    case 32:
2925       break;
2926    case 64:
2927       if (ctx->stage != MESA_SHADER_VERTEX || is_output) {
2928          unreachable("64-bit IO should have been lowered");
2929          return NULL;
2930       }
2931       break;
2932    default:
2933       unreachable("unhandled load type");
2934       return NULL;
2935    }
2936 
2937    if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
2938       component_type = LLVMGetElementType(dest_type);
2939    else
2940       component_type = dest_type;
2941 
2942    if (nir_src_is_const(offset))
2943       assert(nir_src_as_uint(offset) == 0);
2944    else
2945       indir_index = get_src(ctx, offset);
2946 
2947    if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2948       LLVMValueRef result = ctx->abi->load_tess_varyings(ctx->abi, component_type,
2949                                                          vertex_index, indir_index,
2950                                                          base, component,
2951                                                          count, !is_output);
2952       if (instr->def.bit_size == 16) {
2953          result = ac_to_integer(&ctx->ac, result);
2954          result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
2955       }
2956       return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
2957    }
2958 
2959    /* No indirect indexing is allowed after this point. */
2960    assert(!indir_index);
2961 
2962    /* Other non-fragment cases have outputs in temporaries. */
2963    if (is_output && (ctx->stage == MESA_SHADER_VERTEX || ctx->stage == MESA_SHADER_TESS_EVAL)) {
2964       assert(is_output);
2965 
2966       for (unsigned chan = component; chan < count + component; chan++)
2967          values[chan] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32,
2968                                        ctx->abi->outputs[base * 4 + chan], "");
2969 
2970       LLVMValueRef result = ac_build_varying_gather_values(&ctx->ac, values, count, component);
2971       return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
2972    }
2973 
2974    /* Fragment shader inputs. */
2975    assert(ctx->stage == MESA_SHADER_FRAGMENT);
2976    unsigned vertex_id = 0; /* P0 */
2977 
2978    if (instr->intrinsic == nir_intrinsic_load_input_vertex)
2979       vertex_id = nir_src_as_uint(instr->src[0]);
2980 
2981    LLVMValueRef attr_number = LLVMConstInt(ctx->ac.i32, base, false);
2982 
2983    for (unsigned chan = 0; chan < count; chan++) {
2984       LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, (component + chan) % 4, false);
2985       values[chan] = ac_build_fs_interp_mov(&ctx->ac, vertex_id, llvm_chan, attr_number,
2986                                             ac_get_arg(&ctx->ac, ctx->args->prim_mask));
2987       values[chan] = LLVMBuildBitCast(ctx->ac.builder, values[chan], ctx->ac.i32, "");
2988       if (instr->def.bit_size == 16 &&
2989           nir_intrinsic_io_semantics(instr).high_16bits)
2990          values[chan] = LLVMBuildLShr(ctx->ac.builder, values[chan], LLVMConstInt(ctx->ac.i32, 16, 0), "");
2991       values[chan] =
2992          LLVMBuildTruncOrBitCast(ctx->ac.builder, values[chan],
2993                                  instr->def.bit_size == 16 ? ctx->ac.i16 : ctx->ac.i32, "");
2994    }
2995 
2996    LLVMValueRef result = ac_build_gather_values(&ctx->ac, values, count);
2997    return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
2998 }
2999 
3000 static LLVMValueRef
emit_load_frag_shading_rate(struct ac_nir_context * ctx)3001 emit_load_frag_shading_rate(struct ac_nir_context *ctx)
3002 {
3003    LLVMValueRef x_rate, y_rate, cond;
3004 
3005    /* VRS Rate X = Ancillary[2:3]
3006     * VRS Rate Y = Ancillary[4:5]
3007     */
3008    x_rate = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ancillary), 2, 2);
3009    y_rate = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ancillary), 4, 2);
3010 
3011    /* xRate = xRate == 0x1 ? Horizontal2Pixels : None. */
3012    cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, x_rate, ctx->ac.i32_1, "");
3013    x_rate = LLVMBuildSelect(ctx->ac.builder, cond,
3014                             LLVMConstInt(ctx->ac.i32, 4, false), ctx->ac.i32_0, "");
3015 
3016    /* yRate = yRate == 0x1 ? Vertical2Pixels : None. */
3017    cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, y_rate, ctx->ac.i32_1, "");
3018    y_rate = LLVMBuildSelect(ctx->ac.builder, cond,
3019                             ctx->ac.i32_1, ctx->ac.i32_0, "");
3020 
3021    return LLVMBuildOr(ctx->ac.builder, x_rate, y_rate, "");
3022 }
3023 
3024 static LLVMValueRef
emit_load_frag_coord(struct ac_nir_context * ctx)3025 emit_load_frag_coord(struct ac_nir_context *ctx)
3026 {
3027    LLVMValueRef values[4] = {
3028       ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]), ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]),
3029       ac_get_arg(&ctx->ac, ctx->args->frag_pos[2]),
3030       ac_build_fdiv(&ctx->ac, ctx->ac.f32_1, ac_get_arg(&ctx->ac, ctx->args->frag_pos[3]))};
3031 
3032    return ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, 4));
3033 }
3034 
visit_intrinsic(struct ac_nir_context * ctx,nir_intrinsic_instr * instr)3035 static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *instr)
3036 {
3037    LLVMValueRef result = NULL;
3038 
3039    switch (instr->intrinsic) {
3040    case nir_intrinsic_ballot:
3041    case nir_intrinsic_ballot_relaxed:
3042       result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
3043       if (instr->def.bit_size > ctx->ac.wave_size) {
3044          LLVMTypeRef dest_type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
3045          result = LLVMBuildZExt(ctx->ac.builder, result, dest_type, "");
3046       }
3047       break;
3048    case nir_intrinsic_inverse_ballot: {
3049       LLVMValueRef src = get_src(ctx, instr->src[0]);
3050       if (instr->src[0].ssa->bit_size > ctx->ac.wave_size) {
3051          LLVMTypeRef src_type = LLVMIntTypeInContext(ctx->ac.context, ctx->ac.wave_size);
3052          src = LLVMBuildTrunc(ctx->ac.builder, src, src_type, "");
3053       }
3054       result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.inverse.ballot", ctx->ac.i1, &src, 1, 0);
3055       break;
3056    }
3057    case nir_intrinsic_read_invocation:
3058       result =
3059          ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), get_src(ctx, instr->src[1]));
3060       break;
3061    case nir_intrinsic_read_first_invocation:
3062    case nir_intrinsic_as_uniform:
3063       result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), NULL);
3064       break;
3065    case nir_intrinsic_load_subgroup_invocation:
3066       result = ac_get_thread_id(&ctx->ac);
3067       break;
3068    case nir_intrinsic_load_workgroup_id: {
3069       LLVMValueRef values[3] = {ctx->ac.i32_0, ctx->ac.i32_0, ctx->ac.i32_0};
3070 
3071       for (int i = 0; i < 3; i++) {
3072          if (ctx->args->workgroup_ids[i].used)
3073             values[i] = ac_get_arg(&ctx->ac, ctx->args->workgroup_ids[i]);
3074       }
3075       result = ac_build_gather_values(&ctx->ac, values, 3);
3076       break;
3077    }
3078    case nir_intrinsic_load_base_vertex:
3079    case nir_intrinsic_load_first_vertex:
3080    case nir_intrinsic_load_tess_rel_patch_id_amd:
3081    case nir_intrinsic_load_ring_attr_amd:
3082    case nir_intrinsic_load_lds_ngg_scratch_base_amd:
3083    case nir_intrinsic_load_lds_ngg_gs_out_vertex_base_amd:
3084       result = ctx->abi->intrinsic_load(ctx->abi, instr);
3085       break;
3086    case nir_intrinsic_load_vertex_id_zero_base:
3087       result = ctx->abi->vertex_id_replaced ? ctx->abi->vertex_id_replaced : ctx->abi->vertex_id;
3088       break;
3089    case nir_intrinsic_load_local_invocation_id: {
3090       LLVMValueRef ids = ac_get_arg(&ctx->ac, ctx->args->local_invocation_ids);
3091 
3092       if (LLVMGetTypeKind(LLVMTypeOf(ids)) == LLVMIntegerTypeKind) {
3093          /* Thread IDs are packed in VGPR0, 10 bits per component. */
3094          LLVMValueRef id[3];
3095 
3096          for (unsigned i = 0; i < 3; i++)
3097             id[i] = ac_unpack_param(&ctx->ac, ids, i * 10, 10);
3098 
3099          result = ac_build_gather_values(&ctx->ac, id, 3);
3100       } else {
3101          result = ids;
3102       }
3103       break;
3104    }
3105    case nir_intrinsic_load_base_instance:
3106       result = ac_get_arg(&ctx->ac, ctx->args->start_instance);
3107       break;
3108    case nir_intrinsic_load_draw_id:
3109       result = ac_get_arg(&ctx->ac, ctx->args->draw_id);
3110       break;
3111    case nir_intrinsic_load_view_index:
3112       result = ac_get_arg(&ctx->ac, ctx->args->view_index);
3113       break;
3114    case nir_intrinsic_load_invocation_id:
3115       assert(ctx->stage == MESA_SHADER_TESS_CTRL || ctx->stage == MESA_SHADER_GEOMETRY);
3116       if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3117          result = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->tcs_rel_ids), 8, 5);
3118       } else if (ctx->ac.gfx_level >= GFX10) {
3119          result = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id), 0, 7);
3120       } else {
3121          result = ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id);
3122       }
3123       break;
3124    case nir_intrinsic_load_primitive_id:
3125       if (ctx->stage == MESA_SHADER_GEOMETRY) {
3126          result = ac_get_arg(&ctx->ac, ctx->args->gs_prim_id);
3127       } else if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3128          result = ac_get_arg(&ctx->ac, ctx->args->tcs_patch_id);
3129       } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
3130          result = ctx->abi->tes_patch_id_replaced ?
3131             ctx->abi->tes_patch_id_replaced : ac_get_arg(&ctx->ac, ctx->args->tes_patch_id);
3132       } else if (ctx->stage == MESA_SHADER_VERTEX) {
3133          if (ctx->args->vs_prim_id.used)
3134             result = ac_get_arg(&ctx->ac, ctx->args->vs_prim_id); /* legacy */
3135          else
3136             result = ac_get_arg(&ctx->ac, ctx->args->gs_prim_id); /* NGG */
3137       } else
3138          fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
3139       break;
3140    case nir_intrinsic_load_sample_id:
3141       result = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ancillary), 8, 4);
3142       break;
3143    case nir_intrinsic_load_sample_pos:
3144       result = load_sample_pos(ctx);
3145       break;
3146    case nir_intrinsic_load_frag_coord:
3147       result = emit_load_frag_coord(ctx);
3148       break;
3149    case nir_intrinsic_load_frag_shading_rate:
3150       result = emit_load_frag_shading_rate(ctx);
3151       break;
3152    case nir_intrinsic_load_front_face:
3153       result = emit_i2b(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->front_face));
3154       break;
3155    case nir_intrinsic_load_helper_invocation:
3156    case nir_intrinsic_is_helper_invocation:
3157       result = ac_build_load_helper_invocation(&ctx->ac);
3158       break;
3159    case nir_intrinsic_load_instance_id:
3160       result = ctx->abi->instance_id_replaced ?
3161          ctx->abi->instance_id_replaced : ctx->abi->instance_id;
3162       break;
3163    case nir_intrinsic_load_num_workgroups:
3164       if (ctx->abi->load_grid_size_from_user_sgpr) {
3165          result = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
3166       } else {
3167          result = ac_build_load_invariant(&ctx->ac,
3168             ac_get_ptr_arg(&ctx->ac, ctx->args, ctx->args->num_work_groups), ctx->ac.i32_0);
3169       }
3170       break;
3171    case nir_intrinsic_load_local_invocation_index:
3172       result = visit_load_local_invocation_index(ctx);
3173       break;
3174    case nir_intrinsic_first_invocation:
3175       result = visit_first_invocation(ctx);
3176       break;
3177    case nir_intrinsic_load_push_constant:
3178       result = visit_load_push_constant(ctx, instr);
3179       break;
3180    case nir_intrinsic_store_ssbo:
3181       visit_store_ssbo(ctx, instr);
3182       break;
3183    case nir_intrinsic_load_ssbo:
3184       result = visit_load_buffer(ctx, instr);
3185       break;
3186    case nir_intrinsic_load_global_constant:
3187    case nir_intrinsic_load_global:
3188    case nir_intrinsic_load_global_amd:
3189       result = visit_load_global(ctx, instr);
3190       break;
3191    case nir_intrinsic_store_global:
3192    case nir_intrinsic_store_global_amd:
3193       visit_store_global(ctx, instr);
3194       break;
3195    case nir_intrinsic_global_atomic:
3196    case nir_intrinsic_global_atomic_swap:
3197    case nir_intrinsic_global_atomic_amd:
3198    case nir_intrinsic_global_atomic_swap_amd:
3199       result = visit_global_atomic(ctx, instr);
3200       break;
3201    case nir_intrinsic_ssbo_atomic:
3202    case nir_intrinsic_ssbo_atomic_swap:
3203       result = visit_atomic_ssbo(ctx, instr);
3204       break;
3205    case nir_intrinsic_load_ubo:
3206       result = visit_load_ubo_buffer(ctx, instr);
3207       break;
3208    case nir_intrinsic_get_ssbo_size:
3209       result = visit_get_ssbo_size(ctx, instr);
3210       break;
3211    case nir_intrinsic_load_input:
3212    case nir_intrinsic_load_input_vertex:
3213    case nir_intrinsic_load_per_vertex_input:
3214       result = visit_load(ctx, instr, false);
3215       break;
3216    case nir_intrinsic_load_output:
3217    case nir_intrinsic_load_per_vertex_output:
3218       result = visit_load(ctx, instr, true);
3219       break;
3220    case nir_intrinsic_store_output:
3221    case nir_intrinsic_store_per_vertex_output:
3222       visit_store_output(ctx, instr);
3223       break;
3224    case nir_intrinsic_load_shared:
3225       result = visit_load_shared(ctx, instr);
3226       break;
3227    case nir_intrinsic_store_shared:
3228       visit_store_shared(ctx, instr);
3229       break;
3230    case nir_intrinsic_load_shared2_amd:
3231       result = visit_load_shared2_amd(ctx, instr);
3232       break;
3233    case nir_intrinsic_store_shared2_amd:
3234       visit_store_shared2_amd(ctx, instr);
3235       break;
3236    case nir_intrinsic_bindless_image_load:
3237    case nir_intrinsic_bindless_image_sparse_load:
3238    case nir_intrinsic_bindless_image_fragment_mask_load_amd:
3239       result = visit_image_load(ctx, instr);
3240       break;
3241    case nir_intrinsic_bindless_image_store:
3242       visit_image_store(ctx, instr);
3243       break;
3244    case nir_intrinsic_bindless_image_atomic:
3245    case nir_intrinsic_bindless_image_atomic_swap:
3246       result = visit_image_atomic(ctx, instr);
3247       break;
3248    case nir_intrinsic_shader_clock:
3249       result = ac_build_shader_clock(&ctx->ac, nir_intrinsic_memory_scope(instr));
3250       break;
3251    case nir_intrinsic_discard:
3252    case nir_intrinsic_discard_if:
3253    case nir_intrinsic_terminate:
3254    case nir_intrinsic_terminate_if:
3255       emit_discard(ctx, instr);
3256       break;
3257    case nir_intrinsic_demote:
3258    case nir_intrinsic_demote_if:
3259       emit_demote(ctx, instr);
3260       break;
3261    case nir_intrinsic_barrier: {
3262       assert(!(nir_intrinsic_memory_semantics(instr) &
3263                (NIR_MEMORY_MAKE_AVAILABLE | NIR_MEMORY_MAKE_VISIBLE)));
3264 
3265       nir_variable_mode modes = nir_intrinsic_memory_modes(instr);
3266 
3267       unsigned wait_flags = 0;
3268       if (modes & (nir_var_mem_global | nir_var_mem_ssbo | nir_var_image))
3269          wait_flags |= AC_WAIT_VLOAD | AC_WAIT_VSTORE;
3270       if (modes & nir_var_mem_shared)
3271          wait_flags |= AC_WAIT_LGKM;
3272 
3273       if (wait_flags)
3274          ac_build_waitcnt(&ctx->ac, wait_flags);
3275 
3276       if (nir_intrinsic_execution_scope(instr) == SCOPE_WORKGROUP)
3277          ac_build_s_barrier(&ctx->ac, ctx->stage);
3278       break;
3279    }
3280    case nir_intrinsic_optimization_barrier_vgpr_amd:
3281       result = get_src(ctx, instr->src[0]);
3282       ac_build_optimization_barrier(&ctx->ac, &result, false);
3283       break;
3284    case nir_intrinsic_shared_atomic:
3285    case nir_intrinsic_shared_atomic_swap: {
3286       LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0], 0);
3287       result = visit_var_atomic(ctx, instr, ptr, 1);
3288       break;
3289    }
3290    case nir_intrinsic_load_barycentric_pixel:
3291       result = barycentric_center(ctx, nir_intrinsic_interp_mode(instr));
3292       break;
3293    case nir_intrinsic_load_barycentric_centroid:
3294       result = barycentric_centroid(ctx, nir_intrinsic_interp_mode(instr));
3295       break;
3296    case nir_intrinsic_load_barycentric_sample:
3297       result = barycentric_sample(ctx, nir_intrinsic_interp_mode(instr));
3298       break;
3299    case nir_intrinsic_load_barycentric_model:
3300       result = barycentric_model(ctx);
3301       break;
3302    case nir_intrinsic_load_barycentric_at_offset: {
3303       LLVMValueRef offset = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
3304       result = barycentric_offset(ctx, nir_intrinsic_interp_mode(instr), offset);
3305       break;
3306    }
3307    case nir_intrinsic_load_interpolated_input: {
3308       /* We assume any indirect loads have been lowered away */
3309       ASSERTED nir_const_value *offset = nir_src_as_const_value(instr->src[1]);
3310       assert(offset);
3311       assert(offset[0].i32 == 0);
3312 
3313       LLVMValueRef interp_param = get_src(ctx, instr->src[0]);
3314       unsigned index = nir_intrinsic_base(instr);
3315       unsigned component = nir_intrinsic_component(instr);
3316       result = load_interpolated_input(ctx, interp_param, index, component,
3317                                        instr->def.num_components, instr->def.bit_size,
3318                                        nir_intrinsic_io_semantics(instr).high_16bits);
3319       break;
3320    }
3321    case nir_intrinsic_sendmsg_amd: {
3322       unsigned imm = nir_intrinsic_base(instr);
3323       LLVMValueRef m0_content = get_src(ctx, instr->src[0]);
3324       ac_build_sendmsg(&ctx->ac, imm, m0_content);
3325       break;
3326    }
3327    case nir_intrinsic_load_gs_wave_id_amd: {
3328       if (ctx->args->merged_wave_info.used)
3329          result = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->merged_wave_info), 16, 8);
3330       else if (ctx->args->gs_wave_id.used)
3331          result = ac_get_arg(&ctx->ac, ctx->args->gs_wave_id);
3332       else
3333          unreachable("Shader doesn't have GS wave ID.");
3334       break;
3335    }
3336    case nir_intrinsic_load_tess_coord: {
3337       LLVMValueRef coord[] = {
3338          ctx->abi->tes_u_replaced ? ctx->abi->tes_u_replaced : ac_get_arg(&ctx->ac, ctx->args->tes_u),
3339          ctx->abi->tes_v_replaced ? ctx->abi->tes_v_replaced : ac_get_arg(&ctx->ac, ctx->args->tes_v),
3340          ctx->ac.f32_0,
3341       };
3342 
3343       /* For triangles, the vector should be (u, v, 1-u-v). */
3344       if (ctx->info->tess._primitive_mode == TESS_PRIMITIVE_TRIANGLES) {
3345          coord[2] = LLVMBuildFSub(ctx->ac.builder, ctx->ac.f32_1,
3346                                   LLVMBuildFAdd(ctx->ac.builder, coord[0], coord[1], ""), "");
3347       }
3348       result = ac_build_gather_values(&ctx->ac, coord, 3);
3349       break;
3350    }
3351    case nir_intrinsic_vote_all: {
3352       result = ac_build_vote_all(&ctx->ac, get_src(ctx, instr->src[0]));
3353       break;
3354    }
3355    case nir_intrinsic_vote_any: {
3356       result = ac_build_vote_any(&ctx->ac, get_src(ctx, instr->src[0]));
3357       break;
3358    }
3359    case nir_intrinsic_quad_vote_any: {
3360       result = ac_build_wqm_vote(&ctx->ac, get_src(ctx, instr->src[0]));
3361       break;
3362    }
3363    case nir_intrinsic_quad_vote_all: {
3364       LLVMValueRef src = LLVMBuildNot(ctx->ac.builder, get_src(ctx, instr->src[0]), "");
3365       result = LLVMBuildNot(ctx->ac.builder, ac_build_wqm_vote(&ctx->ac, src), "");
3366       break;
3367    }
3368    case nir_intrinsic_shuffle:
3369       if (ctx->ac.gfx_level == GFX8 || ctx->ac.gfx_level == GFX9 ||
3370           (ctx->ac.gfx_level >= GFX10 && ctx->ac.wave_size == 32)) {
3371          result =
3372             ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]), get_src(ctx, instr->src[1]));
3373       } else {
3374          LLVMValueRef src = get_src(ctx, instr->src[0]);
3375          LLVMValueRef index = get_src(ctx, instr->src[1]);
3376          LLVMTypeRef type = LLVMTypeOf(src);
3377          struct waterfall_context wctx;
3378          LLVMValueRef index_val;
3379 
3380          index_val = enter_waterfall(ctx, &wctx, index, true);
3381 
3382          src = LLVMBuildZExt(ctx->ac.builder, src, ctx->ac.i32, "");
3383 
3384          result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.readlane", ctx->ac.i32,
3385                                      (LLVMValueRef[]){src, index_val}, 2, 0);
3386 
3387          result = LLVMBuildTrunc(ctx->ac.builder, result, type, "");
3388 
3389          result = exit_waterfall(ctx, &wctx, result);
3390       }
3391       break;
3392    case nir_intrinsic_reduce:
3393       result = ac_build_reduce(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0],
3394                                instr->const_index[1]);
3395       break;
3396    case nir_intrinsic_inclusive_scan:
3397       result =
3398          ac_build_inclusive_scan(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0]);
3399       break;
3400    case nir_intrinsic_exclusive_scan:
3401       result =
3402          ac_build_exclusive_scan(&ctx->ac, get_src(ctx, instr->src[0]), instr->const_index[0]);
3403       break;
3404    case nir_intrinsic_quad_broadcast: {
3405       unsigned lane = nir_src_as_uint(instr->src[1]);
3406       result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), lane, lane, lane, lane);
3407       result = ac_build_wqm(&ctx->ac, result);
3408       break;
3409    }
3410    case nir_intrinsic_quad_swap_horizontal:
3411       result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 1, 0, 3, 2);
3412       result = ac_build_wqm(&ctx->ac, result);
3413       break;
3414    case nir_intrinsic_quad_swap_vertical:
3415       result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 2, 3, 0, 1);
3416       result = ac_build_wqm(&ctx->ac, result);
3417       break;
3418    case nir_intrinsic_quad_swap_diagonal:
3419       result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1, 0);
3420       result = ac_build_wqm(&ctx->ac, result);
3421       break;
3422    case nir_intrinsic_quad_swizzle_amd: {
3423       uint32_t mask = nir_intrinsic_swizzle_mask(instr);
3424       result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask & 0x3,
3425                                      (mask >> 2) & 0x3, (mask >> 4) & 0x3, (mask >> 6) & 0x3);
3426       result = ac_build_wqm(&ctx->ac, result);
3427       break;
3428    }
3429    case nir_intrinsic_masked_swizzle_amd: {
3430       uint32_t mask = nir_intrinsic_swizzle_mask(instr);
3431       result = ac_build_ds_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask);
3432       break;
3433    }
3434    case nir_intrinsic_write_invocation_amd:
3435       result = ac_build_writelane(&ctx->ac, get_src(ctx, instr->src[0]),
3436                                   get_src(ctx, instr->src[1]), get_src(ctx, instr->src[2]));
3437       break;
3438    case nir_intrinsic_mbcnt_amd:
3439       result = ac_build_mbcnt_add(&ctx->ac, get_src(ctx, instr->src[0]), get_src(ctx, instr->src[1]));
3440       break;
3441    case nir_intrinsic_load_scratch: {
3442       LLVMValueRef offset = get_src(ctx, instr->src[0]);
3443       LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch, offset);
3444       LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
3445       LLVMTypeRef vec_type = instr->def.num_components == 1
3446                                 ? comp_type
3447                                 : LLVMVectorType(comp_type, instr->def.num_components);
3448       result = LLVMBuildLoad2(ctx->ac.builder, vec_type, ptr, "");
3449       break;
3450    }
3451    case nir_intrinsic_store_scratch: {
3452       LLVMValueRef offset = get_src(ctx, instr->src[1]);
3453       LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch, offset);
3454       LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
3455       LLVMValueRef src = get_src(ctx, instr->src[0]);
3456       unsigned wrmask = nir_intrinsic_write_mask(instr);
3457       while (wrmask) {
3458          int start, count;
3459          u_bit_scan_consecutive_range(&wrmask, &start, &count);
3460 
3461          LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, start, false);
3462          LLVMValueRef offset_ptr = LLVMBuildGEP2(ctx->ac.builder, comp_type, ptr, &offset, 1, "");
3463          LLVMValueRef offset_src = ac_extract_components(&ctx->ac, src, start, count);
3464          LLVMBuildStore(ctx->ac.builder, offset_src, offset_ptr);
3465       }
3466       break;
3467    }
3468    case nir_intrinsic_load_constant: {
3469       unsigned base = nir_intrinsic_base(instr);
3470       unsigned range = nir_intrinsic_range(instr);
3471 
3472       LLVMValueRef offset = get_src(ctx, instr->src[0]);
3473       offset = LLVMBuildAdd(ctx->ac.builder, offset, LLVMConstInt(ctx->ac.i32, base, false), "");
3474 
3475       /* Clamp the offset to avoid out-of-bound access because global
3476        * instructions can't handle them.
3477        */
3478       LLVMValueRef size = LLVMConstInt(ctx->ac.i32, base + range, false);
3479       LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
3480       offset = LLVMBuildSelect(ctx->ac.builder, cond, offset, size, "");
3481 
3482       LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->constant_data, offset);
3483       LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
3484       LLVMTypeRef vec_type = instr->def.num_components == 1
3485                                 ? comp_type
3486                                 : LLVMVectorType(comp_type, instr->def.num_components);
3487       result = LLVMBuildLoad2(ctx->ac.builder, vec_type, ptr, "");
3488       break;
3489    }
3490    case nir_intrinsic_set_vertex_and_primitive_count:
3491       /* Currently ignored. */
3492       break;
3493    case nir_intrinsic_load_typed_buffer_amd:
3494    case nir_intrinsic_load_buffer_amd:
3495    case nir_intrinsic_store_buffer_amd: {
3496       unsigned src_base = instr->intrinsic == nir_intrinsic_store_buffer_amd ? 1 : 0;
3497       bool idxen = !nir_src_is_const(instr->src[src_base + 3]) ||
3498                    nir_src_as_uint(instr->src[src_base + 3]);
3499 
3500       LLVMValueRef store_data = get_src(ctx, instr->src[0]);
3501       LLVMValueRef descriptor = get_src(ctx, instr->src[src_base + 0]);
3502       LLVMValueRef addr_voffset = get_src(ctx, instr->src[src_base + 1]);
3503       LLVMValueRef addr_soffset = get_src(ctx, instr->src[src_base + 2]);
3504       LLVMValueRef vidx = idxen ? get_src(ctx, instr->src[src_base + 3]) : NULL;
3505       unsigned num_components = instr->def.num_components;
3506       unsigned const_offset = nir_intrinsic_base(instr);
3507       bool reorder = nir_intrinsic_can_reorder(instr);
3508       enum gl_access_qualifier access = ac_get_mem_access_flags(instr);
3509       bool uses_format = access & ACCESS_USES_FORMAT_AMD;
3510 
3511       LLVMValueRef voffset = LLVMBuildAdd(ctx->ac.builder, addr_voffset,
3512                                           LLVMConstInt(ctx->ac.i32, const_offset, 0), "");
3513 
3514       if (instr->intrinsic == nir_intrinsic_load_buffer_amd && uses_format) {
3515          assert(instr->def.bit_size == 16 || instr->def.bit_size == 32);
3516          result = ac_build_buffer_load_format(&ctx->ac, descriptor, vidx, voffset, num_components,
3517                                               access, reorder,
3518                                               instr->def.bit_size == 16, false);
3519          result = ac_to_integer(&ctx->ac, result);
3520       } else if (instr->intrinsic == nir_intrinsic_store_buffer_amd && uses_format) {
3521          assert(instr->src[0].ssa->bit_size == 16 || instr->src[0].ssa->bit_size == 32);
3522          ac_build_buffer_store_format(&ctx->ac, descriptor, store_data, vidx, voffset, access);
3523       } else if (instr->intrinsic == nir_intrinsic_load_buffer_amd ||
3524                  instr->intrinsic == nir_intrinsic_load_typed_buffer_amd) {
3525          /* LLVM is unable to select instructions for larger than 32-bit channel types.
3526           * Workaround by using i32 and casting to the correct type later.
3527           */
3528          const unsigned fetch_num_components =
3529             num_components * MAX2(32, instr->def.bit_size) / 32;
3530 
3531          LLVMTypeRef channel_type =
3532             LLVMIntTypeInContext(ctx->ac.context, MIN2(32, instr->def.bit_size));
3533 
3534          if (instr->intrinsic == nir_intrinsic_load_buffer_amd) {
3535             result = ac_build_buffer_load(&ctx->ac, descriptor, fetch_num_components, vidx, voffset,
3536                                           addr_soffset, channel_type, access, reorder, false);
3537          } else {
3538             const unsigned align_offset = nir_intrinsic_align_offset(instr);
3539             const unsigned align_mul = nir_intrinsic_align_mul(instr);
3540             const enum pipe_format format = nir_intrinsic_format(instr);
3541 
3542             result =
3543                ac_build_safe_tbuffer_load(&ctx->ac, descriptor, vidx, addr_voffset, addr_soffset,
3544                                           format, MIN2(32, instr->def.bit_size), const_offset, align_offset,
3545                                           align_mul, fetch_num_components, access, reorder);
3546          }
3547 
3548          /* Trim to needed vector components. */
3549          result = ac_trim_vector(&ctx->ac, result, fetch_num_components);
3550 
3551          /* Cast to larger than 32-bit sized components if needed. */
3552          if (instr->def.bit_size > 32) {
3553             LLVMTypeRef cast_channel_type =
3554                LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
3555             LLVMTypeRef cast_type =
3556                num_components == 1 ? cast_channel_type :
3557                LLVMVectorType(cast_channel_type, num_components);
3558             result = LLVMBuildBitCast(ctx->ac.builder, result, cast_type, "");
3559          }
3560 
3561          /* Cast the result to an integer (or vector of integers). */
3562          result = ac_to_integer(&ctx->ac, result);
3563       } else {
3564          unsigned writemask = nir_intrinsic_write_mask(instr);
3565          while (writemask) {
3566             int start, count;
3567             u_bit_scan_consecutive_range(&writemask, &start, &count);
3568 
3569             LLVMValueRef voffset = LLVMBuildAdd(
3570                ctx->ac.builder, addr_voffset,
3571                LLVMConstInt(ctx->ac.i32, const_offset + start * 4, 0), "");
3572 
3573             LLVMValueRef data = extract_vector_range(&ctx->ac, store_data, start, count);
3574             ac_build_buffer_store_dword(&ctx->ac, descriptor, data, vidx, voffset, addr_soffset,
3575                                         access);
3576          }
3577       }
3578       break;
3579    }
3580    case nir_intrinsic_is_subgroup_invocation_lt_amd: {
3581       LLVMValueRef count = LLVMBuildAnd(ctx->ac.builder, get_src(ctx, instr->src[0]),
3582                                         LLVMConstInt(ctx->ac.i32, 0xff, 0), "");
3583       result = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, ac_get_thread_id(&ctx->ac), count, "");
3584       break;
3585    }
3586    case nir_intrinsic_overwrite_vs_arguments_amd:
3587       ctx->abi->vertex_id_replaced = get_src(ctx, instr->src[0]);
3588       ctx->abi->instance_id_replaced = get_src(ctx, instr->src[1]);
3589       break;
3590    case nir_intrinsic_overwrite_tes_arguments_amd:
3591       ctx->abi->tes_u_replaced = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
3592       ctx->abi->tes_v_replaced = ac_to_float(&ctx->ac, get_src(ctx, instr->src[1]));
3593       ctx->abi->tes_rel_patch_id_replaced = get_src(ctx, instr->src[3]);
3594       ctx->abi->tes_patch_id_replaced = get_src(ctx, instr->src[2]);
3595       break;
3596    case nir_intrinsic_gds_atomic_add_amd: {
3597       LLVMValueRef store_val = get_src(ctx, instr->src[0]);
3598       LLVMValueRef addr = get_src(ctx, instr->src[1]);
3599       LLVMTypeRef gds_ptr_type = LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_GDS);
3600       LLVMValueRef gds_base = LLVMBuildIntToPtr(ctx->ac.builder, addr, gds_ptr_type, "");
3601       ac_build_atomic_rmw(&ctx->ac, LLVMAtomicRMWBinOpAdd, gds_base, store_val, "workgroup-one-as");
3602       break;
3603    }
3604    case nir_intrinsic_elect:
3605       result = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, visit_first_invocation(ctx),
3606                              ac_get_thread_id(&ctx->ac), "");
3607       break;
3608    case nir_intrinsic_lane_permute_16_amd:
3609       result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.permlane16", ctx->ac.i32,
3610                                   (LLVMValueRef[]){get_src(ctx, instr->src[0]),
3611                                                    get_src(ctx, instr->src[0]),
3612                                                    get_src(ctx, instr->src[1]),
3613                                                    get_src(ctx, instr->src[2]),
3614                                                    ctx->ac.i1false,
3615                                                    ctx->ac.i1false}, 6, 0);
3616       break;
3617    case nir_intrinsic_load_scalar_arg_amd:
3618    case nir_intrinsic_load_vector_arg_amd: {
3619       assert(nir_intrinsic_base(instr) < AC_MAX_ARGS);
3620       struct ac_arg arg;
3621       arg.arg_index = nir_intrinsic_base(instr);
3622       arg.used = true;
3623       result = ac_to_integer(&ctx->ac, ac_get_arg(&ctx->ac, arg));
3624       if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(result)) != 32)
3625          result = LLVMBuildBitCast(ctx->ac.builder, result, get_def_type(ctx, &instr->def), "");
3626       break;
3627    }
3628    case nir_intrinsic_load_smem_amd: {
3629       LLVMValueRef base = get_src(ctx, instr->src[0]);
3630       LLVMValueRef offset = get_src(ctx, instr->src[1]);
3631 
3632       bool is_addr_32bit = nir_src_bit_size(instr->src[0]) == 32;
3633       int addr_space = is_addr_32bit ? AC_ADDR_SPACE_CONST_32BIT : AC_ADDR_SPACE_CONST;
3634 
3635       LLVMTypeRef result_type = get_def_type(ctx, &instr->def);
3636       LLVMTypeRef byte_ptr_type = LLVMPointerType(ctx->ac.i8, addr_space);
3637 
3638       LLVMValueRef addr = LLVMBuildIntToPtr(ctx->ac.builder, base, byte_ptr_type, "");
3639       /* see ac_build_load_custom() for 32bit/64bit addr GEP difference */
3640       addr = is_addr_32bit ?
3641          LLVMBuildInBoundsGEP2(ctx->ac.builder, ctx->ac.i8, addr, &offset, 1, "") :
3642          LLVMBuildGEP2(ctx->ac.builder, ctx->ac.i8, addr, &offset, 1, "");
3643 
3644       LLVMSetMetadata(addr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
3645       result = LLVMBuildLoad2(ctx->ac.builder, result_type, addr, "");
3646       LLVMSetMetadata(result, ctx->ac.invariant_load_md_kind, ctx->ac.empty_md);
3647       break;
3648    }
3649    case nir_intrinsic_ordered_xfb_counter_add_amd: {
3650       /* must be called in a single lane of a workgroup. */
3651       LLVMTypeRef gdsptr = LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_GDS);
3652 
3653       /* Gfx11 GDS instructions only operate on the first active lane. All other lanes are
3654        * ignored. So are their EXEC bits. This uses the mutex feature of ds_ordered_count
3655        * to emulate a multi-dword atomic.
3656        *
3657        * This is the expected code:
3658        *    ds_ordered_count release=0 done=0   // lock mutex
3659        *    if (gfx_level >= GFX11) {
3660        *       ds_add_gs_reg_rtn GDS_STRMOUT_DWORDS_WRITTEN_0
3661        *       ds_add_gs_reg_rtn GDS_STRMOUT_DWORDS_WRITTEN_1
3662        *       ds_add_gs_reg_rtn GDS_STRMOUT_DWORDS_WRITTEN_2
3663        *       ds_add_gs_reg_rtn GDS_STRMOUT_DWORDS_WRITTEN_3
3664        *    } else {
3665        *       ds_add_rtn_u32 dwords_written0
3666        *       ds_add_rtn_u32 dwords_written1
3667        *       ds_add_rtn_u32 dwords_written2
3668        *       ds_add_rtn_u32 dwords_written3
3669        *    }
3670        *    ds_ordered_count release=1 done=1   // unlock mutex
3671        *
3672        * GDS_STRMOUT_DWORDS_WRITTEN_n are just general-purpose global registers. We use them
3673        * because MCBP (mid-command-buffer preemption) saves and restores them, and it doesn't
3674        * save and restore GDS memory.
3675        */
3676       LLVMValueRef args[8] = {
3677          LLVMBuildIntToPtr(ctx->ac.builder, get_src(ctx, instr->src[0]), gdsptr, ""),
3678          ctx->ac.i32_0,                             /* value to add */
3679          ctx->ac.i32_0,                             /* ordering */
3680          ctx->ac.i32_0,                             /* scope */
3681          ctx->ac.i1false,                           /* isVolatile */
3682          LLVMConstInt(ctx->ac.i32, 1 << 24, false), /* OA index, bits 24+: lane count */
3683          ctx->ac.i1false,                           /* wave release */
3684          ctx->ac.i1false,                           /* wave done */
3685       };
3686 
3687       /* Set release=0 to start a GDS mutex. Set done=0 because it's not the last one. */
3688       ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ds.ordered.add", ctx->ac.i32,
3689                          args, ARRAY_SIZE(args), 0);
3690       ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM);
3691 
3692       LLVMValueRef global_count[4];
3693       LLVMValueRef count_vec = get_src(ctx, instr->src[1]);
3694       unsigned write_mask = nir_intrinsic_write_mask(instr);
3695       for (unsigned i = 0; i < instr->num_components; i++) {
3696          LLVMValueRef value =
3697             LLVMBuildExtractElement(ctx->ac.builder, count_vec,
3698                                     LLVMConstInt(ctx->ac.i32, i, false), "");
3699          if (write_mask & (1 << i)) {
3700             /* The offset is a relative offset from GDS_STRMOUT_DWORDS_WRITTEN_0. */
3701             global_count[i] =
3702                ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ds.add.gs.reg.rtn.i32", ctx->ac.i32,
3703                                   (LLVMValueRef[]){value, LLVMConstInt(ctx->ac.i32, i * 4, 0)},
3704                                   2, 0);
3705          } else {
3706             global_count[i] = LLVMGetUndef(ctx->ac.i32);
3707          }
3708       }
3709 
3710       ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM);
3711 
3712       /* Set release=1 to end a GDS mutex. Set done=1 because it's the last one. */
3713       args[6] = args[7] = ctx->ac.i1true;
3714       ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ds.ordered.add", ctx->ac.i32,
3715                          args, ARRAY_SIZE(args), 0);
3716       result = ac_build_gather_values(&ctx->ac, global_count, instr->num_components);
3717       break;
3718    }
3719    case nir_intrinsic_xfb_counter_sub_amd: {
3720       /* must be called in a single lane of a workgroup. */
3721       LLVMValueRef sub_vec = get_src(ctx, instr->src[0]);
3722       unsigned write_mask = nir_intrinsic_write_mask(instr);
3723 
3724       for (unsigned i = 0; i < instr->num_components; i++) {
3725          if (write_mask & (1 << i)) {
3726             LLVMValueRef value =
3727                LLVMBuildExtractElement(ctx->ac.builder, sub_vec,
3728                                        LLVMConstInt(ctx->ac.i32, i, false), "");
3729             /* The offset is a relative offset from GDS_STRMOUT_DWORDS_WRITTEN_0. */
3730             ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ds.sub.gs.reg.rtn.i32", ctx->ac.i32,
3731                                (LLVMValueRef[]){value, LLVMConstInt(ctx->ac.i32, i * 4, 0)},
3732                                2, 0);
3733          }
3734       }
3735       break;
3736    }
3737    case nir_intrinsic_export_amd: {
3738       unsigned flags = nir_intrinsic_flags(instr);
3739       unsigned target = nir_intrinsic_base(instr);
3740       unsigned write_mask = nir_intrinsic_write_mask(instr);
3741 
3742       struct ac_export_args args = {
3743          .target = target,
3744          .enabled_channels = write_mask,
3745          .compr = flags & AC_EXP_FLAG_COMPRESSED,
3746          .done = flags & AC_EXP_FLAG_DONE,
3747          .valid_mask = flags & AC_EXP_FLAG_VALID_MASK,
3748       };
3749 
3750       LLVMValueRef value = get_src(ctx, instr->src[0]);
3751       int num_components = ac_get_llvm_num_components(value);
3752       for (int i = 0; i < num_components; i++)
3753          args.out[i] = ac_llvm_extract_elem(&ctx->ac, value, i);
3754 
3755       ac_build_export(&ctx->ac, &args);
3756       break;
3757    }
3758    case nir_intrinsic_bvh64_intersect_ray_amd: {
3759       LLVMValueRef desc = get_src(ctx, instr->src[0]);
3760       LLVMValueRef node_id =
3761          LLVMBuildBitCast(ctx->ac.builder, get_src(ctx, instr->src[1]), ctx->ac.i64, "");
3762       LLVMValueRef t_max =
3763          LLVMBuildBitCast(ctx->ac.builder, get_src(ctx, instr->src[2]), ctx->ac.f32, "");
3764       LLVMValueRef origin =
3765          LLVMBuildBitCast(ctx->ac.builder, get_src(ctx, instr->src[3]), ctx->ac.v3f32, "");
3766       LLVMValueRef dir =
3767          LLVMBuildBitCast(ctx->ac.builder, get_src(ctx, instr->src[4]), ctx->ac.v3f32, "");
3768       LLVMValueRef inv_dir =
3769          LLVMBuildBitCast(ctx->ac.builder, get_src(ctx, instr->src[5]), ctx->ac.v3f32, "");
3770 
3771       LLVMValueRef args[6] = {
3772          node_id, t_max, origin, dir, inv_dir, desc,
3773       };
3774 
3775       result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.image.bvh.intersect.ray.i64.v3f32",
3776                                   ctx->ac.v4i32, args, ARRAY_SIZE(args), 0);
3777       break;
3778    }
3779    default:
3780       fprintf(stderr, "Unknown intrinsic: ");
3781       nir_print_instr(&instr->instr, stderr);
3782       fprintf(stderr, "\n");
3783       return false;
3784    }
3785    if (result) {
3786       ctx->ssa_defs[instr->def.index] = result;
3787    }
3788    return true;
3789 }
3790 
3791 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
3792  *
3793  * GFX6-GFX7:
3794  *   If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
3795  *   filtering manually. The driver sets img7 to a mask clearing
3796  *   MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
3797  *     s_and_b32 samp0, samp0, img7
3798  *
3799  * GFX8:
3800  *   The ANISO_OVERRIDE sampler field enables this fix in TA.
3801  */
sici_fix_sampler_aniso(struct ac_nir_context * ctx,LLVMValueRef res,LLVMValueRef samp)3802 static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx, LLVMValueRef res,
3803                                            LLVMValueRef samp)
3804 {
3805    LLVMBuilderRef builder = ctx->ac.builder;
3806    LLVMValueRef img7, samp0;
3807 
3808    if (ctx->ac.gfx_level >= GFX8)
3809       return samp;
3810 
3811    img7 = LLVMBuildExtractElement(builder, res, LLVMConstInt(ctx->ac.i32, 7, 0), "");
3812    samp0 = LLVMBuildExtractElement(builder, samp, ctx->ac.i32_0, "");
3813    samp0 = LLVMBuildAnd(builder, samp0, img7, "");
3814    return LLVMBuildInsertElement(builder, samp, samp0, ctx->ac.i32_0, "");
3815 }
3816 
tex_fetch_ptrs(struct ac_nir_context * ctx,nir_tex_instr * instr,struct waterfall_context * wctx,LLVMValueRef * res_ptr,LLVMValueRef * samp_ptr)3817 static void tex_fetch_ptrs(struct ac_nir_context *ctx, nir_tex_instr *instr,
3818                            struct waterfall_context *wctx, LLVMValueRef *res_ptr,
3819                            LLVMValueRef *samp_ptr)
3820 {
3821    LLVMValueRef texture_dynamic_handle = NULL;
3822    LLVMValueRef sampler_dynamic_handle = NULL;
3823    int plane = -1;
3824 
3825    *res_ptr = NULL;
3826    *samp_ptr = NULL;
3827    for (unsigned i = 0; i < instr->num_srcs; i++) {
3828       switch (instr->src[i].src_type) {
3829       case nir_tex_src_texture_handle:
3830       case nir_tex_src_sampler_handle: {
3831          LLVMValueRef val = get_src(ctx, instr->src[i].src);
3832          if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMVectorTypeKind) {
3833             if (instr->src[i].src_type == nir_tex_src_texture_handle)
3834                *res_ptr = val;
3835             else
3836                *samp_ptr = val;
3837          } else {
3838             if (instr->src[i].src_type == nir_tex_src_texture_handle)
3839                texture_dynamic_handle = val;
3840             else
3841                sampler_dynamic_handle = val;
3842          }
3843          break;
3844       }
3845       case nir_tex_src_plane:
3846          plane = nir_src_as_int(instr->src[i].src);
3847          break;
3848       default:
3849          break;
3850       }
3851    }
3852 
3853    enum ac_descriptor_type main_descriptor =
3854       instr->sampler_dim == GLSL_SAMPLER_DIM_BUF ? AC_DESC_BUFFER : AC_DESC_IMAGE;
3855 
3856    if (plane >= 0) {
3857       assert(instr->op != nir_texop_txf_ms);
3858       assert(instr->sampler_dim != GLSL_SAMPLER_DIM_BUF);
3859 
3860       main_descriptor = AC_DESC_PLANE_0 + plane;
3861    }
3862 
3863    if (instr->op == nir_texop_fragment_mask_fetch_amd) {
3864       /* The fragment mask is fetched from the compressed
3865        * multisampled surface.
3866        */
3867       assert(ctx->ac.gfx_level < GFX11);
3868       main_descriptor = AC_DESC_FMASK;
3869    }
3870 
3871    /* descriptor handles given through nir_tex_src_{texture,sampler}_handle */
3872    if (instr->texture_non_uniform)
3873       texture_dynamic_handle = enter_waterfall(ctx, &wctx[0], texture_dynamic_handle, true);
3874 
3875    if (instr->sampler_non_uniform)
3876       sampler_dynamic_handle = enter_waterfall(ctx, &wctx[1], sampler_dynamic_handle, true);
3877 
3878    if (texture_dynamic_handle)
3879       *res_ptr = ctx->abi->load_sampler_desc(ctx->abi, texture_dynamic_handle, main_descriptor);
3880 
3881    if (sampler_dynamic_handle) {
3882       *samp_ptr = ctx->abi->load_sampler_desc(ctx->abi, sampler_dynamic_handle, AC_DESC_SAMPLER);
3883 
3884       if (ctx->abi->disable_aniso_single_level && instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
3885          *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
3886    }
3887 }
3888 
visit_tex(struct ac_nir_context * ctx,nir_tex_instr * instr)3889 static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
3890 {
3891    LLVMValueRef result = NULL;
3892    struct ac_image_args args = {0};
3893    LLVMValueRef sample_index = NULL;
3894    LLVMValueRef ddx = NULL, ddy = NULL;
3895    struct waterfall_context wctx[2] = {{{0}}};
3896 
3897    tex_fetch_ptrs(ctx, instr, wctx, &args.resource, &args.sampler);
3898 
3899    for (unsigned i = 0; i < instr->num_srcs; i++) {
3900       switch (instr->src[i].src_type) {
3901       case nir_tex_src_coord: {
3902          LLVMValueRef coord = get_src(ctx, instr->src[i].src);
3903          args.a16 = instr->src[i].src.ssa->bit_size == 16;
3904          for (unsigned chan = 0; chan < instr->coord_components; ++chan)
3905             args.coords[chan] = ac_llvm_extract_elem(&ctx->ac, coord, chan);
3906          break;
3907       }
3908       case nir_tex_src_projector:
3909          break;
3910       case nir_tex_src_comparator:
3911          if (instr->is_shadow) {
3912             args.compare = get_src(ctx, instr->src[i].src);
3913             args.compare = ac_to_float(&ctx->ac, args.compare);
3914             assert(instr->src[i].src.ssa->bit_size == 32);
3915          }
3916          break;
3917       case nir_tex_src_offset:
3918          args.offset = get_src(ctx, instr->src[i].src);
3919          /* We pack it with bit shifts, so we need it to be 32-bit. */
3920          assert(ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.offset)) == 32);
3921          break;
3922       case nir_tex_src_bias:
3923          args.bias = get_src(ctx, instr->src[i].src);
3924          assert(ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.bias)) == 32);
3925          break;
3926       case nir_tex_src_lod:
3927          if (nir_src_is_const(instr->src[i].src) && nir_src_as_uint(instr->src[i].src) == 0)
3928             args.level_zero = true;
3929          else
3930             args.lod = get_src(ctx, instr->src[i].src);
3931          break;
3932       case nir_tex_src_ms_index:
3933          sample_index = get_src(ctx, instr->src[i].src);
3934          break;
3935       case nir_tex_src_ddx:
3936          ddx = get_src(ctx, instr->src[i].src);
3937          args.g16 = instr->src[i].src.ssa->bit_size == 16;
3938          break;
3939       case nir_tex_src_ddy:
3940          ddy = get_src(ctx, instr->src[i].src);
3941          assert(LLVMTypeOf(ddy) == LLVMTypeOf(ddx));
3942          break;
3943       case nir_tex_src_min_lod:
3944          args.min_lod = get_src(ctx, instr->src[i].src);
3945          break;
3946       case nir_tex_src_texture_offset:
3947       case nir_tex_src_sampler_offset:
3948       case nir_tex_src_plane:
3949       default:
3950          break;
3951       }
3952    }
3953 
3954    if (args.offset) {
3955       /* offset for txf has been lowered in nir. */
3956       assert(instr->op != nir_texop_txf);
3957 
3958       LLVMValueRef offset[3], pack;
3959       for (unsigned chan = 0; chan < 3; ++chan)
3960          offset[chan] = ctx->ac.i32_0;
3961 
3962       unsigned num_components = ac_get_llvm_num_components(args.offset);
3963       for (unsigned chan = 0; chan < num_components; chan++) {
3964          offset[chan] = ac_llvm_extract_elem(&ctx->ac, args.offset, chan);
3965          offset[chan] =
3966             LLVMBuildAnd(ctx->ac.builder, offset[chan], LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
3967          if (chan)
3968             offset[chan] = LLVMBuildShl(ctx->ac.builder, offset[chan],
3969                                         LLVMConstInt(ctx->ac.i32, chan * 8, false), "");
3970       }
3971       pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
3972       pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
3973       args.offset = pack;
3974    }
3975 
3976    /* Section 8.23.1 (Depth Texture Comparison Mode) of the
3977     * OpenGL 4.5 spec says:
3978     *
3979     *    "If the texture’s internal format indicates a fixed-point
3980     *     depth texture, then D_t and D_ref are clamped to the
3981     *     range [0, 1]; otherwise no clamping is performed."
3982     *
3983     * TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
3984     * so the depth comparison value isn't clamped for Z16 and
3985     * Z24 anymore. Do it manually here for GFX8-9; GFX10 has
3986     * an explicitly clamped 32-bit float format.
3987     */
3988    if (args.compare && ctx->ac.gfx_level >= GFX8 && ctx->ac.gfx_level <= GFX9 &&
3989        ctx->abi->clamp_shadow_reference) {
3990       LLVMValueRef upgraded, clamped;
3991 
3992       upgraded = LLVMBuildExtractElement(ctx->ac.builder, args.sampler,
3993                                          LLVMConstInt(ctx->ac.i32, 3, false), "");
3994       upgraded = LLVMBuildLShr(ctx->ac.builder, upgraded, LLVMConstInt(ctx->ac.i32, 29, false), "");
3995       upgraded = LLVMBuildTrunc(ctx->ac.builder, upgraded, ctx->ac.i1, "");
3996       clamped = ac_build_clamp(&ctx->ac, args.compare);
3997       args.compare = LLVMBuildSelect(ctx->ac.builder, upgraded, clamped, args.compare, "");
3998    }
3999 
4000    /* pack derivatives */
4001    if (ddx || ddy) {
4002       int num_deriv_channels;
4003       switch (instr->sampler_dim) {
4004       case GLSL_SAMPLER_DIM_3D:
4005          num_deriv_channels = 3;
4006          break;
4007       case GLSL_SAMPLER_DIM_2D:
4008       case GLSL_SAMPLER_DIM_CUBE:
4009       default:
4010          num_deriv_channels = 2;
4011          break;
4012       case GLSL_SAMPLER_DIM_1D:
4013          num_deriv_channels = 1;
4014          break;
4015       }
4016 
4017       for (unsigned i = 0; i < num_deriv_channels; i++) {
4018          args.derivs[i] = ac_to_float(&ctx->ac, ac_llvm_extract_elem(&ctx->ac, ddx, i));
4019          args.derivs[num_deriv_channels + i] =
4020             ac_to_float(&ctx->ac, ac_llvm_extract_elem(&ctx->ac, ddy, i));
4021       }
4022    }
4023 
4024    /* Pack sample index */
4025    if (sample_index && (instr->op == nir_texop_txf_ms || instr->op == nir_texop_fragment_fetch_amd))
4026       args.coords[instr->coord_components] = sample_index;
4027 
4028    /* DMASK was repurposed for GATHER4. 4 components are always
4029     * returned and DMASK works like a swizzle - it selects
4030     * the component to fetch. The only valid DMASK values are
4031     * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4032     * (red,red,red,red) etc.) The ISA document doesn't mention
4033     * this.
4034     */
4035    args.dmask = 0xf;
4036    if (instr->op == nir_texop_tg4) {
4037       if (instr->is_shadow)
4038          args.dmask = 1;
4039       else
4040          args.dmask = 1 << instr->component;
4041    }
4042 
4043    if (instr->sampler_dim != GLSL_SAMPLER_DIM_BUF) {
4044       args.dim = ac_get_sampler_dim(ctx->ac.gfx_level, instr->sampler_dim, instr->is_array);
4045       args.unorm = instr->sampler_dim == GLSL_SAMPLER_DIM_RECT;
4046    }
4047 
4048    /* Adjust the number of coordinates because we only need (x,y) for 2D
4049     * multisampled images and (x,y,layer) for 2D multisampled layered
4050     * images or for multisampled input attachments.
4051     */
4052    if (instr->op == nir_texop_fragment_mask_fetch_amd) {
4053       if (args.dim == ac_image_2dmsaa) {
4054          args.dim = ac_image_2d;
4055       } else {
4056          assert(args.dim == ac_image_2darraymsaa);
4057          args.dim = ac_image_2darray;
4058       }
4059    }
4060 
4061    /* Set TRUNC_COORD=0 for textureGather(). */
4062    if (instr->op == nir_texop_tg4 && !ctx->ac.info->conformant_trunc_coord) {
4063       LLVMValueRef dword0 = LLVMBuildExtractElement(ctx->ac.builder, args.sampler, ctx->ac.i32_0, "");
4064       dword0 = LLVMBuildAnd(ctx->ac.builder, dword0, LLVMConstInt(ctx->ac.i32, C_008F30_TRUNC_COORD, 0), "");
4065       args.sampler = LLVMBuildInsertElement(ctx->ac.builder, args.sampler, dword0, ctx->ac.i32_0, "");
4066    }
4067 
4068    args.d16 = instr->def.bit_size == 16;
4069    args.tfe = instr->is_sparse;
4070 
4071    result = build_tex_intrinsic(ctx, instr, &args);
4072 
4073    LLVMValueRef code = NULL;
4074    if (instr->is_sparse) {
4075       code = ac_llvm_extract_elem(&ctx->ac, result, 4);
4076       result = ac_trim_vector(&ctx->ac, result, 4);
4077    }
4078 
4079    if (instr->is_shadow && instr->is_new_style_shadow &&
4080        instr->op != nir_texop_lod && instr->op != nir_texop_tg4)
4081       result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
4082    else if (instr->op == nir_texop_fragment_mask_fetch_amd) {
4083       /* Use 0x76543210 if the image doesn't have FMASK. */
4084       LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, args.resource, ctx->ac.v8i32, "");
4085       tmp = LLVMBuildExtractElement(ctx->ac.builder, tmp, ctx->ac.i32_1, "");
4086       tmp = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, tmp, ctx->ac.i32_0, "");
4087       result = LLVMBuildSelect(ctx->ac.builder, tmp,
4088                                LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, ""),
4089                                LLVMConstInt(ctx->ac.i32, 0x76543210, false), "");
4090    } else if (nir_tex_instr_result_size(instr) != 4)
4091       result = ac_trim_vector(&ctx->ac, result, instr->def.num_components);
4092 
4093    if (instr->is_sparse)
4094       result = ac_build_concat(&ctx->ac, result, code);
4095 
4096    if (result) {
4097       result = ac_to_integer(&ctx->ac, result);
4098 
4099       for (int i = ARRAY_SIZE(wctx); --i >= 0;) {
4100          result = exit_waterfall(ctx, wctx + i, result);
4101       }
4102 
4103       ctx->ssa_defs[instr->def.index] = result;
4104    }
4105 }
4106 
visit_phi(struct ac_nir_context * ctx,nir_phi_instr * instr)4107 static void visit_phi(struct ac_nir_context *ctx, nir_phi_instr *instr)
4108 {
4109    LLVMTypeRef type = get_def_type(ctx, &instr->def);
4110    LLVMValueRef result = LLVMBuildPhi(ctx->ac.builder, type, "");
4111 
4112    ctx->ssa_defs[instr->def.index] = result;
4113    _mesa_hash_table_insert(ctx->phis, instr, result);
4114 }
4115 
visit_post_phi(struct ac_nir_context * ctx,nir_phi_instr * instr,LLVMValueRef llvm_phi)4116 static void visit_post_phi(struct ac_nir_context *ctx, nir_phi_instr *instr, LLVMValueRef llvm_phi)
4117 {
4118    nir_foreach_phi_src (src, instr) {
4119       LLVMBasicBlockRef block = get_block(ctx, src->pred);
4120       LLVMValueRef llvm_src = get_src(ctx, src->src);
4121 
4122       LLVMAddIncoming(llvm_phi, &llvm_src, &block, 1);
4123    }
4124 }
4125 
phi_post_pass(struct ac_nir_context * ctx)4126 static void phi_post_pass(struct ac_nir_context *ctx)
4127 {
4128    hash_table_foreach(ctx->phis, entry)
4129    {
4130       visit_post_phi(ctx, (nir_phi_instr *)entry->key, (LLVMValueRef)entry->data);
4131    }
4132 }
4133 
visit_ssa_undef(struct ac_nir_context * ctx,const nir_undef_instr * instr)4134 static void visit_ssa_undef(struct ac_nir_context *ctx, const nir_undef_instr *instr)
4135 {
4136    unsigned num_components = instr->def.num_components;
4137    LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
4138 
4139    LLVMValueRef undef;
4140 
4141    if (num_components == 1)
4142       undef = LLVMGetUndef(type);
4143    else {
4144       undef = LLVMGetUndef(LLVMVectorType(type, num_components));
4145    }
4146    ctx->ssa_defs[instr->def.index] = undef;
4147 }
4148 
visit_jump(struct ac_llvm_context * ctx,const nir_jump_instr * instr)4149 static bool visit_jump(struct ac_llvm_context *ctx, const nir_jump_instr *instr)
4150 {
4151    switch (instr->type) {
4152    case nir_jump_break:
4153       ac_build_break(ctx);
4154       break;
4155    case nir_jump_continue:
4156       ac_build_continue(ctx);
4157       break;
4158    default:
4159       fprintf(stderr, "Unknown NIR jump instr: ");
4160       nir_print_instr(&instr->instr, stderr);
4161       fprintf(stderr, "\n");
4162       return false;
4163    }
4164    return true;
4165 }
4166 
4167 static bool visit_cf_list(struct ac_nir_context *ctx, struct exec_list *list);
4168 
visit_block(struct ac_nir_context * ctx,nir_block * block)4169 static bool visit_block(struct ac_nir_context *ctx, nir_block *block)
4170 {
4171    LLVMBasicBlockRef blockref = LLVMGetInsertBlock(ctx->ac.builder);
4172    LLVMValueRef first = LLVMGetFirstInstruction(blockref);
4173    if (first) {
4174       /* ac_branch_exited() might have already inserted non-phis */
4175       LLVMPositionBuilderBefore(ctx->ac.builder, LLVMGetFirstInstruction(blockref));
4176    }
4177 
4178    nir_foreach_phi(phi, block) {
4179       visit_phi(ctx, phi);
4180    }
4181 
4182    LLVMPositionBuilderAtEnd(ctx->ac.builder, blockref);
4183 
4184    nir_foreach_instr (instr, block) {
4185       switch (instr->type) {
4186       case nir_instr_type_alu:
4187          if (!visit_alu(ctx, nir_instr_as_alu(instr)))
4188             return false;
4189          break;
4190       case nir_instr_type_load_const:
4191          if (!visit_load_const(ctx, nir_instr_as_load_const(instr)))
4192             return false;
4193          break;
4194       case nir_instr_type_intrinsic:
4195          if (!visit_intrinsic(ctx, nir_instr_as_intrinsic(instr)))
4196             return false;
4197          break;
4198       case nir_instr_type_tex:
4199          visit_tex(ctx, nir_instr_as_tex(instr));
4200          break;
4201       case nir_instr_type_phi:
4202          break;
4203       case nir_instr_type_undef:
4204          visit_ssa_undef(ctx, nir_instr_as_undef(instr));
4205          break;
4206       case nir_instr_type_jump:
4207          if (!visit_jump(&ctx->ac, nir_instr_as_jump(instr)))
4208             return false;
4209          break;
4210       case nir_instr_type_deref:
4211          assert (!nir_deref_mode_is_one_of(nir_instr_as_deref(instr),
4212                                            nir_var_mem_shared | nir_var_mem_global));
4213          break;
4214       default:
4215          fprintf(stderr, "Unknown NIR instr type: ");
4216          nir_print_instr(instr, stderr);
4217          fprintf(stderr, "\n");
4218          return false;
4219       }
4220    }
4221 
4222    _mesa_hash_table_insert(ctx->defs, block, LLVMGetInsertBlock(ctx->ac.builder));
4223 
4224    return true;
4225 }
4226 
visit_if(struct ac_nir_context * ctx,nir_if * if_stmt)4227 static bool visit_if(struct ac_nir_context *ctx, nir_if *if_stmt)
4228 {
4229    LLVMValueRef value = get_src(ctx, if_stmt->condition);
4230 
4231    nir_block *then_block = (nir_block *)exec_list_get_head(&if_stmt->then_list);
4232 
4233    ac_build_ifcc(&ctx->ac, value, then_block->index);
4234 
4235    if (!visit_cf_list(ctx, &if_stmt->then_list))
4236       return false;
4237 
4238    if (!exec_list_is_empty(&if_stmt->else_list)) {
4239       nir_block *else_block = (nir_block *)exec_list_get_head(&if_stmt->else_list);
4240 
4241       ac_build_else(&ctx->ac, else_block->index);
4242       if (!visit_cf_list(ctx, &if_stmt->else_list))
4243          return false;
4244    }
4245 
4246    ac_build_endif(&ctx->ac, then_block->index);
4247    return true;
4248 }
4249 
visit_loop(struct ac_nir_context * ctx,nir_loop * loop)4250 static bool visit_loop(struct ac_nir_context *ctx, nir_loop *loop)
4251 {
4252    assert(!nir_loop_has_continue_construct(loop));
4253    nir_block *first_loop_block = (nir_block *)exec_list_get_head(&loop->body);
4254 
4255    ac_build_bgnloop(&ctx->ac, first_loop_block->index);
4256 
4257    if (!visit_cf_list(ctx, &loop->body))
4258       return false;
4259 
4260    ac_build_endloop(&ctx->ac, first_loop_block->index);
4261    return true;
4262 }
4263 
visit_cf_list(struct ac_nir_context * ctx,struct exec_list * list)4264 static bool visit_cf_list(struct ac_nir_context *ctx, struct exec_list *list)
4265 {
4266    foreach_list_typed(nir_cf_node, node, node, list)
4267    {
4268       switch (node->type) {
4269       case nir_cf_node_block:
4270          if (!visit_block(ctx, nir_cf_node_as_block(node)))
4271             return false;
4272          break;
4273 
4274       case nir_cf_node_if:
4275          if (!visit_if(ctx, nir_cf_node_as_if(node)))
4276             return false;
4277          break;
4278 
4279       case nir_cf_node_loop:
4280          if (!visit_loop(ctx, nir_cf_node_as_loop(node)))
4281             return false;
4282          break;
4283 
4284       default:
4285          return false;
4286       }
4287    }
4288    return true;
4289 }
4290 
setup_scratch(struct ac_nir_context * ctx,struct nir_shader * shader)4291 static void setup_scratch(struct ac_nir_context *ctx, struct nir_shader *shader)
4292 {
4293    if (shader->scratch_size == 0)
4294       return;
4295 
4296    LLVMTypeRef type = LLVMArrayType(ctx->ac.i8, shader->scratch_size);
4297    ctx->scratch = (struct ac_llvm_pointer) {
4298       .value = ac_build_alloca_undef(&ctx->ac, type, "scratch"),
4299       .pointee_type = type
4300    };
4301 }
4302 
setup_constant_data(struct ac_nir_context * ctx,struct nir_shader * shader)4303 static void setup_constant_data(struct ac_nir_context *ctx, struct nir_shader *shader)
4304 {
4305    if (!shader->constant_data)
4306       return;
4307 
4308    LLVMValueRef data = LLVMConstStringInContext(ctx->ac.context, shader->constant_data,
4309                                                 shader->constant_data_size, true);
4310    LLVMTypeRef type = LLVMArrayType(ctx->ac.i8, shader->constant_data_size);
4311    LLVMValueRef global =
4312       LLVMAddGlobalInAddressSpace(ctx->ac.module, type, "const_data", AC_ADDR_SPACE_CONST);
4313 
4314    LLVMSetInitializer(global, data);
4315    LLVMSetGlobalConstant(global, true);
4316    LLVMSetVisibility(global, LLVMHiddenVisibility);
4317    ctx->constant_data = (struct ac_llvm_pointer) {
4318       .value = global,
4319       .pointee_type = type
4320    };
4321 }
4322 
setup_shared(struct ac_nir_context * ctx,struct nir_shader * nir)4323 static void setup_shared(struct ac_nir_context *ctx, struct nir_shader *nir)
4324 {
4325    if (ctx->ac.lds.value)
4326       return;
4327 
4328    LLVMTypeRef type = LLVMArrayType(ctx->ac.i8, nir->info.shared_size);
4329 
4330    LLVMValueRef lds =
4331       LLVMAddGlobalInAddressSpace(ctx->ac.module, type, "compute_lds", AC_ADDR_SPACE_LDS);
4332    LLVMSetAlignment(lds, 64 * 1024);
4333 
4334    ctx->ac.lds = (struct ac_llvm_pointer) {
4335       .value = lds,
4336       .pointee_type = type
4337    };
4338 }
4339 
setup_gds(struct ac_nir_context * ctx,nir_function_impl * impl)4340 static void setup_gds(struct ac_nir_context *ctx, nir_function_impl *impl)
4341 {
4342    bool has_gds_atomic = false;
4343 
4344    if (ctx->ac.gfx_level >= GFX10 &&
4345        (ctx->stage == MESA_SHADER_VERTEX ||
4346         ctx->stage == MESA_SHADER_TESS_EVAL ||
4347         ctx->stage == MESA_SHADER_GEOMETRY)) {
4348 
4349       nir_foreach_block(block, impl) {
4350          nir_foreach_instr(instr, block) {
4351             if (instr->type != nir_instr_type_intrinsic)
4352                continue;
4353 
4354             nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
4355             has_gds_atomic |= intrin->intrinsic == nir_intrinsic_gds_atomic_add_amd;
4356          }
4357       }
4358    }
4359 
4360    unsigned gds_size = has_gds_atomic ? 0x100 : 0;
4361 
4362    if (gds_size)
4363       ac_llvm_add_target_dep_function_attr(ctx->main_function, "amdgpu-gds-size", gds_size);
4364 }
4365 
ac_nir_translate(struct ac_llvm_context * ac,struct ac_shader_abi * abi,const struct ac_shader_args * args,struct nir_shader * nir)4366 bool ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
4367                       const struct ac_shader_args *args, struct nir_shader *nir)
4368 {
4369    struct ac_nir_context ctx = {0};
4370    struct nir_function *func;
4371 
4372    ctx.ac = *ac;
4373    ctx.abi = abi;
4374    ctx.args = args;
4375 
4376    ctx.stage = nir->info.stage;
4377    ctx.info = &nir->info;
4378 
4379    ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
4380 
4381    ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
4382    ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
4383 
4384    if (ctx.abi->kill_ps_if_inf_interp)
4385       ctx.verified_interp =
4386          _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal);
4387 
4388    func = (struct nir_function *)exec_list_get_head(&nir->functions);
4389 
4390    nir_index_ssa_defs(func->impl);
4391    ctx.ssa_defs = calloc(func->impl->ssa_alloc, sizeof(LLVMValueRef));
4392 
4393    setup_scratch(&ctx, nir);
4394    setup_constant_data(&ctx, nir);
4395    setup_gds(&ctx, func->impl);
4396 
4397    if (gl_shader_stage_is_compute(nir->info.stage))
4398       setup_shared(&ctx, nir);
4399 
4400    if (!visit_cf_list(&ctx, &func->impl->body))
4401       return false;
4402 
4403    phi_post_pass(&ctx);
4404 
4405    free(ctx.ssa_defs);
4406    ralloc_free(ctx.defs);
4407    ralloc_free(ctx.phis);
4408    if (ctx.abi->kill_ps_if_inf_interp)
4409       ralloc_free(ctx.verified_interp);
4410 
4411    return true;
4412 }
4413 
4414 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
ac_fixup_ls_hs_input_vgprs(struct ac_llvm_context * ac,struct ac_shader_abi * abi,const struct ac_shader_args * args)4415 void ac_fixup_ls_hs_input_vgprs(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
4416                                 const struct ac_shader_args *args)
4417 {
4418    LLVMValueRef count = ac_unpack_param(ac, ac_get_arg(ac, args->merged_wave_info), 8, 8);
4419    LLVMValueRef hs_empty = LLVMBuildICmp(ac->builder, LLVMIntEQ, count, ac->i32_0, "");
4420 
4421    abi->instance_id =
4422       LLVMBuildSelect(ac->builder, hs_empty, ac_get_arg(ac, args->vertex_id),
4423                       abi->instance_id, "");
4424 
4425    abi->vs_rel_patch_id =
4426       LLVMBuildSelect(ac->builder, hs_empty, ac_get_arg(ac, args->tcs_rel_ids),
4427                       abi->vs_rel_patch_id, "");
4428 
4429    abi->vertex_id =
4430       LLVMBuildSelect(ac->builder, hs_empty, ac_get_arg(ac, args->tcs_patch_id),
4431                       abi->vertex_id, "");
4432 }
4433