1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
11 *
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
19 *
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
22 * of the Software.
23 *
24 */
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
27
28 #include <llvm-c/Core.h>
29
30 #include "c11/threads.h"
31
32 #include <assert.h>
33 #include <stdio.h>
34
35 #include "ac_llvm_util.h"
36 #include "ac_exp_param.h"
37 #include "util/bitscan.h"
38 #include "util/macros.h"
39 #include "util/u_atomic.h"
40 #include "sid.h"
41
42 #include "shader_enums.h"
43
44 #define AC_LLVM_INITIAL_CF_DEPTH 4
45
46 /* Data for if/else/endif and bgnloop/endloop control flow structures.
47 */
48 struct ac_llvm_flow {
49 /* Loop exit or next part of if/else/endif. */
50 LLVMBasicBlockRef next_block;
51 LLVMBasicBlockRef loop_entry_block;
52 };
53
54 /* Initialize module-independent parts of the context.
55 *
56 * The caller is responsible for initializing ctx::module and ctx::builder.
57 */
58 void
ac_llvm_context_init(struct ac_llvm_context * ctx,LLVMContextRef context,enum chip_class chip_class,enum radeon_family family)59 ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context,
60 enum chip_class chip_class, enum radeon_family family)
61 {
62 LLVMValueRef args[1];
63
64 ctx->chip_class = chip_class;
65 ctx->family = family;
66
67 ctx->context = context;
68 ctx->module = NULL;
69 ctx->builder = NULL;
70
71 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
72 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
73 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
74 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
75 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
76 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
77 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
78 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
79 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
80 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
81 ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
82 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
83 ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
84 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
85 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
86
87 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
88 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
89 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
90 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
91 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
92 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
93 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
94 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
95
96 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
97 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
98
99 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
100 "range", 5);
101
102 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
103 "invariant.load", 14);
104
105 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
106
107 args[0] = LLVMConstReal(ctx->f32, 2.5);
108 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
109
110 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
111 "amdgpu.uniform", 14);
112
113 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
114 }
115
116 void
ac_llvm_context_dispose(struct ac_llvm_context * ctx)117 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
118 {
119 free(ctx->flow);
120 ctx->flow = NULL;
121 ctx->flow_depth_max = 0;
122 }
123
124 int
ac_get_llvm_num_components(LLVMValueRef value)125 ac_get_llvm_num_components(LLVMValueRef value)
126 {
127 LLVMTypeRef type = LLVMTypeOf(value);
128 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
129 ? LLVMGetVectorSize(type)
130 : 1;
131 return num_components;
132 }
133
134 LLVMValueRef
ac_llvm_extract_elem(struct ac_llvm_context * ac,LLVMValueRef value,int index)135 ac_llvm_extract_elem(struct ac_llvm_context *ac,
136 LLVMValueRef value,
137 int index)
138 {
139 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
140 assert(index == 0);
141 return value;
142 }
143
144 return LLVMBuildExtractElement(ac->builder, value,
145 LLVMConstInt(ac->i32, index, false), "");
146 }
147
148 unsigned
ac_get_type_size(LLVMTypeRef type)149 ac_get_type_size(LLVMTypeRef type)
150 {
151 LLVMTypeKind kind = LLVMGetTypeKind(type);
152
153 switch (kind) {
154 case LLVMIntegerTypeKind:
155 return LLVMGetIntTypeWidth(type) / 8;
156 case LLVMFloatTypeKind:
157 return 4;
158 case LLVMDoubleTypeKind:
159 case LLVMPointerTypeKind:
160 return 8;
161 case LLVMVectorTypeKind:
162 return LLVMGetVectorSize(type) *
163 ac_get_type_size(LLVMGetElementType(type));
164 case LLVMArrayTypeKind:
165 return LLVMGetArrayLength(type) *
166 ac_get_type_size(LLVMGetElementType(type));
167 default:
168 assert(0);
169 return 0;
170 }
171 }
172
to_integer_type_scalar(struct ac_llvm_context * ctx,LLVMTypeRef t)173 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
174 {
175 if (t == ctx->f16 || t == ctx->i16)
176 return ctx->i16;
177 else if (t == ctx->f32 || t == ctx->i32)
178 return ctx->i32;
179 else if (t == ctx->f64 || t == ctx->i64)
180 return ctx->i64;
181 else
182 unreachable("Unhandled integer size");
183 }
184
185 LLVMTypeRef
ac_to_integer_type(struct ac_llvm_context * ctx,LLVMTypeRef t)186 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
187 {
188 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
189 LLVMTypeRef elem_type = LLVMGetElementType(t);
190 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
191 LLVMGetVectorSize(t));
192 }
193 return to_integer_type_scalar(ctx, t);
194 }
195
196 LLVMValueRef
ac_to_integer(struct ac_llvm_context * ctx,LLVMValueRef v)197 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
198 {
199 LLVMTypeRef type = LLVMTypeOf(v);
200 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
201 }
202
to_float_type_scalar(struct ac_llvm_context * ctx,LLVMTypeRef t)203 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
204 {
205 if (t == ctx->i16 || t == ctx->f16)
206 return ctx->f16;
207 else if (t == ctx->i32 || t == ctx->f32)
208 return ctx->f32;
209 else if (t == ctx->i64 || t == ctx->f64)
210 return ctx->f64;
211 else
212 unreachable("Unhandled float size");
213 }
214
215 LLVMTypeRef
ac_to_float_type(struct ac_llvm_context * ctx,LLVMTypeRef t)216 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
217 {
218 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
219 LLVMTypeRef elem_type = LLVMGetElementType(t);
220 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
221 LLVMGetVectorSize(t));
222 }
223 return to_float_type_scalar(ctx, t);
224 }
225
226 LLVMValueRef
ac_to_float(struct ac_llvm_context * ctx,LLVMValueRef v)227 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
228 {
229 LLVMTypeRef type = LLVMTypeOf(v);
230 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
231 }
232
233
234 LLVMValueRef
ac_build_intrinsic(struct ac_llvm_context * ctx,const char * name,LLVMTypeRef return_type,LLVMValueRef * params,unsigned param_count,unsigned attrib_mask)235 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
236 LLVMTypeRef return_type, LLVMValueRef *params,
237 unsigned param_count, unsigned attrib_mask)
238 {
239 LLVMValueRef function, call;
240 bool set_callsite_attrs = HAVE_LLVM >= 0x0400 &&
241 !(attrib_mask & AC_FUNC_ATTR_LEGACY);
242
243 function = LLVMGetNamedFunction(ctx->module, name);
244 if (!function) {
245 LLVMTypeRef param_types[32], function_type;
246 unsigned i;
247
248 assert(param_count <= 32);
249
250 for (i = 0; i < param_count; ++i) {
251 assert(params[i]);
252 param_types[i] = LLVMTypeOf(params[i]);
253 }
254 function_type =
255 LLVMFunctionType(return_type, param_types, param_count, 0);
256 function = LLVMAddFunction(ctx->module, name, function_type);
257
258 LLVMSetFunctionCallConv(function, LLVMCCallConv);
259 LLVMSetLinkage(function, LLVMExternalLinkage);
260
261 if (!set_callsite_attrs)
262 ac_add_func_attributes(ctx->context, function, attrib_mask);
263 }
264
265 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
266 if (set_callsite_attrs)
267 ac_add_func_attributes(ctx->context, call, attrib_mask);
268 return call;
269 }
270
271 /**
272 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
273 * intrinsic names).
274 */
ac_build_type_name_for_intr(LLVMTypeRef type,char * buf,unsigned bufsize)275 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
276 {
277 LLVMTypeRef elem_type = type;
278
279 assert(bufsize >= 8);
280
281 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
282 int ret = snprintf(buf, bufsize, "v%u",
283 LLVMGetVectorSize(type));
284 if (ret < 0) {
285 char *type_name = LLVMPrintTypeToString(type);
286 fprintf(stderr, "Error building type name for: %s\n",
287 type_name);
288 return;
289 }
290 elem_type = LLVMGetElementType(type);
291 buf += ret;
292 bufsize -= ret;
293 }
294 switch (LLVMGetTypeKind(elem_type)) {
295 default: break;
296 case LLVMIntegerTypeKind:
297 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
298 break;
299 case LLVMFloatTypeKind:
300 snprintf(buf, bufsize, "f32");
301 break;
302 case LLVMDoubleTypeKind:
303 snprintf(buf, bufsize, "f64");
304 break;
305 }
306 }
307
308 /**
309 * Helper function that builds an LLVM IR PHI node and immediately adds
310 * incoming edges.
311 */
312 LLVMValueRef
ac_build_phi(struct ac_llvm_context * ctx,LLVMTypeRef type,unsigned count_incoming,LLVMValueRef * values,LLVMBasicBlockRef * blocks)313 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
314 unsigned count_incoming, LLVMValueRef *values,
315 LLVMBasicBlockRef *blocks)
316 {
317 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
318 LLVMAddIncoming(phi, values, blocks, count_incoming);
319 return phi;
320 }
321
322 /* Prevent optimizations (at least of memory accesses) across the current
323 * point in the program by emitting empty inline assembly that is marked as
324 * having side effects.
325 *
326 * Optionally, a value can be passed through the inline assembly to prevent
327 * LLVM from hoisting calls to ReadNone functions.
328 */
329 void
ac_build_optimization_barrier(struct ac_llvm_context * ctx,LLVMValueRef * pvgpr)330 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
331 LLVMValueRef *pvgpr)
332 {
333 static int counter = 0;
334
335 LLVMBuilderRef builder = ctx->builder;
336 char code[16];
337
338 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
339
340 if (!pvgpr) {
341 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
342 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
343 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
344 } else {
345 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
346 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
347 LLVMValueRef vgpr = *pvgpr;
348 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
349 unsigned vgpr_size = ac_get_type_size(vgpr_type);
350 LLVMValueRef vgpr0;
351
352 assert(vgpr_size % 4 == 0);
353
354 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
355 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
356 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
357 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
358 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
359
360 *pvgpr = vgpr;
361 }
362 }
363
364 LLVMValueRef
ac_build_ballot(struct ac_llvm_context * ctx,LLVMValueRef value)365 ac_build_ballot(struct ac_llvm_context *ctx,
366 LLVMValueRef value)
367 {
368 LLVMValueRef args[3] = {
369 value,
370 ctx->i32_0,
371 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
372 };
373
374 /* We currently have no other way to prevent LLVM from lifting the icmp
375 * calls to a dominating basic block.
376 */
377 ac_build_optimization_barrier(ctx, &args[0]);
378
379 if (LLVMTypeOf(args[0]) != ctx->i32)
380 args[0] = LLVMBuildBitCast(ctx->builder, args[0], ctx->i32, "");
381
382 return ac_build_intrinsic(ctx,
383 "llvm.amdgcn.icmp.i32",
384 ctx->i64, args, 3,
385 AC_FUNC_ATTR_NOUNWIND |
386 AC_FUNC_ATTR_READNONE |
387 AC_FUNC_ATTR_CONVERGENT);
388 }
389
390 LLVMValueRef
ac_build_vote_all(struct ac_llvm_context * ctx,LLVMValueRef value)391 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
392 {
393 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
394 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
395 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
396 }
397
398 LLVMValueRef
ac_build_vote_any(struct ac_llvm_context * ctx,LLVMValueRef value)399 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
400 {
401 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
402 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
403 LLVMConstInt(ctx->i64, 0, 0), "");
404 }
405
406 LLVMValueRef
ac_build_vote_eq(struct ac_llvm_context * ctx,LLVMValueRef value)407 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
408 {
409 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
410 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
411
412 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
413 vote_set, active_set, "");
414 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
415 vote_set,
416 LLVMConstInt(ctx->i64, 0, 0), "");
417 return LLVMBuildOr(ctx->builder, all, none, "");
418 }
419
420 LLVMValueRef
ac_build_varying_gather_values(struct ac_llvm_context * ctx,LLVMValueRef * values,unsigned value_count,unsigned component)421 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
422 unsigned value_count, unsigned component)
423 {
424 LLVMValueRef vec = NULL;
425
426 if (value_count == 1) {
427 return values[component];
428 } else if (!value_count)
429 unreachable("value_count is 0");
430
431 for (unsigned i = component; i < value_count + component; i++) {
432 LLVMValueRef value = values[i];
433
434 if (i == component)
435 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
436 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
437 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
438 }
439 return vec;
440 }
441
442 LLVMValueRef
ac_build_gather_values_extended(struct ac_llvm_context * ctx,LLVMValueRef * values,unsigned value_count,unsigned value_stride,bool load,bool always_vector)443 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
444 LLVMValueRef *values,
445 unsigned value_count,
446 unsigned value_stride,
447 bool load,
448 bool always_vector)
449 {
450 LLVMBuilderRef builder = ctx->builder;
451 LLVMValueRef vec = NULL;
452 unsigned i;
453
454 if (value_count == 1 && !always_vector) {
455 if (load)
456 return LLVMBuildLoad(builder, values[0], "");
457 return values[0];
458 } else if (!value_count)
459 unreachable("value_count is 0");
460
461 for (i = 0; i < value_count; i++) {
462 LLVMValueRef value = values[i * value_stride];
463 if (load)
464 value = LLVMBuildLoad(builder, value, "");
465
466 if (!i)
467 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
468 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
469 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
470 }
471 return vec;
472 }
473
474 LLVMValueRef
ac_build_gather_values(struct ac_llvm_context * ctx,LLVMValueRef * values,unsigned value_count)475 ac_build_gather_values(struct ac_llvm_context *ctx,
476 LLVMValueRef *values,
477 unsigned value_count)
478 {
479 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
480 }
481
482 LLVMValueRef
ac_build_fdiv(struct ac_llvm_context * ctx,LLVMValueRef num,LLVMValueRef den)483 ac_build_fdiv(struct ac_llvm_context *ctx,
484 LLVMValueRef num,
485 LLVMValueRef den)
486 {
487 LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
488
489 /* Use v_rcp_f32 instead of precise division. */
490 if (!LLVMIsConstant(ret))
491 LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
492 return ret;
493 }
494
495 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
496 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
497 * already multiplied by two. id is the cube face number.
498 */
499 struct cube_selection_coords {
500 LLVMValueRef stc[2];
501 LLVMValueRef ma;
502 LLVMValueRef id;
503 };
504
505 static void
build_cube_intrinsic(struct ac_llvm_context * ctx,LLVMValueRef in[3],struct cube_selection_coords * out)506 build_cube_intrinsic(struct ac_llvm_context *ctx,
507 LLVMValueRef in[3],
508 struct cube_selection_coords *out)
509 {
510 LLVMTypeRef f32 = ctx->f32;
511
512 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
513 f32, in, 3, AC_FUNC_ATTR_READNONE);
514 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
515 f32, in, 3, AC_FUNC_ATTR_READNONE);
516 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
517 f32, in, 3, AC_FUNC_ATTR_READNONE);
518 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
519 f32, in, 3, AC_FUNC_ATTR_READNONE);
520 }
521
522 /**
523 * Build a manual selection sequence for cube face sc/tc coordinates and
524 * major axis vector (multiplied by 2 for consistency) for the given
525 * vec3 \p coords, for the face implied by \p selcoords.
526 *
527 * For the major axis, we always adjust the sign to be in the direction of
528 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
529 * the selcoords major axis.
530 */
build_cube_select(struct ac_llvm_context * ctx,const struct cube_selection_coords * selcoords,const LLVMValueRef * coords,LLVMValueRef * out_st,LLVMValueRef * out_ma)531 static void build_cube_select(struct ac_llvm_context *ctx,
532 const struct cube_selection_coords *selcoords,
533 const LLVMValueRef *coords,
534 LLVMValueRef *out_st,
535 LLVMValueRef *out_ma)
536 {
537 LLVMBuilderRef builder = ctx->builder;
538 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
539 LLVMValueRef is_ma_positive;
540 LLVMValueRef sgn_ma;
541 LLVMValueRef is_ma_z, is_not_ma_z;
542 LLVMValueRef is_ma_y;
543 LLVMValueRef is_ma_x;
544 LLVMValueRef sgn;
545 LLVMValueRef tmp;
546
547 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
548 selcoords->ma, LLVMConstReal(f32, 0.0), "");
549 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
550 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
551
552 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
553 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
554 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
555 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
556 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
557
558 /* Select sc */
559 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
560 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
561 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
562 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
563 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
564
565 /* Select tc */
566 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
567 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
568 LLVMConstReal(f32, -1.0), "");
569 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
570
571 /* Select ma */
572 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
573 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
574 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
575 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
576 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
577 }
578
579 void
ac_prepare_cube_coords(struct ac_llvm_context * ctx,bool is_deriv,bool is_array,bool is_lod,LLVMValueRef * coords_arg,LLVMValueRef * derivs_arg)580 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
581 bool is_deriv, bool is_array, bool is_lod,
582 LLVMValueRef *coords_arg,
583 LLVMValueRef *derivs_arg)
584 {
585
586 LLVMBuilderRef builder = ctx->builder;
587 struct cube_selection_coords selcoords;
588 LLVMValueRef coords[3];
589 LLVMValueRef invma;
590
591 if (is_array && !is_lod) {
592 LLVMValueRef tmp = coords_arg[3];
593 tmp = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &tmp, 1, 0);
594
595 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
596 *
597 * "For Array forms, the array layer used will be
598 *
599 * max(0, min(d−1, floor(layer+0.5)))
600 *
601 * where d is the depth of the texture array and layer
602 * comes from the component indicated in the tables below.
603 * Workaroudn for an issue where the layer is taken from a
604 * helper invocation which happens to fall on a different
605 * layer due to extrapolation."
606 *
607 * VI and earlier attempt to implement this in hardware by
608 * clamping the value of coords[2] = (8 * layer) + face.
609 * Unfortunately, this means that the we end up with the wrong
610 * face when clamping occurs.
611 *
612 * Clamp the layer earlier to work around the issue.
613 */
614 if (ctx->chip_class <= VI) {
615 LLVMValueRef ge0;
616 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
617 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
618 }
619
620 coords_arg[3] = tmp;
621 }
622
623 build_cube_intrinsic(ctx, coords_arg, &selcoords);
624
625 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
626 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
627 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
628
629 for (int i = 0; i < 2; ++i)
630 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
631
632 coords[2] = selcoords.id;
633
634 if (is_deriv && derivs_arg) {
635 LLVMValueRef derivs[4];
636 int axis;
637
638 /* Convert cube derivatives to 2D derivatives. */
639 for (axis = 0; axis < 2; axis++) {
640 LLVMValueRef deriv_st[2];
641 LLVMValueRef deriv_ma;
642
643 /* Transform the derivative alongside the texture
644 * coordinate. Mathematically, the correct formula is
645 * as follows. Assume we're projecting onto the +Z face
646 * and denote by dx/dh the derivative of the (original)
647 * X texture coordinate with respect to horizontal
648 * window coordinates. The projection onto the +Z face
649 * plane is:
650 *
651 * f(x,z) = x/z
652 *
653 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
654 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
655 *
656 * This motivatives the implementation below.
657 *
658 * Whether this actually gives the expected results for
659 * apps that might feed in derivatives obtained via
660 * finite differences is anyone's guess. The OpenGL spec
661 * seems awfully quiet about how textureGrad for cube
662 * maps should be handled.
663 */
664 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
665 deriv_st, &deriv_ma);
666
667 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
668
669 for (int i = 0; i < 2; ++i)
670 derivs[axis * 2 + i] =
671 LLVMBuildFSub(builder,
672 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
673 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
674 }
675
676 memcpy(derivs_arg, derivs, sizeof(derivs));
677 }
678
679 /* Shift the texture coordinate. This must be applied after the
680 * derivative calculation.
681 */
682 for (int i = 0; i < 2; ++i)
683 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
684
685 if (is_array) {
686 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
687 /* coords_arg.w component - array_index for cube arrays */
688 LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
689 coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
690 }
691
692 memcpy(coords_arg, coords, sizeof(coords));
693 }
694
695
696 LLVMValueRef
ac_build_fs_interp(struct ac_llvm_context * ctx,LLVMValueRef llvm_chan,LLVMValueRef attr_number,LLVMValueRef params,LLVMValueRef i,LLVMValueRef j)697 ac_build_fs_interp(struct ac_llvm_context *ctx,
698 LLVMValueRef llvm_chan,
699 LLVMValueRef attr_number,
700 LLVMValueRef params,
701 LLVMValueRef i,
702 LLVMValueRef j)
703 {
704 LLVMValueRef args[5];
705 LLVMValueRef p1;
706
707 if (HAVE_LLVM < 0x0400) {
708 LLVMValueRef ij[2];
709 ij[0] = LLVMBuildBitCast(ctx->builder, i, ctx->i32, "");
710 ij[1] = LLVMBuildBitCast(ctx->builder, j, ctx->i32, "");
711
712 args[0] = llvm_chan;
713 args[1] = attr_number;
714 args[2] = params;
715 args[3] = ac_build_gather_values(ctx, ij, 2);
716 return ac_build_intrinsic(ctx, "llvm.SI.fs.interp",
717 ctx->f32, args, 4,
718 AC_FUNC_ATTR_READNONE);
719 }
720
721 args[0] = i;
722 args[1] = llvm_chan;
723 args[2] = attr_number;
724 args[3] = params;
725
726 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
727 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
728
729 args[0] = p1;
730 args[1] = j;
731 args[2] = llvm_chan;
732 args[3] = attr_number;
733 args[4] = params;
734
735 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
736 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
737 }
738
739 LLVMValueRef
ac_build_fs_interp_mov(struct ac_llvm_context * ctx,LLVMValueRef parameter,LLVMValueRef llvm_chan,LLVMValueRef attr_number,LLVMValueRef params)740 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
741 LLVMValueRef parameter,
742 LLVMValueRef llvm_chan,
743 LLVMValueRef attr_number,
744 LLVMValueRef params)
745 {
746 LLVMValueRef args[4];
747 if (HAVE_LLVM < 0x0400) {
748 args[0] = llvm_chan;
749 args[1] = attr_number;
750 args[2] = params;
751
752 return ac_build_intrinsic(ctx,
753 "llvm.SI.fs.constant",
754 ctx->f32, args, 3,
755 AC_FUNC_ATTR_READNONE);
756 }
757
758 args[0] = parameter;
759 args[1] = llvm_chan;
760 args[2] = attr_number;
761 args[3] = params;
762
763 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
764 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
765 }
766
767 LLVMValueRef
ac_build_gep0(struct ac_llvm_context * ctx,LLVMValueRef base_ptr,LLVMValueRef index)768 ac_build_gep0(struct ac_llvm_context *ctx,
769 LLVMValueRef base_ptr,
770 LLVMValueRef index)
771 {
772 LLVMValueRef indices[2] = {
773 LLVMConstInt(ctx->i32, 0, 0),
774 index,
775 };
776 return LLVMBuildGEP(ctx->builder, base_ptr,
777 indices, 2, "");
778 }
779
780 void
ac_build_indexed_store(struct ac_llvm_context * ctx,LLVMValueRef base_ptr,LLVMValueRef index,LLVMValueRef value)781 ac_build_indexed_store(struct ac_llvm_context *ctx,
782 LLVMValueRef base_ptr, LLVMValueRef index,
783 LLVMValueRef value)
784 {
785 LLVMBuildStore(ctx->builder, value,
786 ac_build_gep0(ctx, base_ptr, index));
787 }
788
789 /**
790 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
791 * It's equivalent to doing a load from &base_ptr[index].
792 *
793 * \param base_ptr Where the array starts.
794 * \param index The element index into the array.
795 * \param uniform Whether the base_ptr and index can be assumed to be
796 * dynamically uniform (i.e. load to an SGPR)
797 * \param invariant Whether the load is invariant (no other opcodes affect it)
798 */
799 static LLVMValueRef
ac_build_load_custom(struct ac_llvm_context * ctx,LLVMValueRef base_ptr,LLVMValueRef index,bool uniform,bool invariant)800 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
801 LLVMValueRef index, bool uniform, bool invariant)
802 {
803 LLVMValueRef pointer, result;
804
805 pointer = ac_build_gep0(ctx, base_ptr, index);
806 if (uniform)
807 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
808 result = LLVMBuildLoad(ctx->builder, pointer, "");
809 if (invariant)
810 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
811 return result;
812 }
813
ac_build_load(struct ac_llvm_context * ctx,LLVMValueRef base_ptr,LLVMValueRef index)814 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
815 LLVMValueRef index)
816 {
817 return ac_build_load_custom(ctx, base_ptr, index, false, false);
818 }
819
ac_build_load_invariant(struct ac_llvm_context * ctx,LLVMValueRef base_ptr,LLVMValueRef index)820 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
821 LLVMValueRef base_ptr, LLVMValueRef index)
822 {
823 return ac_build_load_custom(ctx, base_ptr, index, false, true);
824 }
825
ac_build_load_to_sgpr(struct ac_llvm_context * ctx,LLVMValueRef base_ptr,LLVMValueRef index)826 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
827 LLVMValueRef base_ptr, LLVMValueRef index)
828 {
829 return ac_build_load_custom(ctx, base_ptr, index, true, true);
830 }
831
832 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
833 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
834 * or v4i32 (num_channels=3,4).
835 */
836 void
ac_build_buffer_store_dword(struct ac_llvm_context * ctx,LLVMValueRef rsrc,LLVMValueRef vdata,unsigned num_channels,LLVMValueRef voffset,LLVMValueRef soffset,unsigned inst_offset,bool glc,bool slc,bool writeonly_memory,bool swizzle_enable_hint)837 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
838 LLVMValueRef rsrc,
839 LLVMValueRef vdata,
840 unsigned num_channels,
841 LLVMValueRef voffset,
842 LLVMValueRef soffset,
843 unsigned inst_offset,
844 bool glc,
845 bool slc,
846 bool writeonly_memory,
847 bool swizzle_enable_hint)
848 {
849 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
850 * (voffset is swizzled, but soffset isn't swizzled).
851 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
852 */
853 if (!swizzle_enable_hint) {
854 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
855 * intrinsics. */
856 if (num_channels == 3) {
857 LLVMValueRef v[3], v01;
858
859 for (int i = 0; i < 3; i++) {
860 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
861 LLVMConstInt(ctx->i32, i, 0), "");
862 }
863 v01 = ac_build_gather_values(ctx, v, 2);
864
865 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
866 soffset, inst_offset, glc, slc,
867 writeonly_memory, swizzle_enable_hint);
868 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
869 soffset, inst_offset + 8,
870 glc, slc,
871 writeonly_memory, swizzle_enable_hint);
872 return;
873 }
874
875 unsigned func = CLAMP(num_channels, 1, 3) - 1;
876 static const char *types[] = {"f32", "v2f32", "v4f32"};
877 char name[256];
878 LLVMValueRef offset = soffset;
879
880 if (inst_offset)
881 offset = LLVMBuildAdd(ctx->builder, offset,
882 LLVMConstInt(ctx->i32, inst_offset, 0), "");
883 if (voffset)
884 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
885
886 LLVMValueRef args[] = {
887 ac_to_float(ctx, vdata),
888 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
889 LLVMConstInt(ctx->i32, 0, 0),
890 offset,
891 LLVMConstInt(ctx->i1, glc, 0),
892 LLVMConstInt(ctx->i1, slc, 0),
893 };
894
895 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
896 types[func]);
897
898 ac_build_intrinsic(ctx, name, ctx->voidt,
899 args, ARRAY_SIZE(args),
900 writeonly_memory ?
901 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
902 AC_FUNC_ATTR_WRITEONLY);
903 return;
904 }
905
906 static unsigned dfmt[] = {
907 V_008F0C_BUF_DATA_FORMAT_32,
908 V_008F0C_BUF_DATA_FORMAT_32_32,
909 V_008F0C_BUF_DATA_FORMAT_32_32_32,
910 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
911 };
912 assert(num_channels >= 1 && num_channels <= 4);
913
914 LLVMValueRef args[] = {
915 rsrc,
916 vdata,
917 LLVMConstInt(ctx->i32, num_channels, 0),
918 voffset ? voffset : LLVMGetUndef(ctx->i32),
919 soffset,
920 LLVMConstInt(ctx->i32, inst_offset, 0),
921 LLVMConstInt(ctx->i32, dfmt[num_channels - 1], 0),
922 LLVMConstInt(ctx->i32, V_008F0C_BUF_NUM_FORMAT_UINT, 0),
923 LLVMConstInt(ctx->i32, voffset != NULL, 0),
924 LLVMConstInt(ctx->i32, 0, 0), /* idxen */
925 LLVMConstInt(ctx->i32, glc, 0),
926 LLVMConstInt(ctx->i32, slc, 0),
927 LLVMConstInt(ctx->i32, 0, 0), /* tfe*/
928 };
929
930 /* The instruction offset field has 12 bits */
931 assert(voffset || inst_offset < (1 << 12));
932
933 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
934 unsigned func = CLAMP(num_channels, 1, 3) - 1;
935 const char *types[] = {"i32", "v2i32", "v4i32"};
936 char name[256];
937 snprintf(name, sizeof(name), "llvm.SI.tbuffer.store.%s", types[func]);
938
939 ac_build_intrinsic(ctx, name, ctx->voidt,
940 args, ARRAY_SIZE(args),
941 AC_FUNC_ATTR_LEGACY);
942 }
943
944 LLVMValueRef
ac_build_buffer_load(struct ac_llvm_context * ctx,LLVMValueRef rsrc,int num_channels,LLVMValueRef vindex,LLVMValueRef voffset,LLVMValueRef soffset,unsigned inst_offset,unsigned glc,unsigned slc,bool can_speculate,bool allow_smem)945 ac_build_buffer_load(struct ac_llvm_context *ctx,
946 LLVMValueRef rsrc,
947 int num_channels,
948 LLVMValueRef vindex,
949 LLVMValueRef voffset,
950 LLVMValueRef soffset,
951 unsigned inst_offset,
952 unsigned glc,
953 unsigned slc,
954 bool can_speculate,
955 bool allow_smem)
956 {
957 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
958 if (voffset)
959 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
960 if (soffset)
961 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
962
963 /* TODO: VI and later generations can use SMEM with GLC=1.*/
964 if (allow_smem && !glc && !slc) {
965 assert(vindex == NULL);
966
967 LLVMValueRef result[4];
968
969 for (int i = 0; i < num_channels; i++) {
970 if (i) {
971 offset = LLVMBuildAdd(ctx->builder, offset,
972 LLVMConstInt(ctx->i32, 4, 0), "");
973 }
974 LLVMValueRef args[2] = {rsrc, offset};
975 result[i] = ac_build_intrinsic(ctx, "llvm.SI.load.const.v4i32",
976 ctx->f32, args, 2,
977 AC_FUNC_ATTR_READNONE |
978 AC_FUNC_ATTR_LEGACY);
979 }
980 if (num_channels == 1)
981 return result[0];
982
983 if (num_channels == 3)
984 result[num_channels++] = LLVMGetUndef(ctx->f32);
985 return ac_build_gather_values(ctx, result, num_channels);
986 }
987
988 unsigned func = CLAMP(num_channels, 1, 3) - 1;
989
990 LLVMValueRef args[] = {
991 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
992 vindex ? vindex : LLVMConstInt(ctx->i32, 0, 0),
993 offset,
994 LLVMConstInt(ctx->i1, glc, 0),
995 LLVMConstInt(ctx->i1, slc, 0)
996 };
997
998 LLVMTypeRef types[] = {ctx->f32, LLVMVectorType(ctx->f32, 2),
999 ctx->v4f32};
1000 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1001 char name[256];
1002
1003 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1004 type_names[func]);
1005
1006 return ac_build_intrinsic(ctx, name, types[func], args,
1007 ARRAY_SIZE(args),
1008 ac_get_load_intr_attribs(can_speculate));
1009 }
1010
ac_build_buffer_load_format(struct ac_llvm_context * ctx,LLVMValueRef rsrc,LLVMValueRef vindex,LLVMValueRef voffset,bool can_speculate)1011 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1012 LLVMValueRef rsrc,
1013 LLVMValueRef vindex,
1014 LLVMValueRef voffset,
1015 bool can_speculate)
1016 {
1017 LLVMValueRef args [] = {
1018 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1019 vindex,
1020 voffset,
1021 ctx->i1false, /* glc */
1022 ctx->i1false, /* slc */
1023 };
1024
1025 return ac_build_intrinsic(ctx,
1026 "llvm.amdgcn.buffer.load.format.v4f32",
1027 ctx->v4f32, args, ARRAY_SIZE(args),
1028 ac_get_load_intr_attribs(can_speculate));
1029 }
1030
ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context * ctx,LLVMValueRef rsrc,LLVMValueRef vindex,LLVMValueRef voffset,bool can_speculate)1031 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
1032 LLVMValueRef rsrc,
1033 LLVMValueRef vindex,
1034 LLVMValueRef voffset,
1035 bool can_speculate)
1036 {
1037 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
1038 LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 1, 0), "");
1039 stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
1040
1041 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
1042 LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
1043 elem_count, stride, "");
1044
1045 LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
1046 LLVMConstInt(ctx->i32, 2, 0), "");
1047
1048 return ac_build_buffer_load_format(ctx, new_rsrc, vindex, voffset, can_speculate);
1049 }
1050
1051 /**
1052 * Set range metadata on an instruction. This can only be used on load and
1053 * call instructions. If you know an instruction can only produce the values
1054 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1055 * \p lo is the minimum value inclusive.
1056 * \p hi is the maximum value exclusive.
1057 */
set_range_metadata(struct ac_llvm_context * ctx,LLVMValueRef value,unsigned lo,unsigned hi)1058 static void set_range_metadata(struct ac_llvm_context *ctx,
1059 LLVMValueRef value, unsigned lo, unsigned hi)
1060 {
1061 LLVMValueRef range_md, md_args[2];
1062 LLVMTypeRef type = LLVMTypeOf(value);
1063 LLVMContextRef context = LLVMGetTypeContext(type);
1064
1065 md_args[0] = LLVMConstInt(type, lo, false);
1066 md_args[1] = LLVMConstInt(type, hi, false);
1067 range_md = LLVMMDNodeInContext(context, md_args, 2);
1068 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1069 }
1070
1071 LLVMValueRef
ac_get_thread_id(struct ac_llvm_context * ctx)1072 ac_get_thread_id(struct ac_llvm_context *ctx)
1073 {
1074 LLVMValueRef tid;
1075
1076 LLVMValueRef tid_args[2];
1077 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
1078 tid_args[1] = LLVMConstInt(ctx->i32, 0, false);
1079 tid_args[1] = ac_build_intrinsic(ctx,
1080 "llvm.amdgcn.mbcnt.lo", ctx->i32,
1081 tid_args, 2, AC_FUNC_ATTR_READNONE);
1082
1083 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
1084 ctx->i32, tid_args,
1085 2, AC_FUNC_ATTR_READNONE);
1086 set_range_metadata(ctx, tid, 0, 64);
1087 return tid;
1088 }
1089
1090 /*
1091 * SI implements derivatives using the local data store (LDS)
1092 * All writes to the LDS happen in all executing threads at
1093 * the same time. TID is the Thread ID for the current
1094 * thread and is a value between 0 and 63, representing
1095 * the thread's position in the wavefront.
1096 *
1097 * For the pixel shader threads are grouped into quads of four pixels.
1098 * The TIDs of the pixels of a quad are:
1099 *
1100 * +------+------+
1101 * |4n + 0|4n + 1|
1102 * +------+------+
1103 * |4n + 2|4n + 3|
1104 * +------+------+
1105 *
1106 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1107 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1108 * the current pixel's column, and masking with 0xfffffffe yields the TID
1109 * of the left pixel of the current pixel's row.
1110 *
1111 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1112 * adding 2 yields the TID of the pixel below the top pixel.
1113 */
1114 LLVMValueRef
ac_build_ddxy(struct ac_llvm_context * ctx,uint32_t mask,int idx,LLVMValueRef val)1115 ac_build_ddxy(struct ac_llvm_context *ctx,
1116 uint32_t mask,
1117 int idx,
1118 LLVMValueRef val)
1119 {
1120 LLVMValueRef tl, trbl, args[2];
1121 LLVMValueRef result;
1122
1123 if (ctx->chip_class >= VI) {
1124 LLVMValueRef thread_id, tl_tid, trbl_tid;
1125 thread_id = ac_get_thread_id(ctx);
1126
1127 tl_tid = LLVMBuildAnd(ctx->builder, thread_id,
1128 LLVMConstInt(ctx->i32, mask, false), "");
1129
1130 trbl_tid = LLVMBuildAdd(ctx->builder, tl_tid,
1131 LLVMConstInt(ctx->i32, idx, false), "");
1132
1133 args[0] = LLVMBuildMul(ctx->builder, tl_tid,
1134 LLVMConstInt(ctx->i32, 4, false), "");
1135 args[1] = val;
1136 tl = ac_build_intrinsic(ctx,
1137 "llvm.amdgcn.ds.bpermute", ctx->i32,
1138 args, 2,
1139 AC_FUNC_ATTR_READNONE |
1140 AC_FUNC_ATTR_CONVERGENT);
1141
1142 args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
1143 LLVMConstInt(ctx->i32, 4, false), "");
1144 trbl = ac_build_intrinsic(ctx,
1145 "llvm.amdgcn.ds.bpermute", ctx->i32,
1146 args, 2,
1147 AC_FUNC_ATTR_READNONE |
1148 AC_FUNC_ATTR_CONVERGENT);
1149 } else {
1150 uint32_t masks[2] = {};
1151
1152 switch (mask) {
1153 case AC_TID_MASK_TOP_LEFT:
1154 masks[0] = 0x8000;
1155 if (idx == 1)
1156 masks[1] = 0x8055;
1157 else
1158 masks[1] = 0x80aa;
1159
1160 break;
1161 case AC_TID_MASK_TOP:
1162 masks[0] = 0x8044;
1163 masks[1] = 0x80ee;
1164 break;
1165 case AC_TID_MASK_LEFT:
1166 masks[0] = 0x80a0;
1167 masks[1] = 0x80f5;
1168 break;
1169 default:
1170 assert(0);
1171 }
1172
1173 args[0] = val;
1174 args[1] = LLVMConstInt(ctx->i32, masks[0], false);
1175
1176 tl = ac_build_intrinsic(ctx,
1177 "llvm.amdgcn.ds.swizzle", ctx->i32,
1178 args, 2,
1179 AC_FUNC_ATTR_READNONE |
1180 AC_FUNC_ATTR_CONVERGENT);
1181
1182 args[1] = LLVMConstInt(ctx->i32, masks[1], false);
1183 trbl = ac_build_intrinsic(ctx,
1184 "llvm.amdgcn.ds.swizzle", ctx->i32,
1185 args, 2,
1186 AC_FUNC_ATTR_READNONE |
1187 AC_FUNC_ATTR_CONVERGENT);
1188 }
1189
1190 tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
1191 trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
1192 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
1193 return result;
1194 }
1195
1196 void
ac_build_sendmsg(struct ac_llvm_context * ctx,uint32_t msg,LLVMValueRef wave_id)1197 ac_build_sendmsg(struct ac_llvm_context *ctx,
1198 uint32_t msg,
1199 LLVMValueRef wave_id)
1200 {
1201 LLVMValueRef args[2];
1202 const char *intr_name = (HAVE_LLVM < 0x0400) ? "llvm.SI.sendmsg" : "llvm.amdgcn.s.sendmsg";
1203 args[0] = LLVMConstInt(ctx->i32, msg, false);
1204 args[1] = wave_id;
1205 ac_build_intrinsic(ctx, intr_name, ctx->voidt, args, 2, 0);
1206 }
1207
1208 LLVMValueRef
ac_build_imsb(struct ac_llvm_context * ctx,LLVMValueRef arg,LLVMTypeRef dst_type)1209 ac_build_imsb(struct ac_llvm_context *ctx,
1210 LLVMValueRef arg,
1211 LLVMTypeRef dst_type)
1212 {
1213 const char *intr_name = (HAVE_LLVM < 0x0400) ? "llvm.AMDGPU.flbit.i32" :
1214 "llvm.amdgcn.sffbh.i32";
1215 LLVMValueRef msb = ac_build_intrinsic(ctx, intr_name,
1216 dst_type, &arg, 1,
1217 AC_FUNC_ATTR_READNONE);
1218
1219 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1220 * the index from LSB. Invert it by doing "31 - msb". */
1221 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1222 msb, "");
1223
1224 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
1225 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
1226 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1227 arg, LLVMConstInt(ctx->i32, 0, 0), ""),
1228 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1229 arg, all_ones, ""), "");
1230
1231 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
1232 }
1233
1234 LLVMValueRef
ac_build_umsb(struct ac_llvm_context * ctx,LLVMValueRef arg,LLVMTypeRef dst_type)1235 ac_build_umsb(struct ac_llvm_context *ctx,
1236 LLVMValueRef arg,
1237 LLVMTypeRef dst_type)
1238 {
1239 LLVMValueRef args[2] = {
1240 arg,
1241 ctx->i1true,
1242 };
1243 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.ctlz.i32",
1244 dst_type, args, ARRAY_SIZE(args),
1245 AC_FUNC_ATTR_READNONE);
1246
1247 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1248 * the index from LSB. Invert it by doing "31 - msb". */
1249 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1250 msb, "");
1251
1252 /* check for zero */
1253 return LLVMBuildSelect(ctx->builder,
1254 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg,
1255 LLVMConstInt(ctx->i32, 0, 0), ""),
1256 LLVMConstInt(ctx->i32, -1, true), msb, "");
1257 }
1258
ac_build_fmin(struct ac_llvm_context * ctx,LLVMValueRef a,LLVMValueRef b)1259 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
1260 LLVMValueRef b)
1261 {
1262 LLVMValueRef args[2] = {a, b};
1263 return ac_build_intrinsic(ctx, "llvm.minnum.f32", ctx->f32, args, 2,
1264 AC_FUNC_ATTR_READNONE);
1265 }
1266
ac_build_fmax(struct ac_llvm_context * ctx,LLVMValueRef a,LLVMValueRef b)1267 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
1268 LLVMValueRef b)
1269 {
1270 LLVMValueRef args[2] = {a, b};
1271 return ac_build_intrinsic(ctx, "llvm.maxnum.f32", ctx->f32, args, 2,
1272 AC_FUNC_ATTR_READNONE);
1273 }
1274
ac_build_umin(struct ac_llvm_context * ctx,LLVMValueRef a,LLVMValueRef b)1275 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
1276 LLVMValueRef b)
1277 {
1278 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
1279 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1280 }
1281
ac_build_clamp(struct ac_llvm_context * ctx,LLVMValueRef value)1282 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
1283 {
1284 if (HAVE_LLVM >= 0x0500) {
1285 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, ctx->f32_0),
1286 ctx->f32_1);
1287 }
1288
1289 LLVMValueRef args[3] = {
1290 value,
1291 LLVMConstReal(ctx->f32, 0),
1292 LLVMConstReal(ctx->f32, 1),
1293 };
1294
1295 return ac_build_intrinsic(ctx, "llvm.AMDGPU.clamp.", ctx->f32, args, 3,
1296 AC_FUNC_ATTR_READNONE |
1297 AC_FUNC_ATTR_LEGACY);
1298 }
1299
ac_build_export(struct ac_llvm_context * ctx,struct ac_export_args * a)1300 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
1301 {
1302 LLVMValueRef args[9];
1303
1304 if (HAVE_LLVM >= 0x0500) {
1305 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
1306 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1307
1308 if (a->compr) {
1309 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
1310 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
1311
1312 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
1313 v2i16, "");
1314 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
1315 v2i16, "");
1316 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
1317 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1318
1319 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
1320 ctx->voidt, args, 6, 0);
1321 } else {
1322 args[2] = a->out[0];
1323 args[3] = a->out[1];
1324 args[4] = a->out[2];
1325 args[5] = a->out[3];
1326 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
1327 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1328
1329 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
1330 ctx->voidt, args, 8, 0);
1331 }
1332 return;
1333 }
1334
1335 args[0] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1336 args[1] = LLVMConstInt(ctx->i32, a->valid_mask, 0);
1337 args[2] = LLVMConstInt(ctx->i32, a->done, 0);
1338 args[3] = LLVMConstInt(ctx->i32, a->target, 0);
1339 args[4] = LLVMConstInt(ctx->i32, a->compr, 0);
1340 memcpy(args + 5, a->out, sizeof(a->out[0]) * 4);
1341
1342 ac_build_intrinsic(ctx, "llvm.SI.export", ctx->voidt, args, 9,
1343 AC_FUNC_ATTR_LEGACY);
1344 }
1345
ac_build_image_opcode(struct ac_llvm_context * ctx,struct ac_image_args * a)1346 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
1347 struct ac_image_args *a)
1348 {
1349 LLVMTypeRef dst_type;
1350 LLVMValueRef args[11];
1351 unsigned num_args = 0;
1352 const char *name = NULL;
1353 char intr_name[128], type[64];
1354
1355 if (HAVE_LLVM >= 0x0400) {
1356 bool sample = a->opcode == ac_image_sample ||
1357 a->opcode == ac_image_gather4 ||
1358 a->opcode == ac_image_get_lod;
1359
1360 if (sample)
1361 args[num_args++] = ac_to_float(ctx, a->addr);
1362 else
1363 args[num_args++] = a->addr;
1364
1365 args[num_args++] = a->resource;
1366 if (sample)
1367 args[num_args++] = a->sampler;
1368 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, 0);
1369 if (sample)
1370 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, 0);
1371 args[num_args++] = ctx->i1false; /* glc */
1372 args[num_args++] = ctx->i1false; /* slc */
1373 args[num_args++] = ctx->i1false; /* lwe */
1374 args[num_args++] = LLVMConstInt(ctx->i1, a->da, 0);
1375
1376 switch (a->opcode) {
1377 case ac_image_sample:
1378 name = "llvm.amdgcn.image.sample";
1379 break;
1380 case ac_image_gather4:
1381 name = "llvm.amdgcn.image.gather4";
1382 break;
1383 case ac_image_load:
1384 name = "llvm.amdgcn.image.load";
1385 break;
1386 case ac_image_load_mip:
1387 name = "llvm.amdgcn.image.load.mip";
1388 break;
1389 case ac_image_get_lod:
1390 name = "llvm.amdgcn.image.getlod";
1391 break;
1392 case ac_image_get_resinfo:
1393 name = "llvm.amdgcn.image.getresinfo";
1394 break;
1395 default:
1396 unreachable("invalid image opcode");
1397 }
1398
1399 ac_build_type_name_for_intr(LLVMTypeOf(args[0]), type,
1400 sizeof(type));
1401
1402 snprintf(intr_name, sizeof(intr_name), "%s%s%s%s.v4f32.%s.v8i32",
1403 name,
1404 a->compare ? ".c" : "",
1405 a->bias ? ".b" :
1406 a->lod ? ".l" :
1407 a->deriv ? ".d" :
1408 a->level_zero ? ".lz" : "",
1409 a->offset ? ".o" : "",
1410 type);
1411
1412 LLVMValueRef result =
1413 ac_build_intrinsic(ctx, intr_name,
1414 ctx->v4f32, args, num_args,
1415 AC_FUNC_ATTR_READNONE);
1416 if (!sample) {
1417 result = LLVMBuildBitCast(ctx->builder, result,
1418 ctx->v4i32, "");
1419 }
1420 return result;
1421 }
1422
1423 args[num_args++] = a->addr;
1424 args[num_args++] = a->resource;
1425
1426 if (a->opcode == ac_image_load ||
1427 a->opcode == ac_image_load_mip ||
1428 a->opcode == ac_image_get_resinfo) {
1429 dst_type = ctx->v4i32;
1430 } else {
1431 dst_type = ctx->v4f32;
1432 args[num_args++] = a->sampler;
1433 }
1434
1435 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, 0);
1436 args[num_args++] = LLVMConstInt(ctx->i32, a->unorm, 0);
1437 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* r128 */
1438 args[num_args++] = LLVMConstInt(ctx->i32, a->da, 0);
1439 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* glc */
1440 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* slc */
1441 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* tfe */
1442 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* lwe */
1443
1444 switch (a->opcode) {
1445 case ac_image_sample:
1446 name = "llvm.SI.image.sample";
1447 break;
1448 case ac_image_gather4:
1449 name = "llvm.SI.gather4";
1450 break;
1451 case ac_image_load:
1452 name = "llvm.SI.image.load";
1453 break;
1454 case ac_image_load_mip:
1455 name = "llvm.SI.image.load.mip";
1456 break;
1457 case ac_image_get_lod:
1458 name = "llvm.SI.getlod";
1459 break;
1460 case ac_image_get_resinfo:
1461 name = "llvm.SI.getresinfo";
1462 break;
1463 }
1464
1465 ac_build_type_name_for_intr(LLVMTypeOf(a->addr), type, sizeof(type));
1466 snprintf(intr_name, sizeof(intr_name), "%s%s%s%s.%s",
1467 name,
1468 a->compare ? ".c" : "",
1469 a->bias ? ".b" :
1470 a->lod ? ".l" :
1471 a->deriv ? ".d" :
1472 a->level_zero ? ".lz" : "",
1473 a->offset ? ".o" : "",
1474 type);
1475
1476 return ac_build_intrinsic(ctx, intr_name,
1477 dst_type, args, num_args,
1478 AC_FUNC_ATTR_READNONE |
1479 AC_FUNC_ATTR_LEGACY);
1480 }
1481
ac_build_cvt_pkrtz_f16(struct ac_llvm_context * ctx,LLVMValueRef args[2])1482 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
1483 LLVMValueRef args[2])
1484 {
1485 if (HAVE_LLVM >= 0x0500) {
1486 LLVMTypeRef v2f16 =
1487 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
1488 LLVMValueRef res =
1489 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz",
1490 v2f16, args, 2,
1491 AC_FUNC_ATTR_READNONE);
1492 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1493 }
1494
1495 return ac_build_intrinsic(ctx, "llvm.SI.packf16", ctx->i32, args, 2,
1496 AC_FUNC_ATTR_READNONE |
1497 AC_FUNC_ATTR_LEGACY);
1498 }
1499
ac_build_wqm_vote(struct ac_llvm_context * ctx,LLVMValueRef i1)1500 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
1501 {
1502 assert(HAVE_LLVM >= 0x0600);
1503 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
1504 &i1, 1, AC_FUNC_ATTR_READNONE);
1505 }
1506
ac_build_kill_if_false(struct ac_llvm_context * ctx,LLVMValueRef i1)1507 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
1508 {
1509 if (HAVE_LLVM >= 0x0600) {
1510 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
1511 &i1, 1, 0);
1512 return;
1513 }
1514
1515 LLVMValueRef value = LLVMBuildSelect(ctx->builder, i1,
1516 LLVMConstReal(ctx->f32, 1),
1517 LLVMConstReal(ctx->f32, -1), "");
1518 ac_build_intrinsic(ctx, "llvm.AMDGPU.kill", ctx->voidt,
1519 &value, 1, AC_FUNC_ATTR_LEGACY);
1520 }
1521
ac_build_bfe(struct ac_llvm_context * ctx,LLVMValueRef input,LLVMValueRef offset,LLVMValueRef width,bool is_signed)1522 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
1523 LLVMValueRef offset, LLVMValueRef width,
1524 bool is_signed)
1525 {
1526 LLVMValueRef args[] = {
1527 input,
1528 offset,
1529 width,
1530 };
1531
1532 if (HAVE_LLVM >= 0x0500) {
1533 return ac_build_intrinsic(ctx,
1534 is_signed ? "llvm.amdgcn.sbfe.i32" :
1535 "llvm.amdgcn.ubfe.i32",
1536 ctx->i32, args, 3,
1537 AC_FUNC_ATTR_READNONE);
1538 }
1539
1540 return ac_build_intrinsic(ctx,
1541 is_signed ? "llvm.AMDGPU.bfe.i32" :
1542 "llvm.AMDGPU.bfe.u32",
1543 ctx->i32, args, 3,
1544 AC_FUNC_ATTR_READNONE |
1545 AC_FUNC_ATTR_LEGACY);
1546 }
1547
ac_build_waitcnt(struct ac_llvm_context * ctx,unsigned simm16)1548 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
1549 {
1550 LLVMValueRef args[1] = {
1551 LLVMConstInt(ctx->i32, simm16, false),
1552 };
1553 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
1554 ctx->voidt, args, 1, 0);
1555 }
1556
ac_get_image_intr_name(const char * base_name,LLVMTypeRef data_type,LLVMTypeRef coords_type,LLVMTypeRef rsrc_type,char * out_name,unsigned out_len)1557 void ac_get_image_intr_name(const char *base_name,
1558 LLVMTypeRef data_type,
1559 LLVMTypeRef coords_type,
1560 LLVMTypeRef rsrc_type,
1561 char *out_name, unsigned out_len)
1562 {
1563 char coords_type_name[8];
1564
1565 ac_build_type_name_for_intr(coords_type, coords_type_name,
1566 sizeof(coords_type_name));
1567
1568 if (HAVE_LLVM <= 0x0309) {
1569 snprintf(out_name, out_len, "%s.%s", base_name, coords_type_name);
1570 } else {
1571 char data_type_name[8];
1572 char rsrc_type_name[8];
1573
1574 ac_build_type_name_for_intr(data_type, data_type_name,
1575 sizeof(data_type_name));
1576 ac_build_type_name_for_intr(rsrc_type, rsrc_type_name,
1577 sizeof(rsrc_type_name));
1578 snprintf(out_name, out_len, "%s.%s.%s.%s", base_name,
1579 data_type_name, coords_type_name, rsrc_type_name);
1580 }
1581 }
1582
1583 #define AC_EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
1584 #define AC_EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
1585
1586 enum ac_ir_type {
1587 AC_IR_UNDEF,
1588 AC_IR_CONST,
1589 AC_IR_VALUE,
1590 };
1591
1592 struct ac_vs_exp_chan
1593 {
1594 LLVMValueRef value;
1595 float const_float;
1596 enum ac_ir_type type;
1597 };
1598
1599 struct ac_vs_exp_inst {
1600 unsigned offset;
1601 LLVMValueRef inst;
1602 struct ac_vs_exp_chan chan[4];
1603 };
1604
1605 struct ac_vs_exports {
1606 unsigned num;
1607 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
1608 };
1609
1610 /* Return true if the PARAM export has been eliminated. */
ac_eliminate_const_output(uint8_t * vs_output_param_offset,uint32_t num_outputs,struct ac_vs_exp_inst * exp)1611 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
1612 uint32_t num_outputs,
1613 struct ac_vs_exp_inst *exp)
1614 {
1615 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
1616 bool is_zero[4] = {}, is_one[4] = {};
1617
1618 for (i = 0; i < 4; i++) {
1619 /* It's a constant expression. Undef outputs are eliminated too. */
1620 if (exp->chan[i].type == AC_IR_UNDEF) {
1621 is_zero[i] = true;
1622 is_one[i] = true;
1623 } else if (exp->chan[i].type == AC_IR_CONST) {
1624 if (exp->chan[i].const_float == 0)
1625 is_zero[i] = true;
1626 else if (exp->chan[i].const_float == 1)
1627 is_one[i] = true;
1628 else
1629 return false; /* other constant */
1630 } else
1631 return false;
1632 }
1633
1634 /* Only certain combinations of 0 and 1 can be eliminated. */
1635 if (is_zero[0] && is_zero[1] && is_zero[2])
1636 default_val = is_zero[3] ? 0 : 1;
1637 else if (is_one[0] && is_one[1] && is_one[2])
1638 default_val = is_zero[3] ? 2 : 3;
1639 else
1640 return false;
1641
1642 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
1643 LLVMInstructionEraseFromParent(exp->inst);
1644
1645 /* Change OFFSET to DEFAULT_VAL. */
1646 for (i = 0; i < num_outputs; i++) {
1647 if (vs_output_param_offset[i] == exp->offset) {
1648 vs_output_param_offset[i] =
1649 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
1650 break;
1651 }
1652 }
1653 return true;
1654 }
1655
ac_eliminate_duplicated_output(uint8_t * vs_output_param_offset,uint32_t num_outputs,struct ac_vs_exports * processed,struct ac_vs_exp_inst * exp)1656 static bool ac_eliminate_duplicated_output(uint8_t *vs_output_param_offset,
1657 uint32_t num_outputs,
1658 struct ac_vs_exports *processed,
1659 struct ac_vs_exp_inst *exp)
1660 {
1661 unsigned p, copy_back_channels = 0;
1662
1663 /* See if the output is already in the list of processed outputs.
1664 * The LLVMValueRef comparison relies on SSA.
1665 */
1666 for (p = 0; p < processed->num; p++) {
1667 bool different = false;
1668
1669 for (unsigned j = 0; j < 4; j++) {
1670 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
1671 struct ac_vs_exp_chan *c2 = &exp->chan[j];
1672
1673 /* Treat undef as a match. */
1674 if (c2->type == AC_IR_UNDEF)
1675 continue;
1676
1677 /* If c1 is undef but c2 isn't, we can copy c2 to c1
1678 * and consider the instruction duplicated.
1679 */
1680 if (c1->type == AC_IR_UNDEF) {
1681 copy_back_channels |= 1 << j;
1682 continue;
1683 }
1684
1685 /* Test whether the channels are not equal. */
1686 if (c1->type != c2->type ||
1687 (c1->type == AC_IR_CONST &&
1688 c1->const_float != c2->const_float) ||
1689 (c1->type == AC_IR_VALUE &&
1690 c1->value != c2->value)) {
1691 different = true;
1692 break;
1693 }
1694 }
1695 if (!different)
1696 break;
1697
1698 copy_back_channels = 0;
1699 }
1700 if (p == processed->num)
1701 return false;
1702
1703 /* If a match was found, but the matching export has undef where the new
1704 * one has a normal value, copy the normal value to the undef channel.
1705 */
1706 struct ac_vs_exp_inst *match = &processed->exp[p];
1707
1708 while (copy_back_channels) {
1709 unsigned chan = u_bit_scan(©_back_channels);
1710
1711 assert(match->chan[chan].type == AC_IR_UNDEF);
1712 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
1713 exp->chan[chan].value);
1714 match->chan[chan] = exp->chan[chan];
1715 }
1716
1717 /* The PARAM export is duplicated. Kill it. */
1718 LLVMInstructionEraseFromParent(exp->inst);
1719
1720 /* Change OFFSET to the matching export. */
1721 for (unsigned i = 0; i < num_outputs; i++) {
1722 if (vs_output_param_offset[i] == exp->offset) {
1723 vs_output_param_offset[i] = match->offset;
1724 break;
1725 }
1726 }
1727 return true;
1728 }
1729
ac_optimize_vs_outputs(struct ac_llvm_context * ctx,LLVMValueRef main_fn,uint8_t * vs_output_param_offset,uint32_t num_outputs,uint8_t * num_param_exports)1730 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
1731 LLVMValueRef main_fn,
1732 uint8_t *vs_output_param_offset,
1733 uint32_t num_outputs,
1734 uint8_t *num_param_exports)
1735 {
1736 LLVMBasicBlockRef bb;
1737 bool removed_any = false;
1738 struct ac_vs_exports exports;
1739
1740 exports.num = 0;
1741
1742 /* Process all LLVM instructions. */
1743 bb = LLVMGetFirstBasicBlock(main_fn);
1744 while (bb) {
1745 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
1746
1747 while (inst) {
1748 LLVMValueRef cur = inst;
1749 inst = LLVMGetNextInstruction(inst);
1750 struct ac_vs_exp_inst exp;
1751
1752 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
1753 continue;
1754
1755 LLVMValueRef callee = ac_llvm_get_called_value(cur);
1756
1757 if (!ac_llvm_is_function(callee))
1758 continue;
1759
1760 const char *name = LLVMGetValueName(callee);
1761 unsigned num_args = LLVMCountParams(callee);
1762
1763 /* Check if this is an export instruction. */
1764 if ((num_args != 9 && num_args != 8) ||
1765 (strcmp(name, "llvm.SI.export") &&
1766 strcmp(name, "llvm.amdgcn.exp.f32")))
1767 continue;
1768
1769 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
1770 unsigned target = LLVMConstIntGetZExtValue(arg);
1771
1772 if (target < V_008DFC_SQ_EXP_PARAM)
1773 continue;
1774
1775 target -= V_008DFC_SQ_EXP_PARAM;
1776
1777 /* Parse the instruction. */
1778 memset(&exp, 0, sizeof(exp));
1779 exp.offset = target;
1780 exp.inst = cur;
1781
1782 for (unsigned i = 0; i < 4; i++) {
1783 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
1784
1785 exp.chan[i].value = v;
1786
1787 if (LLVMIsUndef(v)) {
1788 exp.chan[i].type = AC_IR_UNDEF;
1789 } else if (LLVMIsAConstantFP(v)) {
1790 LLVMBool loses_info;
1791 exp.chan[i].type = AC_IR_CONST;
1792 exp.chan[i].const_float =
1793 LLVMConstRealGetDouble(v, &loses_info);
1794 } else {
1795 exp.chan[i].type = AC_IR_VALUE;
1796 }
1797 }
1798
1799 /* Eliminate constant and duplicated PARAM exports. */
1800 if (ac_eliminate_const_output(vs_output_param_offset,
1801 num_outputs, &exp) ||
1802 ac_eliminate_duplicated_output(vs_output_param_offset,
1803 num_outputs, &exports,
1804 &exp)) {
1805 removed_any = true;
1806 } else {
1807 exports.exp[exports.num++] = exp;
1808 }
1809 }
1810 bb = LLVMGetNextBasicBlock(bb);
1811 }
1812
1813 /* Remove holes in export memory due to removed PARAM exports.
1814 * This is done by renumbering all PARAM exports.
1815 */
1816 if (removed_any) {
1817 uint8_t old_offset[VARYING_SLOT_MAX];
1818 unsigned out, i;
1819
1820 /* Make a copy of the offsets. We need the old version while
1821 * we are modifying some of them. */
1822 memcpy(old_offset, vs_output_param_offset,
1823 sizeof(old_offset));
1824
1825 for (i = 0; i < exports.num; i++) {
1826 unsigned offset = exports.exp[i].offset;
1827
1828 /* Update vs_output_param_offset. Multiple outputs can
1829 * have the same offset.
1830 */
1831 for (out = 0; out < num_outputs; out++) {
1832 if (old_offset[out] == offset)
1833 vs_output_param_offset[out] = i;
1834 }
1835
1836 /* Change the PARAM offset in the instruction. */
1837 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
1838 LLVMConstInt(ctx->i32,
1839 V_008DFC_SQ_EXP_PARAM + i, 0));
1840 }
1841 *num_param_exports = exports.num;
1842 }
1843 }
1844
ac_init_exec_full_mask(struct ac_llvm_context * ctx)1845 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
1846 {
1847 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
1848 ac_build_intrinsic(ctx,
1849 "llvm.amdgcn.init.exec", ctx->voidt,
1850 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
1851 }
1852
ac_declare_lds_as_pointer(struct ac_llvm_context * ctx)1853 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
1854 {
1855 unsigned lds_size = ctx->chip_class >= CIK ? 65536 : 32768;
1856 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
1857 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_LOCAL_ADDR_SPACE),
1858 "lds");
1859 }
1860
ac_lds_load(struct ac_llvm_context * ctx,LLVMValueRef dw_addr)1861 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
1862 LLVMValueRef dw_addr)
1863 {
1864 return ac_build_load(ctx, ctx->lds, dw_addr);
1865 }
1866
ac_lds_store(struct ac_llvm_context * ctx,LLVMValueRef dw_addr,LLVMValueRef value)1867 void ac_lds_store(struct ac_llvm_context *ctx,
1868 LLVMValueRef dw_addr,
1869 LLVMValueRef value)
1870 {
1871 value = ac_to_integer(ctx, value);
1872 ac_build_indexed_store(ctx, ctx->lds,
1873 dw_addr, value);
1874 }
1875
ac_find_lsb(struct ac_llvm_context * ctx,LLVMTypeRef dst_type,LLVMValueRef src0)1876 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
1877 LLVMTypeRef dst_type,
1878 LLVMValueRef src0)
1879 {
1880 LLVMValueRef params[2] = {
1881 src0,
1882
1883 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
1884 * add special code to check for x=0. The reason is that
1885 * the LLVM behavior for x=0 is different from what we
1886 * need here. However, LLVM also assumes that ffs(x) is
1887 * in [0, 31], but GLSL expects that ffs(0) = -1, so
1888 * a conditional assignment to handle 0 is still required.
1889 *
1890 * The hardware already implements the correct behavior.
1891 */
1892 LLVMConstInt(ctx->i1, 1, false),
1893 };
1894
1895 LLVMValueRef lsb = ac_build_intrinsic(ctx, "llvm.cttz.i32", ctx->i32,
1896 params, 2,
1897 AC_FUNC_ATTR_READNONE);
1898
1899 /* TODO: We need an intrinsic to skip this conditional. */
1900 /* Check for zero: */
1901 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
1902 LLVMIntEQ, src0,
1903 ctx->i32_0, ""),
1904 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
1905 }
1906
1907 static struct ac_llvm_flow *
get_current_flow(struct ac_llvm_context * ctx)1908 get_current_flow(struct ac_llvm_context *ctx)
1909 {
1910 if (ctx->flow_depth > 0)
1911 return &ctx->flow[ctx->flow_depth - 1];
1912 return NULL;
1913 }
1914
1915 static struct ac_llvm_flow *
get_innermost_loop(struct ac_llvm_context * ctx)1916 get_innermost_loop(struct ac_llvm_context *ctx)
1917 {
1918 for (unsigned i = ctx->flow_depth; i > 0; --i) {
1919 if (ctx->flow[i - 1].loop_entry_block)
1920 return &ctx->flow[i - 1];
1921 }
1922 return NULL;
1923 }
1924
1925 static struct ac_llvm_flow *
push_flow(struct ac_llvm_context * ctx)1926 push_flow(struct ac_llvm_context *ctx)
1927 {
1928 struct ac_llvm_flow *flow;
1929
1930 if (ctx->flow_depth >= ctx->flow_depth_max) {
1931 unsigned new_max = MAX2(ctx->flow_depth << 1,
1932 AC_LLVM_INITIAL_CF_DEPTH);
1933
1934 ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
1935 ctx->flow_depth_max = new_max;
1936 }
1937
1938 flow = &ctx->flow[ctx->flow_depth];
1939 ctx->flow_depth++;
1940
1941 flow->next_block = NULL;
1942 flow->loop_entry_block = NULL;
1943 return flow;
1944 }
1945
set_basicblock_name(LLVMBasicBlockRef bb,const char * base,int label_id)1946 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
1947 int label_id)
1948 {
1949 char buf[32];
1950 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
1951 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
1952 }
1953
1954 /* Append a basic block at the level of the parent flow.
1955 */
append_basic_block(struct ac_llvm_context * ctx,const char * name)1956 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
1957 const char *name)
1958 {
1959 assert(ctx->flow_depth >= 1);
1960
1961 if (ctx->flow_depth >= 2) {
1962 struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
1963
1964 return LLVMInsertBasicBlockInContext(ctx->context,
1965 flow->next_block, name);
1966 }
1967
1968 LLVMValueRef main_fn =
1969 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
1970 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
1971 }
1972
1973 /* Emit a branch to the given default target for the current block if
1974 * applicable -- that is, if the current block does not already contain a
1975 * branch from a break or continue.
1976 */
emit_default_branch(LLVMBuilderRef builder,LLVMBasicBlockRef target)1977 static void emit_default_branch(LLVMBuilderRef builder,
1978 LLVMBasicBlockRef target)
1979 {
1980 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
1981 LLVMBuildBr(builder, target);
1982 }
1983
ac_build_bgnloop(struct ac_llvm_context * ctx,int label_id)1984 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
1985 {
1986 struct ac_llvm_flow *flow = push_flow(ctx);
1987 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
1988 flow->next_block = append_basic_block(ctx, "ENDLOOP");
1989 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
1990 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
1991 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
1992 }
1993
ac_build_break(struct ac_llvm_context * ctx)1994 void ac_build_break(struct ac_llvm_context *ctx)
1995 {
1996 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
1997 LLVMBuildBr(ctx->builder, flow->next_block);
1998 }
1999
ac_build_continue(struct ac_llvm_context * ctx)2000 void ac_build_continue(struct ac_llvm_context *ctx)
2001 {
2002 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2003 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2004 }
2005
ac_build_else(struct ac_llvm_context * ctx,int label_id)2006 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
2007 {
2008 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2009 LLVMBasicBlockRef endif_block;
2010
2011 assert(!current_branch->loop_entry_block);
2012
2013 endif_block = append_basic_block(ctx, "ENDIF");
2014 emit_default_branch(ctx->builder, endif_block);
2015
2016 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2017 set_basicblock_name(current_branch->next_block, "else", label_id);
2018
2019 current_branch->next_block = endif_block;
2020 }
2021
ac_build_endif(struct ac_llvm_context * ctx,int label_id)2022 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
2023 {
2024 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2025
2026 assert(!current_branch->loop_entry_block);
2027
2028 emit_default_branch(ctx->builder, current_branch->next_block);
2029 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2030 set_basicblock_name(current_branch->next_block, "endif", label_id);
2031
2032 ctx->flow_depth--;
2033 }
2034
ac_build_endloop(struct ac_llvm_context * ctx,int label_id)2035 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
2036 {
2037 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
2038
2039 assert(current_loop->loop_entry_block);
2040
2041 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
2042
2043 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
2044 set_basicblock_name(current_loop->next_block, "endloop", label_id);
2045 ctx->flow_depth--;
2046 }
2047
if_cond_emit(struct ac_llvm_context * ctx,LLVMValueRef cond,int label_id)2048 static void if_cond_emit(struct ac_llvm_context *ctx, LLVMValueRef cond,
2049 int label_id)
2050 {
2051 struct ac_llvm_flow *flow = push_flow(ctx);
2052 LLVMBasicBlockRef if_block;
2053
2054 if_block = append_basic_block(ctx, "IF");
2055 flow->next_block = append_basic_block(ctx, "ELSE");
2056 set_basicblock_name(if_block, "if", label_id);
2057 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
2058 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
2059 }
2060
ac_build_if(struct ac_llvm_context * ctx,LLVMValueRef value,int label_id)2061 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
2062 int label_id)
2063 {
2064 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
2065 value, ctx->f32_0, "");
2066 if_cond_emit(ctx, cond, label_id);
2067 }
2068
ac_build_uif(struct ac_llvm_context * ctx,LLVMValueRef value,int label_id)2069 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
2070 int label_id)
2071 {
2072 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
2073 ac_to_integer(ctx, value),
2074 ctx->i32_0, "");
2075 if_cond_emit(ctx, cond, label_id);
2076 }
2077