• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2016 Red Hat.
3  * Copyright © 2016 Bas Nieuwenhuizen
4  *
5  * based in part on anv driver which is:
6  * Copyright © 2015 Intel Corporation
7  *
8  * Permission is hereby granted, free of charge, to any person obtaining a
9  * copy of this software and associated documentation files (the "Software"),
10  * to deal in the Software without restriction, including without limitation
11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12  * and/or sell copies of the Software, and to permit persons to whom the
13  * Software is furnished to do so, subject to the following conditions:
14  *
15  * The above copyright notice and this permission notice (including the next
16  * paragraph) shall be included in all copies or substantial portions of the
17  * Software.
18  *
19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
22  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25  * IN THE SOFTWARE.
26  */
27 
28 #include "nir/nir.h"
29 #include "radv_debug.h"
30 #include "radv_llvm_helper.h"
31 #include "radv_private.h"
32 #include "radv_shader.h"
33 #include "radv_shader_args.h"
34 
35 #include "ac_binary.h"
36 #include "ac_llvm_build.h"
37 #include "ac_nir.h"
38 #include "ac_nir_to_llvm.h"
39 #include "ac_shader_abi.h"
40 #include "ac_shader_util.h"
41 #include "sid.h"
42 
43 struct radv_shader_context {
44    struct ac_llvm_context ac;
45    const struct nir_shader *shader;
46    struct ac_shader_abi abi;
47    const struct radv_nir_compiler_options *options;
48    const struct radv_shader_info *shader_info;
49    const struct radv_shader_args *args;
50 
51    gl_shader_stage stage;
52 
53    unsigned max_workgroup_size;
54    LLVMContextRef context;
55    struct ac_llvm_pointer main_function;
56 
57    LLVMValueRef descriptor_sets[MAX_SETS];
58 
59    LLVMValueRef gs_wave_id;
60 
61    uint64_t output_mask;
62 };
63 
64 static inline struct radv_shader_context *
radv_shader_context_from_abi(struct ac_shader_abi * abi)65 radv_shader_context_from_abi(struct ac_shader_abi *abi)
66 {
67    return container_of(abi, struct radv_shader_context, abi);
68 }
69 
70 static struct ac_llvm_pointer
create_llvm_function(struct ac_llvm_context * ctx,LLVMModuleRef module,LLVMBuilderRef builder,const struct ac_shader_args * args,enum ac_llvm_calling_convention convention,unsigned max_workgroup_size,const struct radv_nir_compiler_options * options)71 create_llvm_function(struct ac_llvm_context *ctx, LLVMModuleRef module, LLVMBuilderRef builder,
72                      const struct ac_shader_args *args, enum ac_llvm_calling_convention convention,
73                      unsigned max_workgroup_size, const struct radv_nir_compiler_options *options)
74 {
75    struct ac_llvm_pointer main_function = ac_build_main(args, ctx, convention, "main", ctx->voidt, module);
76 
77    if (options->info->address32_hi) {
78       ac_llvm_add_target_dep_function_attr(main_function.value, "amdgpu-32bit-address-high-bits",
79                                            options->info->address32_hi);
80    }
81 
82    ac_llvm_set_workgroup_size(main_function.value, max_workgroup_size);
83    ac_llvm_set_target_features(main_function.value, ctx, true);
84 
85    return main_function;
86 }
87 
88 static void
load_descriptor_sets(struct radv_shader_context * ctx)89 load_descriptor_sets(struct radv_shader_context *ctx)
90 {
91    const struct radv_userdata_locations *user_sgprs_locs = &ctx->shader_info->user_sgprs_locs;
92    uint32_t mask = ctx->shader_info->desc_set_used_mask;
93 
94    if (user_sgprs_locs->shader_data[AC_UD_INDIRECT_DESCRIPTOR_SETS].sgpr_idx != -1) {
95       struct ac_llvm_pointer desc_sets = ac_get_ptr_arg(&ctx->ac, &ctx->args->ac, ctx->args->descriptor_sets[0]);
96       while (mask) {
97          int i = u_bit_scan(&mask);
98 
99          ctx->descriptor_sets[i] = ac_build_load_to_sgpr(&ctx->ac, desc_sets, LLVMConstInt(ctx->ac.i32, i, false));
100          LLVMSetAlignment(ctx->descriptor_sets[i], 4);
101       }
102    } else {
103       while (mask) {
104          int i = u_bit_scan(&mask);
105 
106          ctx->descriptor_sets[i] = ac_get_arg(&ctx->ac, ctx->args->descriptor_sets[i]);
107       }
108    }
109 }
110 
111 static enum ac_llvm_calling_convention
get_llvm_calling_convention(LLVMValueRef func,gl_shader_stage stage)112 get_llvm_calling_convention(LLVMValueRef func, gl_shader_stage stage)
113 {
114    switch (stage) {
115    case MESA_SHADER_VERTEX:
116    case MESA_SHADER_TESS_EVAL:
117       return AC_LLVM_AMDGPU_VS;
118       break;
119    case MESA_SHADER_GEOMETRY:
120       return AC_LLVM_AMDGPU_GS;
121       break;
122    case MESA_SHADER_TESS_CTRL:
123       return AC_LLVM_AMDGPU_HS;
124       break;
125    case MESA_SHADER_FRAGMENT:
126       return AC_LLVM_AMDGPU_PS;
127       break;
128    case MESA_SHADER_COMPUTE:
129       return AC_LLVM_AMDGPU_CS;
130       break;
131    default:
132       unreachable("Unhandle shader type");
133    }
134 }
135 
136 /* Returns whether the stage is a stage that can be directly before the GS */
137 static bool
is_pre_gs_stage(gl_shader_stage stage)138 is_pre_gs_stage(gl_shader_stage stage)
139 {
140    return stage == MESA_SHADER_VERTEX || stage == MESA_SHADER_TESS_EVAL;
141 }
142 
143 static void
create_function(struct radv_shader_context * ctx,gl_shader_stage stage,bool has_previous_stage)144 create_function(struct radv_shader_context *ctx, gl_shader_stage stage, bool has_previous_stage)
145 {
146    if (ctx->ac.gfx_level >= GFX10) {
147       if (is_pre_gs_stage(stage) && ctx->shader_info->is_ngg) {
148          /* On GFX10+, VS and TES are merged into GS for NGG. */
149          stage = MESA_SHADER_GEOMETRY;
150          has_previous_stage = true;
151       }
152    }
153 
154    ctx->main_function = create_llvm_function(&ctx->ac, ctx->ac.module, ctx->ac.builder, &ctx->args->ac,
155                                              get_llvm_calling_convention(ctx->main_function.value, stage),
156                                              ctx->max_workgroup_size, ctx->options);
157 
158    load_descriptor_sets(ctx);
159 
160    if (stage == MESA_SHADER_TESS_CTRL || (stage == MESA_SHADER_VERTEX && ctx->shader_info->vs.as_ls) ||
161        ctx->shader_info->is_ngg ||
162        /* GFX9 has the ESGS ring buffer in LDS. */
163        (stage == MESA_SHADER_GEOMETRY && has_previous_stage)) {
164       ac_declare_lds_as_pointer(&ctx->ac);
165    }
166 }
167 
168 static LLVMValueRef
radv_load_base_vertex(struct ac_shader_abi * abi,bool non_indexed_is_zero)169 radv_load_base_vertex(struct ac_shader_abi *abi, bool non_indexed_is_zero)
170 {
171    struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
172    return ac_get_arg(&ctx->ac, ctx->args->ac.base_vertex);
173 }
174 
175 static LLVMValueRef
radv_load_rsrc(struct radv_shader_context * ctx,LLVMValueRef ptr,LLVMTypeRef type)176 radv_load_rsrc(struct radv_shader_context *ctx, LLVMValueRef ptr, LLVMTypeRef type)
177 {
178    if (ptr && LLVMTypeOf(ptr) == ctx->ac.i32) {
179       LLVMValueRef result;
180 
181       LLVMTypeRef ptr_type = LLVMPointerType(type, AC_ADDR_SPACE_CONST_32BIT);
182       ptr = LLVMBuildIntToPtr(ctx->ac.builder, ptr, ptr_type, "");
183       LLVMSetMetadata(ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
184 
185       result = LLVMBuildLoad2(ctx->ac.builder, type, ptr, "");
186       LLVMSetMetadata(result, ctx->ac.invariant_load_md_kind, ctx->ac.empty_md);
187 
188       return result;
189    }
190 
191    return ptr;
192 }
193 
194 static LLVMValueRef
radv_load_ubo(struct ac_shader_abi * abi,LLVMValueRef buffer_ptr)195 radv_load_ubo(struct ac_shader_abi *abi, LLVMValueRef buffer_ptr)
196 {
197    struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
198    return radv_load_rsrc(ctx, buffer_ptr, ctx->ac.v4i32);
199 }
200 
201 static LLVMValueRef
radv_load_ssbo(struct ac_shader_abi * abi,LLVMValueRef buffer_ptr,bool write,bool non_uniform)202 radv_load_ssbo(struct ac_shader_abi *abi, LLVMValueRef buffer_ptr, bool write, bool non_uniform)
203 {
204    struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
205    return radv_load_rsrc(ctx, buffer_ptr, ctx->ac.v4i32);
206 }
207 
208 static LLVMValueRef
radv_get_sampler_desc(struct ac_shader_abi * abi,LLVMValueRef index,enum ac_descriptor_type desc_type)209 radv_get_sampler_desc(struct ac_shader_abi *abi, LLVMValueRef index, enum ac_descriptor_type desc_type)
210 {
211    struct radv_shader_context *ctx = radv_shader_context_from_abi(abi);
212 
213    /* 3 plane formats always have same size and format for plane 1 & 2, so
214     * use the tail from plane 1 so that we can store only the first 16 bytes
215     * of the last plane. */
216    if (desc_type == AC_DESC_PLANE_2 && index && LLVMTypeOf(index) == ctx->ac.i32) {
217       LLVMValueRef plane1_addr = LLVMBuildSub(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i32, 32, false), "");
218       LLVMValueRef descriptor1 = radv_load_rsrc(ctx, plane1_addr, ctx->ac.v8i32);
219       LLVMValueRef descriptor2 = radv_load_rsrc(ctx, index, ctx->ac.v4i32);
220 
221       LLVMValueRef components[8];
222       for (unsigned i = 0; i < 4; ++i)
223          components[i] = ac_llvm_extract_elem(&ctx->ac, descriptor2, i);
224 
225       for (unsigned i = 4; i < 8; ++i)
226          components[i] = ac_llvm_extract_elem(&ctx->ac, descriptor1, i);
227       return ac_build_gather_values(&ctx->ac, components, 8);
228    }
229 
230    bool v4 = desc_type == AC_DESC_BUFFER || desc_type == AC_DESC_SAMPLER;
231    return radv_load_rsrc(ctx, index, v4 ? ctx->ac.v4i32 : ctx->ac.v8i32);
232 }
233 
234 static void
scan_shader_output_decl(struct radv_shader_context * ctx,struct nir_variable * variable,struct nir_shader * shader,gl_shader_stage stage)235 scan_shader_output_decl(struct radv_shader_context *ctx, struct nir_variable *variable, struct nir_shader *shader,
236                         gl_shader_stage stage)
237 {
238    int idx = variable->data.driver_location;
239    unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
240    uint64_t mask_attribs;
241 
242    if (variable->data.compact) {
243       unsigned component_count = variable->data.location_frac + glsl_get_length(variable->type);
244       attrib_count = (component_count + 3) / 4;
245    }
246 
247    mask_attribs = ((1ull << attrib_count) - 1) << idx;
248 
249    ctx->output_mask |= mask_attribs;
250 }
251 
252 static LLVMValueRef
radv_load_output(struct radv_shader_context * ctx,unsigned index,unsigned chan)253 radv_load_output(struct radv_shader_context *ctx, unsigned index, unsigned chan)
254 {
255    int idx = ac_llvm_reg_index_soa(index, chan);
256    LLVMValueRef output = ctx->abi.outputs[idx];
257    LLVMTypeRef type = ctx->abi.is_16bit[idx] ? ctx->ac.f16 : ctx->ac.f32;
258    return LLVMBuildLoad2(ctx->ac.builder, type, output, "");
259 }
260 
261 static void
ac_llvm_finalize_module(struct radv_shader_context * ctx,LLVMPassManagerRef passmgr)262 ac_llvm_finalize_module(struct radv_shader_context *ctx, LLVMPassManagerRef passmgr)
263 {
264    LLVMRunPassManager(passmgr, ctx->ac.module);
265    LLVMDisposeBuilder(ctx->ac.builder);
266 
267    ac_llvm_context_dispose(&ctx->ac);
268 }
269 
270 static void
prepare_gs_input_vgprs(struct radv_shader_context * ctx,bool merged)271 prepare_gs_input_vgprs(struct radv_shader_context *ctx, bool merged)
272 {
273    if (merged) {
274       ctx->gs_wave_id = ac_unpack_param(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args->ac.merged_wave_info), 16, 8);
275    } else {
276       ctx->gs_wave_id = ac_get_arg(&ctx->ac, ctx->args->ac.gs_wave_id);
277    }
278 }
279 
280 /* Ensure that the esgs ring is declared.
281  *
282  * We declare it with 64KB alignment as a hint that the
283  * pointer value will always be 0.
284  */
285 static void
declare_esgs_ring(struct radv_shader_context * ctx)286 declare_esgs_ring(struct radv_shader_context *ctx)
287 {
288    assert(!LLVMGetNamedGlobal(ctx->ac.module, "esgs_ring"));
289 
290    LLVMValueRef esgs_ring =
291       LLVMAddGlobalInAddressSpace(ctx->ac.module, LLVMArrayType(ctx->ac.i32, 0), "esgs_ring", AC_ADDR_SPACE_LDS);
292    LLVMSetLinkage(esgs_ring, LLVMExternalLinkage);
293    LLVMSetAlignment(esgs_ring, 64 * 1024);
294 }
295 
296 static LLVMValueRef
radv_intrinsic_load(struct ac_shader_abi * abi,nir_intrinsic_instr * intrin)297 radv_intrinsic_load(struct ac_shader_abi *abi, nir_intrinsic_instr *intrin)
298 {
299    switch (intrin->intrinsic) {
300    case nir_intrinsic_load_base_vertex:
301    case nir_intrinsic_load_first_vertex:
302       return radv_load_base_vertex(abi, intrin->intrinsic == nir_intrinsic_load_base_vertex);
303    default:
304       return NULL;
305    }
306 }
307 
308 static LLVMModuleRef
ac_translate_nir_to_llvm(struct ac_llvm_compiler * ac_llvm,const struct radv_nir_compiler_options * options,const struct radv_shader_info * info,struct nir_shader * const * shaders,int shader_count,const struct radv_shader_args * args)309 ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm, const struct radv_nir_compiler_options *options,
310                          const struct radv_shader_info *info, struct nir_shader *const *shaders, int shader_count,
311                          const struct radv_shader_args *args)
312 {
313    struct radv_shader_context ctx = {0};
314    ctx.args = args;
315    ctx.options = options;
316    ctx.shader_info = info;
317 
318    enum ac_float_mode float_mode = AC_FLOAT_MODE_DEFAULT;
319 
320    if (shaders[0]->info.float_controls_execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32) {
321       float_mode = AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO;
322    }
323 
324    bool exports_mrtz = false;
325    bool exports_color_null = false;
326    if (shaders[0]->info.stage == MESA_SHADER_FRAGMENT) {
327       exports_mrtz = info->ps.writes_z || info->ps.writes_stencil || info->ps.writes_sample_mask;
328       exports_color_null = !exports_mrtz || (shaders[0]->info.outputs_written & (0xffu << FRAG_RESULT_DATA0));
329    }
330 
331    ac_llvm_context_init(&ctx.ac, ac_llvm, options->info, float_mode, info->wave_size, info->ballot_bit_size,
332                         exports_color_null, exports_mrtz);
333 
334    uint32_t length = 1;
335    for (uint32_t i = 0; i < shader_count; i++)
336       if (shaders[i]->info.name)
337          length += strlen(shaders[i]->info.name) + 1;
338 
339    char *name = malloc(length);
340    if (name) {
341       uint32_t offset = 0;
342       for (uint32_t i = 0; i < shader_count; i++) {
343          if (!shaders[i]->info.name)
344             continue;
345 
346          strcpy(name + offset, shaders[i]->info.name);
347          offset += strlen(shaders[i]->info.name);
348          if (i != shader_count - 1)
349             name[offset++] = ',';
350       }
351 
352       LLVMSetSourceFileName(ctx.ac.module, name, offset);
353    }
354 
355    ctx.context = ctx.ac.context;
356 
357    ctx.max_workgroup_size = info->workgroup_size;
358 
359    create_function(&ctx, shaders[shader_count - 1]->info.stage, shader_count >= 2);
360 
361    ctx.abi.intrinsic_load = radv_intrinsic_load;
362    ctx.abi.load_ubo = radv_load_ubo;
363    ctx.abi.load_ssbo = radv_load_ssbo;
364    ctx.abi.load_sampler_desc = radv_get_sampler_desc;
365    ctx.abi.clamp_shadow_reference = false;
366    ctx.abi.robust_buffer_access = options->robust_buffer_access_llvm;
367    ctx.abi.load_grid_size_from_user_sgpr = args->load_grid_size_from_user_sgpr;
368 
369    bool is_ngg = is_pre_gs_stage(shaders[0]->info.stage) && info->is_ngg;
370    if (shader_count >= 2 || is_ngg)
371       ac_init_exec_full_mask(&ctx.ac);
372 
373    if (args->ac.vertex_id.used)
374       ctx.abi.vertex_id = ac_get_arg(&ctx.ac, args->ac.vertex_id);
375    if (args->ac.vs_rel_patch_id.used)
376       ctx.abi.vs_rel_patch_id = ac_get_arg(&ctx.ac, args->ac.vs_rel_patch_id);
377    if (args->ac.instance_id.used)
378       ctx.abi.instance_id = ac_get_arg(&ctx.ac, args->ac.instance_id);
379 
380    if (options->info->has_ls_vgpr_init_bug && shaders[shader_count - 1]->info.stage == MESA_SHADER_TESS_CTRL)
381       ac_fixup_ls_hs_input_vgprs(&ctx.ac, &ctx.abi, &args->ac);
382 
383    if (is_ngg) {
384       if (!info->is_ngg_passthrough)
385          declare_esgs_ring(&ctx);
386 
387       if (ctx.stage == MESA_SHADER_GEOMETRY) {
388          /* Scratch space used by NGG GS for repacking vertices at the end. */
389          LLVMTypeRef ai32 = LLVMArrayType(ctx.ac.i32, 8);
390          LLVMValueRef gs_ngg_scratch =
391             LLVMAddGlobalInAddressSpace(ctx.ac.module, ai32, "ngg_scratch", AC_ADDR_SPACE_LDS);
392          LLVMSetInitializer(gs_ngg_scratch, LLVMGetUndef(ai32));
393          LLVMSetLinkage(gs_ngg_scratch, LLVMExternalLinkage);
394          LLVMSetAlignment(gs_ngg_scratch, 4);
395 
396          /* Vertex emit space used by NGG GS for storing all vertex attributes. */
397          LLVMValueRef gs_ngg_emit =
398             LLVMAddGlobalInAddressSpace(ctx.ac.module, LLVMArrayType(ctx.ac.i32, 0), "ngg_emit", AC_ADDR_SPACE_LDS);
399          LLVMSetInitializer(gs_ngg_emit, LLVMGetUndef(ai32));
400          LLVMSetLinkage(gs_ngg_emit, LLVMExternalLinkage);
401          LLVMSetAlignment(gs_ngg_emit, 4);
402       }
403 
404       /* GFX10 hang workaround - there needs to be an s_barrier before gs_alloc_req always */
405       if (ctx.ac.gfx_level == GFX10 && shader_count == 1)
406          ac_build_s_barrier(&ctx.ac, shaders[0]->info.stage);
407    }
408 
409    for (int shader_idx = 0; shader_idx < shader_count; ++shader_idx) {
410       ctx.stage = shaders[shader_idx]->info.stage;
411       ctx.shader = shaders[shader_idx];
412       ctx.output_mask = 0;
413 
414       if (shader_idx && !(shaders[shader_idx]->info.stage == MESA_SHADER_GEOMETRY && info->is_ngg)) {
415          /* Execute a barrier before the second shader in
416           * a merged shader.
417           *
418           * Execute the barrier inside the conditional block,
419           * so that empty waves can jump directly to s_endpgm,
420           * which will also signal the barrier.
421           *
422           * This is possible in gfx9, because an empty wave
423           * for the second shader does not participate in
424           * the epilogue. With NGG, empty waves may still
425           * be required to export data (e.g. GS output vertices),
426           * so we cannot let them exit early.
427           *
428           * If the shader is TCS and the TCS epilog is present
429           * and contains a barrier, it will wait there and then
430           * reach s_endpgm.
431           */
432          ac_build_waitcnt(&ctx.ac, AC_WAIT_LGKM);
433          ac_build_s_barrier(&ctx.ac, shaders[shader_idx]->info.stage);
434       }
435 
436       nir_foreach_shader_out_variable (variable, shaders[shader_idx])
437          scan_shader_output_decl(&ctx, variable, shaders[shader_idx], shaders[shader_idx]->info.stage);
438 
439       bool check_merged_wave_info = shader_count >= 2 && !(is_ngg && shader_idx == 1);
440       LLVMBasicBlockRef merge_block = NULL;
441 
442       if (check_merged_wave_info) {
443          LLVMValueRef fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
444          LLVMBasicBlockRef then_block = LLVMAppendBasicBlockInContext(ctx.ac.context, fn, "");
445          merge_block = LLVMAppendBasicBlockInContext(ctx.ac.context, fn, "");
446 
447          LLVMValueRef count =
448             ac_unpack_param(&ctx.ac, ac_get_arg(&ctx.ac, args->ac.merged_wave_info), 8 * shader_idx, 8);
449          LLVMValueRef thread_id = ac_get_thread_id(&ctx.ac);
450          LLVMValueRef cond = LLVMBuildICmp(ctx.ac.builder, LLVMIntULT, thread_id, count, "");
451          LLVMBuildCondBr(ctx.ac.builder, cond, then_block, merge_block);
452 
453          LLVMPositionBuilderAtEnd(ctx.ac.builder, then_block);
454       }
455 
456       if (shaders[shader_idx]->info.stage == MESA_SHADER_GEOMETRY && !info->is_ngg)
457          prepare_gs_input_vgprs(&ctx, shader_count >= 2);
458 
459       if (!ac_nir_translate(&ctx.ac, &ctx.abi, &args->ac, shaders[shader_idx])) {
460          abort();
461       }
462 
463       if (check_merged_wave_info) {
464          LLVMBuildBr(ctx.ac.builder, merge_block);
465          LLVMPositionBuilderAtEnd(ctx.ac.builder, merge_block);
466       }
467    }
468 
469    LLVMBuildRetVoid(ctx.ac.builder);
470 
471    if (options->dump_preoptir) {
472       fprintf(stderr, "%s LLVM IR:\n\n", radv_get_shader_name(info, shaders[shader_count - 1]->info.stage));
473       ac_dump_module(ctx.ac.module);
474       fprintf(stderr, "\n");
475    }
476 
477    ac_llvm_finalize_module(&ctx, ac_llvm->passmgr);
478 
479    free(name);
480 
481    return ctx.ac.module;
482 }
483 
484 static void
ac_diagnostic_handler(LLVMDiagnosticInfoRef di,void * context)485 ac_diagnostic_handler(LLVMDiagnosticInfoRef di, void *context)
486 {
487    unsigned *retval = (unsigned *)context;
488    LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
489    char *description = LLVMGetDiagInfoDescription(di);
490 
491    if (severity == LLVMDSError) {
492       *retval = 1;
493       fprintf(stderr, "LLVM triggered Diagnostic Handler: %s\n", description);
494    }
495 
496    LLVMDisposeMessage(description);
497 }
498 
499 static unsigned
radv_llvm_compile(LLVMModuleRef M,char ** pelf_buffer,size_t * pelf_size,struct ac_llvm_compiler * ac_llvm)500 radv_llvm_compile(LLVMModuleRef M, char **pelf_buffer, size_t *pelf_size, struct ac_llvm_compiler *ac_llvm)
501 {
502    unsigned retval = 0;
503    LLVMContextRef llvm_ctx;
504 
505    /* Setup Diagnostic Handler*/
506    llvm_ctx = LLVMGetModuleContext(M);
507 
508    LLVMContextSetDiagnosticHandler(llvm_ctx, ac_diagnostic_handler, &retval);
509 
510    /* Compile IR*/
511    if (!radv_compile_to_elf(ac_llvm, M, pelf_buffer, pelf_size))
512       retval = 1;
513    return retval;
514 }
515 
516 static void
ac_compile_llvm_module(struct ac_llvm_compiler * ac_llvm,LLVMModuleRef llvm_module,struct radv_shader_binary ** rbinary,const char * name,const struct radv_nir_compiler_options * options)517 ac_compile_llvm_module(struct ac_llvm_compiler *ac_llvm, LLVMModuleRef llvm_module, struct radv_shader_binary **rbinary,
518                        const char *name, const struct radv_nir_compiler_options *options)
519 {
520    char *elf_buffer = NULL;
521    size_t elf_size = 0;
522    char *llvm_ir_string = NULL;
523 
524    if (options->dump_shader) {
525       fprintf(stderr, "%s LLVM IR:\n\n", name);
526       ac_dump_module(llvm_module);
527       fprintf(stderr, "\n");
528    }
529 
530    if (options->record_ir) {
531       char *llvm_ir = LLVMPrintModuleToString(llvm_module);
532       llvm_ir_string = strdup(llvm_ir);
533       LLVMDisposeMessage(llvm_ir);
534    }
535 
536    int v = radv_llvm_compile(llvm_module, &elf_buffer, &elf_size, ac_llvm);
537    if (v) {
538       fprintf(stderr, "compile failed\n");
539    }
540 
541    LLVMContextRef ctx = LLVMGetModuleContext(llvm_module);
542    LLVMDisposeModule(llvm_module);
543    LLVMContextDispose(ctx);
544 
545    size_t llvm_ir_size = llvm_ir_string ? strlen(llvm_ir_string) : 0;
546    size_t alloc_size = sizeof(struct radv_shader_binary_rtld) + elf_size + llvm_ir_size + 1;
547    struct radv_shader_binary_rtld *rbin = calloc(1, alloc_size);
548    memcpy(rbin->data, elf_buffer, elf_size);
549    if (llvm_ir_string)
550       memcpy(rbin->data + elf_size, llvm_ir_string, llvm_ir_size + 1);
551 
552    rbin->base.type = RADV_BINARY_TYPE_RTLD;
553    rbin->base.total_size = alloc_size;
554    rbin->elf_size = elf_size;
555    rbin->llvm_ir_size = llvm_ir_size;
556    *rbinary = &rbin->base;
557 
558    free(llvm_ir_string);
559    free(elf_buffer);
560 }
561 
562 static void
radv_compile_nir_shader(struct ac_llvm_compiler * ac_llvm,const struct radv_nir_compiler_options * options,const struct radv_shader_info * info,struct radv_shader_binary ** rbinary,const struct radv_shader_args * args,struct nir_shader * const * nir,int nir_count)563 radv_compile_nir_shader(struct ac_llvm_compiler *ac_llvm, const struct radv_nir_compiler_options *options,
564                         const struct radv_shader_info *info, struct radv_shader_binary **rbinary,
565                         const struct radv_shader_args *args, struct nir_shader *const *nir, int nir_count)
566 {
567 
568    LLVMModuleRef llvm_module;
569 
570    llvm_module = ac_translate_nir_to_llvm(ac_llvm, options, info, nir, nir_count, args);
571 
572    ac_compile_llvm_module(ac_llvm, llvm_module, rbinary, radv_get_shader_name(info, nir[nir_count - 1]->info.stage),
573                           options);
574 }
575 
576 void
llvm_compile_shader(const struct radv_nir_compiler_options * options,const struct radv_shader_info * info,unsigned shader_count,struct nir_shader * const * shaders,struct radv_shader_binary ** binary,const struct radv_shader_args * args)577 llvm_compile_shader(const struct radv_nir_compiler_options *options, const struct radv_shader_info *info,
578                     unsigned shader_count, struct nir_shader *const *shaders, struct radv_shader_binary **binary,
579                     const struct radv_shader_args *args)
580 {
581    enum ac_target_machine_options tm_options = 0;
582    struct ac_llvm_compiler ac_llvm;
583 
584    tm_options |= AC_TM_SUPPORTS_SPILL;
585    if (options->check_ir)
586       tm_options |= AC_TM_CHECK_IR;
587 
588    radv_init_llvm_compiler(&ac_llvm, options->info->family, tm_options, info->wave_size);
589 
590    radv_compile_nir_shader(&ac_llvm, options, info, binary, args, shaders, shader_count);
591 }
592