LLVMBasicBlockRef loop_entry_block;
};
-#define CPU_STRING_LEN 30
-#define FS_STRING_LEN 30
-#define TRIPLE_STRING_LEN 7
-
-/**
- * Shader types for the LLVM backend.
- */
-enum si_llvm_shader_type {
- RADEON_LLVM_SHADER_PS = 0,
- RADEON_LLVM_SHADER_VS = 1,
- RADEON_LLVM_SHADER_GS = 2,
- RADEON_LLVM_SHADER_CS = 3,
-};
-
enum si_llvm_calling_convention {
RADEON_LLVM_AMDGPU_VS = 87,
RADEON_LLVM_AMDGPU_GS = 88,
LLVMAddTargetDependentFunctionAttr(F, name, str);
}
-/**
- * Set the shader type we want to compile
- *
- * @param type shader type to set
- */
-void si_llvm_shader_type(LLVMValueRef F, unsigned type)
-{
- enum si_llvm_shader_type llvm_type;
- enum si_llvm_calling_convention calling_conv;
-
- switch (type) {
- case PIPE_SHADER_VERTEX:
- case PIPE_SHADER_TESS_CTRL:
- case PIPE_SHADER_TESS_EVAL:
- llvm_type = RADEON_LLVM_SHADER_VS;
- calling_conv = RADEON_LLVM_AMDGPU_VS;
- break;
- case PIPE_SHADER_GEOMETRY:
- llvm_type = RADEON_LLVM_SHADER_GS;
- calling_conv = RADEON_LLVM_AMDGPU_GS;
- break;
- case PIPE_SHADER_FRAGMENT:
- llvm_type = RADEON_LLVM_SHADER_PS;
- calling_conv = RADEON_LLVM_AMDGPU_PS;
- break;
- case PIPE_SHADER_COMPUTE:
- llvm_type = RADEON_LLVM_SHADER_CS;
- calling_conv = RADEON_LLVM_AMDGPU_CS;
- break;
- default:
- unreachable("Unhandle shader type");
- }
-
- if (HAVE_LLVM >= 0x309)
- LLVMSetFunctionCallConv(F, calling_conv);
- else
- si_llvm_add_attribute(F, "ShaderType", llvm_type);
-}
-
static void init_amdgpu_target()
{
gallivm_init_llvm_targets();
LLVMInitializeAMDGPUTargetMC();
LLVMInitializeAMDGPUAsmPrinter();
+ /* For inline assembly. */
+ LLVMInitializeAMDGPUAsmParser();
+
if (HAVE_LLVM >= 0x0400) {
/*
* Workaround for bug in llvm 4.0 that causes image intrinsics
unsigned swizzle)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
-
+ LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef cval = ctx->system_values[reg->Register.Index];
+
+ if (tgsi_type_is_64bit(type)) {
+ LLVMValueRef lo, hi;
+
+ assert(swizzle == 0 || swizzle == 2);
+
+ lo = LLVMBuildExtractElement(
+ builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
+ hi = LLVMBuildExtractElement(
+ builder, cval, LLVMConstInt(ctx->i32, swizzle + 1, 0), "");
+
+ return si_llvm_emit_fetch_64bit(bld_base, type, lo, hi);
+ }
+
if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
- cval = LLVMBuildExtractElement(gallivm->builder, cval,
- LLVMConstInt(ctx->i32, swizzle, 0), "");
+ cval = LLVMBuildExtractElement(
+ builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
+ } else {
+ assert(swizzle == 0);
}
+
return bitcast(bld_base, type, cval);
}
* FIXME: We shouldn't need to have the non-alloca
* code path for arrays. LLVM should be smart enough to
* promote allocas into registers when profitable.
- *
- * LLVM 3.8 crashes with this.
*/
- if ((HAVE_LLVM >= 0x0309 && array_size > 16) ||
+ if (array_size > 16 ||
/* TODO: VGPR indexing is buggy on GFX9. */
ctx->screen->b.chip_class == GFX9) {
array_alloca = LLVMBuildAlloca(builder,
void si_llvm_context_init(struct si_shader_context *ctx,
struct si_screen *sscreen,
- struct si_shader *shader,
- LLVMTargetMachineRef tm,
- const struct tgsi_shader_info *info,
- const struct tgsi_token *tokens)
+ LLVMTargetMachineRef tm)
{
struct lp_type type;
* helper functions in the gallivm module.
*/
memset(ctx, 0, sizeof(*ctx));
- ctx->shader = shader;
ctx->screen = sscreen;
ctx->tm = tm;
- ctx->type = info ? info->processor : -1;
ctx->gallivm.context = LLVMContextCreate();
ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
ctx->gallivm.context);
LLVMSetTarget(ctx->gallivm.module, "amdgcn--");
-#if HAVE_LLVM >= 0x0309
LLVMTargetDataRef data_layout = LLVMCreateTargetDataLayout(tm);
char *data_layout_str = LLVMCopyStringRepOfTargetData(data_layout);
LLVMSetDataLayout(ctx->gallivm.module, data_layout_str);
LLVMDisposeTargetData(data_layout);
LLVMDisposeMessage(data_layout_str);
-#endif
bool unsafe_fpmath = (sscreen->b.debug_flags & DBG_UNSAFE_MATH) != 0;
enum lp_float_mode float_mode =
struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
- bld_base->info = info;
-
- if (info && info->array_max[TGSI_FILE_TEMPORARY] > 0) {
- int size = info->array_max[TGSI_FILE_TEMPORARY];
-
- ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
- ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
-
- if (tokens)
- tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
- ctx->temp_arrays);
- }
-
- if (info && info->file_max[TGSI_FILE_IMMEDIATE] >= 0) {
- int size = info->file_max[TGSI_FILE_IMMEDIATE] + 1;
- ctx->imms = MALLOC(size * TGSI_NUM_CHANNELS * sizeof(LLVMValueRef));
- }
-
type.floating = true;
type.fixed = false;
type.sign = true;
lp_build_context_init(&ctx->bld_base.int64_bld, &ctx->gallivm, lp_int_type(type));
bld_base->soa = 1;
- bld_base->emit_store = si_llvm_emit_store;
bld_base->emit_swizzle = emit_swizzle;
bld_base->emit_declaration = emit_declaration;
bld_base->emit_immediate = emit_immediate;
- bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = si_llvm_emit_fetch;
- bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = si_llvm_emit_fetch;
- bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = si_llvm_emit_fetch;
- bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = si_llvm_emit_fetch;
- bld_base->emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
-
/* metadata allowing 2.5 ULP */
ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->gallivm.context,
"fpmath", 6);
ctx->i64 = LLVMInt64TypeInContext(ctx->gallivm.context);
ctx->i128 = LLVMIntTypeInContext(ctx->gallivm.context, 128);
ctx->f32 = LLVMFloatTypeInContext(ctx->gallivm.context);
- ctx->v16i8 = LLVMVectorType(ctx->i8, 16);
ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
ctx->i32_1 = LLVMConstInt(ctx->i32, 1, 0);
}
+/* Set the context to a certain TGSI shader. Can be called repeatedly
+ * to change the shader. */
+void si_llvm_context_set_tgsi(struct si_shader_context *ctx,
+ struct si_shader *shader)
+{
+ const struct tgsi_shader_info *info = NULL;
+ const struct tgsi_token *tokens = NULL;
+
+ if (shader && shader->selector) {
+ info = &shader->selector->info;
+ tokens = shader->selector->tokens;
+ }
+
+ ctx->shader = shader;
+ ctx->type = info ? info->processor : -1;
+ ctx->bld_base.info = info;
+
+ /* Clean up the old contents. */
+ FREE(ctx->temp_arrays);
+ ctx->temp_arrays = NULL;
+ FREE(ctx->temp_array_allocas);
+ ctx->temp_array_allocas = NULL;
+
+ FREE(ctx->imms);
+ ctx->imms = NULL;
+ ctx->imms_num = 0;
+
+ FREE(ctx->temps);
+ ctx->temps = NULL;
+ ctx->temps_count = 0;
+
+ if (!info || !tokens)
+ return;
+
+ if (info->array_max[TGSI_FILE_TEMPORARY] > 0) {
+ int size = info->array_max[TGSI_FILE_TEMPORARY];
+
+ ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
+ ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
+
+ tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
+ ctx->temp_arrays);
+ }
+ if (info->file_max[TGSI_FILE_IMMEDIATE] >= 0) {
+ int size = info->file_max[TGSI_FILE_IMMEDIATE] + 1;
+ ctx->imms = MALLOC(size * TGSI_NUM_CHANNELS * sizeof(LLVMValueRef));
+ }
+
+ /* Re-set these to start with a clean slate. */
+ ctx->bld_base.num_instructions = 0;
+ ctx->bld_base.pc = 0;
+ memset(ctx->outputs, 0, sizeof(ctx->outputs));
+
+ ctx->bld_base.emit_store = si_llvm_emit_store;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = si_llvm_emit_fetch;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = si_llvm_emit_fetch;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = si_llvm_emit_fetch;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_OUTPUT] = si_llvm_emit_fetch;
+ ctx->bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
+}
+
void si_llvm_create_func(struct si_shader_context *ctx,
const char *name,
LLVMTypeRef *return_types, unsigned num_return_elems,
{
LLVMTypeRef main_fn_type, ret_type;
LLVMBasicBlockRef main_fn_body;
+ enum si_llvm_calling_convention call_conv;
if (num_return_elems)
ret_type = LLVMStructTypeInContext(ctx->gallivm.context,
main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
ctx->main_fn, "main_body");
LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
+
+ switch (ctx->type) {
+ case PIPE_SHADER_VERTEX:
+ case PIPE_SHADER_TESS_CTRL:
+ case PIPE_SHADER_TESS_EVAL:
+ call_conv = RADEON_LLVM_AMDGPU_VS;
+ break;
+ case PIPE_SHADER_GEOMETRY:
+ call_conv = RADEON_LLVM_AMDGPU_GS;
+ break;
+ case PIPE_SHADER_FRAGMENT:
+ call_conv = RADEON_LLVM_AMDGPU_PS;
+ break;
+ case PIPE_SHADER_COMPUTE:
+ call_conv = RADEON_LLVM_AMDGPU_CS;
+ break;
+ default:
+ unreachable("Unhandle shader type");
+ }
+
+ LLVMSetFunctionCallConv(ctx->main_fn, call_conv);
}
-void si_llvm_finalize_module(struct si_shader_context *ctx,
- bool run_verifier)
+void si_llvm_optimize_module(struct si_shader_context *ctx)
{
struct gallivm_state *gallivm = &ctx->gallivm;
const char *triple = LLVMGetTarget(gallivm->module);
LLVMTargetLibraryInfoRef target_library_info;
+ /* Dump LLVM IR before any optimization passes */
+ if (ctx->screen->b.debug_flags & DBG_PREOPT_IR &&
+ r600_can_dump_shader(&ctx->screen->b, ctx->type))
+ LLVMDumpModule(ctx->gallivm.module);
+
/* Create the pass manager */
gallivm->passmgr = LLVMCreatePassManager();
target_library_info = gallivm_create_target_library_info(triple);
LLVMAddTargetLibraryInfo(target_library_info, gallivm->passmgr);
- if (run_verifier)
+ if (r600_extra_shader_checks(&ctx->screen->b, ctx->type))
LLVMAddVerifierPass(gallivm->passmgr);
LLVMAddAlwaysInlinerPass(gallivm->passmgr);