#include <stdio.h>
#include <llvm-c/Transforms/IPO.h>
#include <llvm-c/Transforms/Scalar.h>
-#include <llvm-c/Support.h>
/* Data for if/else/endif and bgnloop/endloop control flow structures.
*/
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,
RADEON_LLVM_AMDGPU_PS = 89,
RADEON_LLVM_AMDGPU_CS = 90,
+ RADEON_LLVM_AMDGPU_HS = 93,
};
void si_llvm_add_attribute(LLVMValueRef F, const char *name, int value)
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();
- LLVMInitializeAMDGPUTargetInfo();
- LLVMInitializeAMDGPUTarget();
- LLVMInitializeAMDGPUTargetMC();
- LLVMInitializeAMDGPUAsmPrinter();
-
- if (HAVE_LLVM >= 0x0400) {
- /*
- * Workaround for bug in llvm 4.0 that causes image intrinsics
- * to disappear.
- * https://reviews.llvm.org/D26348
- */
- const char *argv[2] = {"mesa", "-simplifycfg-sink-common=false"};
- LLVMParseCommandLineOptions(2, argv, NULL);
- }
-}
-
-static once_flag init_amdgpu_target_once_flag = ONCE_FLAG_INIT;
-
-LLVMTargetRef si_llvm_get_amdgpu_target(const char *triple)
-{
- LLVMTargetRef target = NULL;
- char *err_message = NULL;
-
- call_once(&init_amdgpu_target_once_flag, init_amdgpu_target);
-
- if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
- fprintf(stderr, "Cannot find target for triple %s ", triple);
- if (err_message) {
- fprintf(stderr, "%s\n", err_message);
- }
- LLVMDisposeMessage(err_message);
- return NULL;
- }
- return target;
-}
-
struct si_llvm_diagnostics {
struct pipe_debug_callback *debug;
unsigned retval;
buffer_size = LLVMGetBufferSize(out_buffer);
buffer_data = LLVMGetBufferStart(out_buffer);
- ac_elf_read(buffer_data, buffer_size, binary);
+ if (!ac_elf_read(buffer_data, buffer_size, binary)) {
+ fprintf(stderr, "radeonsi: cannot read an ELF shader binary\n");
+ diag.retval = 1;
+ }
/* Clean up */
LLVMDisposeMemoryBuffer(out_buffer);
LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
enum tgsi_opcode_type type)
{
- LLVMContextRef ctx = bld_base->base.gallivm->context;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
switch (type) {
case TGSI_TYPE_UNSIGNED:
case TGSI_TYPE_SIGNED:
- return LLVMInt32TypeInContext(ctx);
+ return ctx->ac.i32;
case TGSI_TYPE_UNSIGNED64:
case TGSI_TYPE_SIGNED64:
- return LLVMInt64TypeInContext(ctx);
+ return ctx->ac.i64;
case TGSI_TYPE_DOUBLE:
- return LLVMDoubleTypeInContext(ctx);
+ return ctx->ac.f64;
case TGSI_TYPE_UNTYPED:
case TGSI_TYPE_FLOAT:
- return LLVMFloatTypeInContext(ctx);
+ return ctx->ac.f32;
default: break;
}
return 0;
LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
enum tgsi_opcode_type type, LLVMValueRef value)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMTypeRef dst_type = tgsi2llvmtype(bld_base, type);
if (dst_type)
- return LLVMBuildBitCast(builder, value, dst_type, "");
+ return LLVMBuildBitCast(ctx->ac.builder, value, dst_type, "");
else
return value;
}
LLVMValueRef index,
unsigned num)
{
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef c_max = LLVMConstInt(ctx->i32, num - 1, 0);
LLVMValueRef cc;
unsigned swizzle_z,
unsigned swizzle_w)
{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef swizzles[4];
- LLVMTypeRef i32t =
- LLVMInt32TypeInContext(bld_base->base.gallivm->context);
- swizzles[0] = LLVMConstInt(i32t, swizzle_x, 0);
- swizzles[1] = LLVMConstInt(i32t, swizzle_y, 0);
- swizzles[2] = LLVMConstInt(i32t, swizzle_z, 0);
- swizzles[3] = LLVMConstInt(i32t, swizzle_w, 0);
+ swizzles[0] = LLVMConstInt(ctx->i32, swizzle_x, 0);
+ swizzles[1] = LLVMConstInt(ctx->i32, swizzle_y, 0);
+ swizzles[2] = LLVMConstInt(ctx->i32, swizzle_z, 0);
+ swizzles[3] = LLVMConstInt(ctx->i32, swizzle_w, 0);
- return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
+ return LLVMBuildShuffleVector(ctx->ac.builder,
value,
LLVMGetUndef(LLVMTypeOf(value)),
LLVMConstVector(swizzles, 4), "");
return range;
}
-static LLVMValueRef
-emit_array_index(struct si_shader_context *ctx,
- const struct tgsi_ind_register *reg,
- unsigned offset)
-{
- struct gallivm_state *gallivm = &ctx->gallivm;
-
- if (!reg) {
- return LLVMConstInt(ctx->i32, offset, 0);
- }
- LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, ctx->addrs[reg->Index][reg->Swizzle], "");
- return LLVMBuildAdd(gallivm->builder, addr, LLVMConstInt(ctx->i32, offset, 0), "");
-}
-
/**
* For indirect registers, construct a pointer directly to the requested
* element using getelementptr if possible.
{
unsigned array_id;
struct tgsi_array_info *array;
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef idxs[2];
LLVMValueRef index;
LLVMValueRef alloca;
if (!(array->writemask & (1 << swizzle)))
return ctx->undef_alloca;
- index = emit_array_index(ctx, reg_indirect,
- reg_index - ctx->temp_arrays[array_id - 1].range.First);
+ index = si_get_indirect_index(ctx, reg_indirect, 1,
+ reg_index - ctx->temp_arrays[array_id - 1].range.First);
/* Ensure that the index is within a valid range, to guard against
* VM faults and overwriting critical data (e.g. spilled resource
"");
idxs[0] = ctx->i32_0;
idxs[1] = index;
- return LLVMBuildGEP(builder, alloca, idxs, 2, "");
+ return LLVMBuildGEP(ctx->ac.builder, alloca, idxs, 2, "");
}
LLVMValueRef
si_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
- enum tgsi_opcode_type type,
+ LLVMTypeRef type,
LLVMValueRef ptr,
LLVMValueRef ptr2)
{
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef result;
- result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
+ result = LLVMGetUndef(LLVMVectorType(ctx->i32, 2));
- result = LLVMBuildInsertElement(builder,
+ result = LLVMBuildInsertElement(ctx->ac.builder,
result,
- bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr),
- bld_base->int_bld.zero, "");
- result = LLVMBuildInsertElement(builder,
+ ac_to_integer(&ctx->ac, ptr),
+ ctx->i32_0, "");
+ result = LLVMBuildInsertElement(ctx->ac.builder,
result,
- bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr2),
- bld_base->int_bld.one, "");
- return bitcast(bld_base, type, result);
+ ac_to_integer(&ctx->ac, ptr2),
+ ctx->i32_1, "");
+ return LLVMBuildBitCast(ctx->ac.builder, result, type, "");
}
static LLVMValueRef
unsigned swizzle)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
-
- LLVMBuilderRef builder = ctx->gallivm.builder;
-
unsigned i, size = range.Last - range.First + 1;
LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
LLVMValueRef result = LLVMGetUndef(vec);
for (i = 0; i < size; ++i) {
tmp_reg.Register.Index = i + range.First;
LLVMValueRef temp = si_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
- result = LLVMBuildInsertElement(builder, result, temp,
+ result = LLVMBuildInsertElement(ctx->ac.builder, result, temp,
LLVMConstInt(ctx->i32, i, 0), "array_vector");
}
return result;
const struct tgsi_ind_register *reg_indirect)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef ptr;
ptr = get_pointer_into_array(ctx, file, swizzle, reg_index, reg_indirect);
LLVMValueRef ptr_hi, val_hi;
ptr_hi = LLVMBuildGEP(builder, ptr, &ctx->i32_1, 1, "");
val_hi = LLVMBuildLoad(builder, ptr_hi, "");
- val = si_llvm_emit_fetch_64bit(bld_base, type, val, val_hi);
+ val = si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
+ val, val_hi);
}
return val;
struct tgsi_declaration_range range =
get_array_range(bld_base, file, reg_index, reg_indirect);
LLVMValueRef index =
- emit_array_index(ctx, reg_indirect, reg_index - range.First);
+ si_get_indirect_index(ctx, reg_indirect, 1, reg_index - range.First);
LLVMValueRef array =
emit_array_fetch(bld_base, file, type, range, swizzle);
return LLVMBuildExtractElement(builder, array, index, "");
const struct tgsi_ind_register *reg_indirect)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMBuilderRef builder = gallivm->builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef ptr;
ptr = get_pointer_into_array(ctx, file, chan_index, reg_index, reg_indirect);
} else {
unsigned i, size;
struct tgsi_declaration_range range = get_array_range(bld_base, file, reg_index, reg_indirect);
- LLVMValueRef index = emit_array_index(ctx, reg_indirect, reg_index - range.First);
+ LLVMValueRef index = si_get_indirect_index(ctx, reg_indirect, 1, reg_index - range.First);
LLVMValueRef array =
emit_array_fetch(bld_base, file, TGSI_TYPE_FLOAT, range, chan_index);
LLVMValueRef temp_ptr;
unsigned swizzle)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- LLVMBuilderRef builder = ctx->gallivm.builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef result = NULL, ptr, ptr2;
if (swizzle == ~0) {
case TGSI_FILE_IMMEDIATE: {
LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
if (tgsi_type_is_64bit(type)) {
- result = LLVMGetUndef(LLVMVectorType(ctx->i32, bld_base->base.type.length * 2));
+ result = LLVMGetUndef(LLVMVectorType(ctx->i32, 2));
result = LLVMConstInsertElement(result,
ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle],
ctx->i32_0);
if (tgsi_type_is_64bit(type)) {
ptr = result;
ptr2 = input[swizzle + 1];
- return si_llvm_emit_fetch_64bit(bld_base, type, ptr, ptr2);
+ return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
+ ptr, ptr2);
}
break;
}
ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
if (tgsi_type_is_64bit(type)) {
ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1];
- return si_llvm_emit_fetch_64bit(bld_base, type,
+ return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
LLVMBuildLoad(builder, ptr, ""),
LLVMBuildLoad(builder, ptr2, ""));
}
ptr = get_output_ptr(bld_base, reg->Register.Index, swizzle);
if (tgsi_type_is_64bit(type)) {
ptr2 = get_output_ptr(bld_base, reg->Register.Index, swizzle + 1);
- return si_llvm_emit_fetch_64bit(bld_base, type,
+ return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
LLVMBuildLoad(builder, ptr, ""),
LLVMBuildLoad(builder, ptr2, ""));
}
unsigned swizzle)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- LLVMBuilderRef builder = ctx->gallivm.builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef cval = ctx->system_values[reg->Register.Index];
if (tgsi_type_is_64bit(type)) {
hi = LLVMBuildExtractElement(
builder, cval, LLVMConstInt(ctx->i32, swizzle + 1, 0), "");
- return si_llvm_emit_fetch_64bit(bld_base, type, lo, hi);
+ return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
+ lo, hi);
}
if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
const struct tgsi_full_declaration *decl)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- LLVMBuilderRef builder = ctx->gallivm.builder;
+ LLVMBuilderRef builder = ctx->ac.builder;
unsigned first, last, i;
switch(decl->Declaration.File) {
case TGSI_FILE_ADDRESS:
* 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) ||
- /* TODO: VGPR indexing is buggy on GFX9. */
- ctx->screen->b.chip_class == GFX9) {
- array_alloca = LLVMBuildAlloca(builder,
+ if (array_size > 16 ||
+ !ctx->screen->llvm_has_working_vgpr_indexing) {
+ array_alloca = lp_build_alloca_undef(&ctx->gallivm,
LLVMArrayType(ctx->f32,
array_size), "array");
ctx->temp_array_allocas[id] = array_alloca;
{
unsigned idx;
for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
- ctx->load_system_value(ctx, idx, decl);
+ si_load_system_value(ctx, idx, decl);
}
}
break;
}
case TGSI_FILE_MEMORY:
- ctx->declare_memory_region(ctx, decl);
+ si_declare_compute_memory(ctx, decl);
break;
default:
void si_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
const struct tgsi_full_instruction *inst,
const struct tgsi_opcode_info *info,
+ unsigned index,
LLVMValueRef dst[4])
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
- const struct tgsi_full_dst_register *reg = &inst->Dst[0];
- LLVMBuilderRef builder = ctx->gallivm.builder;
+ const struct tgsi_full_dst_register *reg = &inst->Dst[index];
+ LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef temp_ptr, temp_ptr2 = NULL;
- unsigned chan, chan_index;
bool is_vec_store = false;
- enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
+ enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode, index);
if (dst[0]) {
LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
if (is_vec_store) {
LLVMValueRef values[4] = {};
- TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan) {
+ uint32_t writemask = reg->Register.WriteMask;
+ while (writemask) {
+ unsigned chan = u_bit_scan(&writemask);
LLVMValueRef index = LLVMConstInt(ctx->i32, chan, 0);
- values[chan] = LLVMBuildExtractElement(gallivm->builder,
+ values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
dst[0], index, "");
}
- bld_base->emit_store(bld_base, inst, info, values);
+ bld_base->emit_store(bld_base, inst, info, index, values);
return;
}
- TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
+ uint32_t writemask = reg->Register.WriteMask;
+ while (writemask) {
+ unsigned chan_index = u_bit_scan(&writemask);
LLVMValueRef value = dst[chan_index];
if (tgsi_type_is_64bit(dtype) && (chan_index == 1 || chan_index == 3))
}
if (!tgsi_type_is_64bit(dtype))
- value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
+ value = ac_to_float(&ctx->ac, value);
if (reg->Register.Indirect) {
unsigned file = reg->Register.File;
val2 = LLVMBuildExtractElement(builder, ptr,
ctx->i32_1, "");
- LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, value), temp_ptr);
- LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, val2), temp_ptr2);
+ LLVMBuildStore(builder, ac_to_float(&ctx->ac, value), temp_ptr);
+ LLVMBuildStore(builder, ac_to_float(&ctx->ac, val2), temp_ptr2);
}
}
}
static LLVMBasicBlockRef append_basic_block(struct si_shader_context *ctx,
const char *name)
{
- struct gallivm_state *gallivm = &ctx->gallivm;
-
assert(ctx->flow_depth >= 1);
if (ctx->flow_depth >= 2) {
struct si_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
- return LLVMInsertBasicBlockInContext(gallivm->context,
+ return LLVMInsertBasicBlockInContext(ctx->ac.context,
flow->next_block, name);
}
- return LLVMAppendBasicBlockInContext(gallivm->context, ctx->main_fn, name);
+ return LLVMAppendBasicBlockInContext(ctx->ac.context, ctx->main_fn, name);
}
/* Emit a branch to the given default target for the current block if
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *flow = push_flow(ctx);
flow->loop_entry_block = append_basic_block(ctx, "LOOP");
flow->next_block = append_basic_block(ctx, "ENDLOOP");
set_basicblock_name(flow->loop_entry_block, "loop", bld_base->pc);
- LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
- LLVMPositionBuilderAtEnd(gallivm->builder, flow->loop_entry_block);
+ LLVMBuildBr(ctx->ac.builder, flow->loop_entry_block);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, flow->loop_entry_block);
}
static void brk_emit(const struct lp_build_tgsi_action *action,
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *flow = get_innermost_loop(ctx);
- LLVMBuildBr(gallivm->builder, flow->next_block);
+ LLVMBuildBr(ctx->ac.builder, flow->next_block);
}
static void cont_emit(const struct lp_build_tgsi_action *action,
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *flow = get_innermost_loop(ctx);
- LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
+ LLVMBuildBr(ctx->ac.builder, flow->loop_entry_block);
}
static void else_emit(const struct lp_build_tgsi_action *action,
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *current_branch = get_current_flow(ctx);
LLVMBasicBlockRef endif_block;
assert(!current_branch->loop_entry_block);
endif_block = append_basic_block(ctx, "ENDIF");
- emit_default_branch(gallivm->builder, endif_block);
+ emit_default_branch(ctx->ac.builder, endif_block);
- LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, current_branch->next_block);
set_basicblock_name(current_branch->next_block, "else", bld_base->pc);
current_branch->next_block = endif_block;
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *current_branch = get_current_flow(ctx);
assert(!current_branch->loop_entry_block);
- emit_default_branch(gallivm->builder, current_branch->next_block);
- LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
+ emit_default_branch(ctx->ac.builder, current_branch->next_block);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, current_branch->next_block);
set_basicblock_name(current_branch->next_block, "endif", bld_base->pc);
ctx->flow_depth--;
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *current_loop = get_current_flow(ctx);
assert(current_loop->loop_entry_block);
- emit_default_branch(gallivm->builder, current_loop->loop_entry_block);
+ emit_default_branch(ctx->ac.builder, current_loop->loop_entry_block);
- LLVMPositionBuilderAtEnd(gallivm->builder, current_loop->next_block);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, current_loop->next_block);
set_basicblock_name(current_loop->next_block, "endloop", bld_base->pc);
ctx->flow_depth--;
}
LLVMValueRef cond)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = &ctx->gallivm;
struct si_llvm_flow *flow = push_flow(ctx);
LLVMBasicBlockRef if_block;
if_block = append_basic_block(ctx, "IF");
flow->next_block = append_basic_block(ctx, "ELSE");
set_basicblock_name(if_block, "if", bld_base->pc);
- LLVMBuildCondBr(gallivm->builder, cond, if_block, flow->next_block);
- LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
+ LLVMBuildCondBr(ctx->ac.builder, cond, if_block, flow->next_block);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, if_block);
}
static void if_emit(const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef cond;
- cond = LLVMBuildFCmp(gallivm->builder, LLVMRealUNE,
+ cond = LLVMBuildFCmp(ctx->ac.builder, LLVMRealUNE,
emit_data->args[0],
- bld_base->base.zero, "");
+ ctx->ac.f32_0, "");
if_cond_emit(action, bld_base, emit_data, cond);
}
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef cond;
- cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
- bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
- bld_base->int_bld.zero, "");
+ cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE,
+ ac_to_integer(&ctx->ac, emit_data->args[0]), ctx->i32_0, "");
if_cond_emit(action, bld_base, emit_data, cond);
}
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 =
- unsafe_fpmath ? LP_FLOAT_MODE_UNSAFE_FP_MATH :
- LP_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH;
+ bool unsafe_fpmath = (sscreen->debug_flags & DBG(UNSAFE_MATH)) != 0;
+ enum ac_float_mode float_mode =
+ unsafe_fpmath ? AC_FLOAT_MODE_UNSAFE_FP_MATH :
+ AC_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH;
- ctx->gallivm.builder = lp_create_builder(ctx->gallivm.context,
+ ctx->gallivm.builder = ac_create_builder(ctx->gallivm.context,
float_mode);
- ac_llvm_context_init(&ctx->ac, ctx->gallivm.context);
+ ac_llvm_context_init(&ctx->ac, ctx->gallivm.context,
+ sscreen->info.chip_class, sscreen->info.family);
ctx->ac.module = ctx->gallivm.module;
ctx->ac.builder = ctx->gallivm.builder;
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);
- LLVMValueRef arg = lp_build_const_float(&ctx->gallivm, 2.5);
- ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->gallivm.context,
- &arg, 1);
-
bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
si_shader_context_init_alu(&ctx->bld_base);
-
- ctx->voidt = LLVMVoidTypeInContext(ctx->gallivm.context);
- ctx->i1 = LLVMInt1TypeInContext(ctx->gallivm.context);
- ctx->i8 = LLVMInt8TypeInContext(ctx->gallivm.context);
- ctx->i32 = LLVMInt32TypeInContext(ctx->gallivm.context);
- 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);
+ si_shader_context_init_mem(ctx);
+
+ ctx->voidt = LLVMVoidTypeInContext(ctx->ac.context);
+ ctx->i1 = LLVMInt1TypeInContext(ctx->ac.context);
+ ctx->i8 = LLVMInt8TypeInContext(ctx->ac.context);
+ ctx->i32 = LLVMInt32TypeInContext(ctx->ac.context);
+ ctx->i64 = LLVMInt64TypeInContext(ctx->ac.context);
+ ctx->i128 = LLVMIntTypeInContext(ctx->ac.context, 128);
+ ctx->f32 = LLVMFloatTypeInContext(ctx->ac.context);
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;
+
+ ctx->num_const_buffers = util_last_bit(info->const_buffers_declared);
+ ctx->num_shader_buffers = util_last_bit(info->shader_buffers_declared);
+ ctx->num_samplers = util_last_bit(info->samplers_declared);
+ ctx->num_images = util_last_bit(info->images_declared);
+}
+
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;
+ unsigned real_shader_type;
if (num_return_elems)
- ret_type = LLVMStructTypeInContext(ctx->gallivm.context,
+ ret_type = LLVMStructTypeInContext(ctx->ac.context,
return_types,
num_return_elems, true);
else
- ret_type = LLVMVoidTypeInContext(ctx->gallivm.context);
+ ret_type = ctx->voidt;
/* Setup the function */
ctx->return_type = ret_type;
main_fn_type = LLVMFunctionType(ret_type, ParamTypes, ParamCount, 0);
ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, name, main_fn_type);
- main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
+ main_fn_body = LLVMAppendBasicBlockInContext(ctx->ac.context,
ctx->main_fn, "main_body");
- LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
+ LLVMPositionBuilderAtEnd(ctx->ac.builder, main_fn_body);
+
+ real_shader_type = ctx->type;
+
+ /* LS is merged into HS (TCS), and ES is merged into GS. */
+ if (ctx->screen->info.chip_class >= GFX9) {
+ if (ctx->shader->key.as_ls)
+ real_shader_type = PIPE_SHADER_TESS_CTRL;
+ else if (ctx->shader->key.as_es)
+ real_shader_type = PIPE_SHADER_GEOMETRY;
+ }
+
+ switch (real_shader_type) {
+ case PIPE_SHADER_VERTEX:
+ case PIPE_SHADER_TESS_EVAL:
+ call_conv = RADEON_LLVM_AMDGPU_VS;
+ break;
+ case PIPE_SHADER_TESS_CTRL:
+ call_conv = HAVE_LLVM >= 0x0500 ? RADEON_LLVM_AMDGPU_HS :
+ 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->debug_flags & DBG(PREOPT_IR) &&
+ si_can_dump_shader(ctx->screen, 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 (si_extra_shader_checks(ctx->screen, ctx->type))
LLVMAddVerifierPass(gallivm->passmgr);
LLVMAddAlwaysInlinerPass(gallivm->passmgr);
LLVMAddLICMPass(gallivm->passmgr);
LLVMAddAggressiveDCEPass(gallivm->passmgr);
LLVMAddCFGSimplificationPass(gallivm->passmgr);
+#if HAVE_LLVM >= 0x0400
+ /* This is recommended by the instruction combining pass. */
+ LLVMAddEarlyCSEMemSSAPass(gallivm->passmgr);
+#endif
LLVMAddInstructionCombiningPass(gallivm->passmgr);
/* Run the pass */
LLVMRunPassManager(gallivm->passmgr, ctx->gallivm.module);
- LLVMDisposeBuilder(gallivm->builder);
+ LLVMDisposeBuilder(ctx->ac.builder);
LLVMDisposePassManager(gallivm->passmgr);
gallivm_dispose_target_library_info(target_library_info);
}
projects / mesa.git / blobdiff