[Mesa-dev] [PATCH v2 61/73] ac/nir: add always_vector argument to ac_build_gather_values_extended
Nicolai Hähnle
nhaehnle at gmail.com
Wed Jul 5 10:48:45 UTC 2017
From: Nicolai Hähnle <nicolai.haehnle at amd.com>
This simplifies a bunch of places that no longer need special treatment
of value_count == 1. We rely on LLVM to optimize away the 1-element vector
types.
This fixes a bunch of bugs where 1-element arrays are indexed indirectly.
---
src/amd/common/ac_llvm_build.c | 7 ++++---
src/amd/common/ac_llvm_build.h | 3 ++-
src/amd/common/ac_nir_to_llvm.c | 22 +++++++---------------
3 files changed, 13 insertions(+), 19 deletions(-)
diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 2cc4eae..2dddf44 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -174,27 +174,28 @@ void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
snprintf(buf, bufsize, "f64");
break;
}
}
LLVMValueRef
ac_build_gather_values_extended(struct ac_llvm_context *ctx,
LLVMValueRef *values,
unsigned value_count,
unsigned value_stride,
- bool load)
+ bool load,
+ bool always_vector)
{
LLVMBuilderRef builder = ctx->builder;
LLVMValueRef vec = NULL;
unsigned i;
- if (value_count == 1) {
+ if (value_count == 1 && !always_vector) {
if (load)
return LLVMBuildLoad(builder, values[0], "");
return values[0];
} else if (!value_count)
unreachable("value_count is 0");
for (i = 0; i < value_count; i++) {
LLVMValueRef value = values[i * value_stride];
if (load)
value = LLVMBuildLoad(builder, value, "");
@@ -205,21 +206,21 @@ ac_build_gather_values_extended(struct ac_llvm_context *ctx,
vec = LLVMBuildInsertElement(builder, vec, value, index, "");
}
return vec;
}
LLVMValueRef
ac_build_gather_values(struct ac_llvm_context *ctx,
LLVMValueRef *values,
unsigned value_count)
{
- return ac_build_gather_values_extended(ctx, values, value_count, 1, false);
+ return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
}
LLVMValueRef
ac_build_fdiv(struct ac_llvm_context *ctx,
LLVMValueRef num,
LLVMValueRef den)
{
LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
if (!LLVMIsConstant(ret))
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 10efabb..f25af8b 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -71,21 +71,22 @@ ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
LLVMTypeRef return_type, LLVMValueRef *params,
unsigned param_count, unsigned attrib_mask);
void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
LLVMValueRef
ac_build_gather_values_extended(struct ac_llvm_context *ctx,
LLVMValueRef *values,
unsigned value_count,
unsigned value_stride,
- bool load);
+ bool load,
+ bool always_vector);
LLVMValueRef
ac_build_gather_values(struct ac_llvm_context *ctx,
LLVMValueRef *values,
unsigned value_count);
LLVMValueRef
ac_build_fdiv(struct ac_llvm_context *ctx,
LLVMValueRef num,
LLVMValueRef den);
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index bc3cb2d..c5b28d1 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -1024,25 +1024,20 @@ static LLVMValueRef trim_vector(struct ac_llvm_context *ctx,
static void
build_store_values_extended(struct ac_llvm_context *ac,
LLVMValueRef *values,
unsigned value_count,
unsigned value_stride,
LLVMValueRef vec)
{
LLVMBuilderRef builder = ac->builder;
unsigned i;
- if (value_count == 1) {
- LLVMBuildStore(builder, vec, values[0]);
- return;
- }
-
for (i = 0; i < value_count; i++) {
LLVMValueRef ptr = values[i * value_stride];
LLVMValueRef index = LLVMConstInt(ac->i32, i, false);
LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
LLVMBuildStore(builder, value, ptr);
}
}
static LLVMTypeRef get_def_type(struct ac_nir_context *ctx,
const nir_ssa_def *def)
@@ -2954,58 +2949,58 @@ static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
return load_gs_input(ctx->nctx, instr);
}
for (unsigned chan = 0; chan < ve; chan++) {
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type,
ctx->stage == MESA_SHADER_VERTEX);
count -= chan / 4;
LLVMValueRef tmp_vec = ac_build_gather_values_extended(
&ctx->ac, ctx->abi->inputs + idx + chan, count,
- 4, false);
+ 4, false, true);
values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
tmp_vec,
indir_index, "");
} else
values[chan] = ctx->abi->inputs[idx + chan + const_index * 4];
}
break;
case nir_var_local:
for (unsigned chan = 0; chan < ve; chan++) {
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type, false);
count -= chan / 4;
LLVMValueRef tmp_vec = ac_build_gather_values_extended(
&ctx->ac, ctx->locals + idx + chan, count,
- 4, true);
+ 4, true, true);
values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
tmp_vec,
indir_index, "");
} else {
values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * 4], "");
}
}
break;
case nir_var_shader_out:
if (ctx->stage == MESA_SHADER_TESS_CTRL)
return load_tcs_output(ctx->nctx, instr);
for (unsigned chan = 0; chan < ve; chan++) {
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type, false);
count -= chan / 4;
LLVMValueRef tmp_vec = ac_build_gather_values_extended(
&ctx->ac, ctx->outputs + idx + chan, count,
- 4, true);
+ 4, true, true);
values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
tmp_vec,
indir_index, "");
} else {
values[chan] = LLVMBuildLoad(ctx->ac.builder,
ctx->outputs[idx + chan + const_index * 4],
"");
}
}
@@ -3077,27 +3072,24 @@ visit_store_var(struct ac_nir_context *ctx,
value = llvm_extract_elem(&ctx->ac, src, chan);
if (instr->variables[0]->var->data.compact)
stride = 1;
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type, false);
count -= chan / 4;
LLVMValueRef tmp_vec = ac_build_gather_values_extended(
&ctx->ac, ctx->outputs + idx + chan, count,
- stride, true);
+ stride, true, true);
- if (get_llvm_num_components(tmp_vec) > 1) {
- tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
- value, indir_index, "");
- } else
- tmp_vec = value;
+ tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
+ value, indir_index, "");
build_store_values_extended(&ctx->ac, ctx->outputs + idx + chan,
count, stride, tmp_vec);
} else {
temp_ptr = ctx->outputs[idx + chan + const_index * stride];
LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
}
}
break;
@@ -3106,21 +3098,21 @@ visit_store_var(struct ac_nir_context *ctx,
if (!(writemask & (1 << chan)))
continue;
value = llvm_extract_elem(&ctx->ac, src, chan);
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type, false);
count -= chan / 4;
LLVMValueRef tmp_vec = ac_build_gather_values_extended(
&ctx->ac, ctx->locals + idx + chan, count,
- 4, true);
+ 4, true, true);
tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
value, indir_index, "");
build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
count, 4, tmp_vec);
} else {
temp_ptr = ctx->locals[idx + chan + const_index * 4];
LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
}
--
2.9.3
More information about the mesa-dev
mailing list