From: Nicolai Hähnle <nicolai.haeh...@amd.com>
... and eliminate the non-ac copies. Mostly straight-forward
search & replace.
---
src/amd/common/ac_nir_to_llvm.c | 89 +++++++++++------------------------------
1 file changed, 24 insertions(+), 65 deletions(-)
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index fefd5e7..5fe9e8c 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -624,53 +624,20 @@ static LLVMValueRef trim_vector(struct
nir_to_llvm_context *ctx,
LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false)};
if (count == 1)
return LLVMBuildExtractElement(ctx->builder, value, masks[0],
"");
LLVMValueRef swizzle = LLVMConstVector(masks, count);
return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
}
-static LLVMValueRef
-build_gather_values_extended(struct nir_to_llvm_context *ctx,
- LLVMValueRef *values,
- unsigned value_count,
- unsigned value_stride,
- bool load)
-{
- LLVMBuilderRef builder = ctx->builder;
- LLVMValueRef vec;
- unsigned i;
-
-
- if (value_count == 1) {
- 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, "");
-
- if (!i)
- vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value),
value_count));
- LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
- vec = LLVMBuildInsertElement(builder, vec, value, index, "");
- }
- return vec;
-}
-
-
static void
build_store_values_extended(struct nir_to_llvm_context *ctx,
LLVMValueRef *values,
unsigned value_count,
unsigned value_stride,
LLVMValueRef vec)
{
LLVMBuilderRef builder = ctx->builder;
unsigned i;
@@ -680,28 +647,20 @@ build_store_values_extended(struct nir_to_llvm_context
*ctx,
}
for (i = 0; i < value_count; i++) {
LLVMValueRef ptr = values[i * value_stride];
LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
LLVMValueRef value = LLVMBuildExtractElement(builder, vec,
index, "");
LLVMBuildStore(builder, value, ptr);
}
}
-static LLVMValueRef
-build_gather_values(struct nir_to_llvm_context *ctx,
- LLVMValueRef *values,
- unsigned value_count)
-{
- return build_gather_values_extended(ctx, values, value_count, 1, false);
-}
-
static LLVMTypeRef get_def_type(struct nir_to_llvm_context *ctx,
nir_ssa_def *def)
{
LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, def->bit_size);
if (def->num_components > 1) {
type = LLVMVectorType(type, def->num_components);
}
return type;
}
@@ -744,21 +703,21 @@ static LLVMValueRef get_alu_src(struct
nir_to_llvm_context *ctx,
LLVMConstInt(ctx->i32, src.swizzle[0], false),
LLVMConstInt(ctx->i32, src.swizzle[1], false),
LLVMConstInt(ctx->i32, src.swizzle[2], false),
LLVMConstInt(ctx->i32, src.swizzle[3], false)};
if (src_components > 1 && num_components == 1) {
value = LLVMBuildExtractElement(ctx->builder, value,
masks[0], "");
} else if (src_components == 1 && num_components > 1) {
LLVMValueRef values[] = {value, value, value, value};
- value = build_gather_values(ctx, values,
num_components);
+ value = ac_build_gather_values(&ctx->ac, values,
num_components);
} else {
LLVMValueRef swizzle = LLVMConstVector(masks,
num_components);
value = LLVMBuildShuffleVector(ctx->builder, value,
value,
swizzle, "");
}
}
assert(!src.negate);
assert(!src.abs);
return value;
}
@@ -1224,21 +1183,21 @@ static LLVMValueRef emit_ddxy_interp(
{
LLVMValueRef result[4], a;
unsigned i;
for (i = 0; i < 2; i++) {
a = LLVMBuildExtractElement(ctx->builder, interp_ij,
LLVMConstInt(ctx->i32, i, false),
"");
result[i] = emit_ddxy(ctx, nir_op_fddx, a);
result[2+i] = emit_ddxy(ctx, nir_op_fddy, a);
}
- return build_gather_values(ctx, result, 4);
+ return ac_build_gather_values(&ctx->ac, result, 4);
}
static LLVMValueRef emit_fdiv(struct nir_to_llvm_context *ctx,
LLVMValueRef num,
LLVMValueRef den)
{
LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
if (!LLVMIsConstant(ret))
LLVMSetMetadata(ret, ctx->fpmath_md_kind,
ctx->fpmath_md_2p5_ulp);
@@ -1476,21 +1435,21 @@ static void visit_alu(struct nir_to_llvm_context *ctx,
nir_alu_instr *instr)
result = ac_emit_llvm_intrinsic(&ctx->ac,
"llvm.bitreverse.i32", ctx->i32, src, 1, AC_FUNC_ATTR_READNONE);
break;
case nir_op_bit_count:
result = ac_emit_llvm_intrinsic(&ctx->ac, "llvm.ctpop.i32",
ctx->i32, src, 1, AC_FUNC_ATTR_READNONE);
break;
case nir_op_vec2:
case nir_op_vec3:
case nir_op_vec4:
for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs;
i++)
src[i] = to_integer(ctx, src[i]);
- result = build_gather_values(ctx, src, num_components);
+ result = ac_build_gather_values(&ctx->ac, src, num_components);
break;
case nir_op_f2i:
src[0] = to_float(ctx, src[0]);
result = LLVMBuildFPToSI(ctx->builder, src[0], ctx->i32, "");
break;
case nir_op_f2u:
src[0] = to_float(ctx, src[0]);
result = LLVMBuildFPToUI(ctx->builder, src[0], ctx->i32, "");
break;
case nir_op_i2f:
@@ -1998,21 +1957,21 @@ static LLVMValueRef visit_load_ubo_buffer(struct
nir_to_llvm_context *ctx,
LLVMValueRef params[] = {
rsrc,
LLVMBuildAdd(ctx->builder, LLVMConstInt(ctx->i32, 4 *
i, 0),
offset, "")
};
results[i] = ac_emit_llvm_intrinsic(&ctx->ac,
"llvm.SI.load.const", ctx->f32,
params, 2,
AC_FUNC_ATTR_READNONE);
}
- ret = build_gather_values(ctx, results, instr->num_components);
+ ret = ac_build_gather_values(&ctx->ac, results, instr->num_components);
return LLVMBuildBitCast(ctx->builder, ret,
get_def_type(ctx, &instr->dest.ssa), "");
}
static void
radv_get_deref_offset(struct nir_to_llvm_context *ctx, nir_deref *tail,
bool vs_in, unsigned *const_out, LLVMValueRef *indir_out)
{
unsigned const_offset = 0;
LLVMValueRef offset = NULL;
@@ -2072,86 +2031,86 @@ static LLVMValueRef visit_load_var(struct
nir_to_llvm_context *ctx,
switch (instr->variables[0]->var->data.mode) {
case nir_var_shader_in:
radv_get_deref_offset(ctx, &instr->variables[0]->deref,
ctx->stage == MESA_SHADER_VERTEX,
&const_index, &indir_index);
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);
- LLVMValueRef tmp_vec =
build_gather_values_extended(
- ctx, ctx->inputs + idx + chan,
count,
+ LLVMValueRef tmp_vec =
ac_build_gather_values_extended(
+ &ctx->ac, ctx->inputs + idx +
chan, count,
4, false);
values[chan] =
LLVMBuildExtractElement(ctx->builder,
tmp_vec,
indir_index, "");
} else
values[chan] = ctx->inputs[idx + chan +
const_index * 4];
}
- return to_integer(ctx, build_gather_values(ctx, values, ve));
+ return to_integer(ctx, ac_build_gather_values(&ctx->ac, values,
ve));
break;
case nir_var_local:
radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
&const_index, &indir_index);
for (unsigned chan = 0; chan < ve; chan++) {
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type, false);
- LLVMValueRef tmp_vec =
build_gather_values_extended(
- ctx, ctx->locals + idx + chan,
count,
+ LLVMValueRef tmp_vec =
ac_build_gather_values_extended(
+ &ctx->ac, ctx->locals + idx +
chan, count,
4, true);
values[chan] =
LLVMBuildExtractElement(ctx->builder,
tmp_vec,
indir_index, "");
} else {
values[chan] = LLVMBuildLoad(ctx->builder,
ctx->locals[idx + chan + const_index * 4], "");
}
}
- return to_integer(ctx, build_gather_values(ctx, values, ve));
+ return to_integer(ctx, ac_build_gather_values(&ctx->ac, values,
ve));
case nir_var_shader_out:
radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
&const_index, &indir_index);
for (unsigned chan = 0; chan < ve; chan++) {
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type,
false);
- LLVMValueRef tmp_vec =
build_gather_values_extended(
- ctx, ctx->outputs + idx + chan,
count,
+ LLVMValueRef tmp_vec =
ac_build_gather_values_extended(
+ &ctx->ac, ctx->outputs + idx +
chan, count,
4, true);
values[chan] =
LLVMBuildExtractElement(ctx->builder,
tmp_vec,
indir_index, "");
} else {
values[chan] = LLVMBuildLoad(ctx->builder,
ctx->outputs[idx + chan +
const_index * 4],
"");
}
}
- return to_integer(ctx, build_gather_values(ctx, values, ve));
+ return to_integer(ctx, ac_build_gather_values(&ctx->ac, values,
ve));
case nir_var_shared: {
radv_get_deref_offset(ctx, &instr->variables[0]->deref, false,
&const_index, &indir_index);
LLVMValueRef ptr = get_shared_memory_ptr(ctx, idx, ctx->i32);
LLVMValueRef derived_ptr;
for (unsigned chan = 0; chan < ve; chan++) {
LLVMValueRef index = LLVMConstInt(ctx->i32, chan,
false);
if (indir_index)
index = LLVMBuildAdd(ctx->builder, index,
indir_index, "");
derived_ptr = LLVMBuildGEP(ctx->builder, ptr, &index,
1, "");
values[chan] = LLVMBuildLoad(ctx->builder, derived_ptr,
"");
}
- return to_integer(ctx, build_gather_values(ctx, values, ve));
+ return to_integer(ctx, ac_build_gather_values(&ctx->ac, values,
ve));
}
default:
break;
}
return NULL;
}
static void
visit_store_var(struct nir_to_llvm_context *ctx,
nir_intrinsic_instr *instr)
@@ -2177,22 +2136,22 @@ visit_store_var(struct nir_to_llvm_context *ctx,
LLVMConstInt(ctx->i32,
chan, false),
"");
if (instr->variables[0]->var->data.location ==
VARYING_SLOT_CLIP_DIST0 ||
instr->variables[0]->var->data.location ==
VARYING_SLOT_CULL_DIST0)
stride = 1;
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type,
false);
- LLVMValueRef tmp_vec =
build_gather_values_extended(
- ctx, ctx->outputs + idx + chan,
count,
+ LLVMValueRef tmp_vec =
ac_build_gather_values_extended(
+ &ctx->ac, ctx->outputs + idx +
chan, count,
stride, true);
if (get_llvm_num_components(tmp_vec) > 1) {
tmp_vec =
LLVMBuildInsertElement(ctx->builder, tmp_vec,
value,
indir_index, "");
} else
tmp_vec = value;
build_store_values_extended(ctx, ctx->outputs +
idx + chan,
count, stride,
tmp_vec);
@@ -2211,22 +2170,22 @@ visit_store_var(struct nir_to_llvm_context *ctx,
continue;
if (get_llvm_num_components(src) == 1)
value = src;
else
value = LLVMBuildExtractElement(ctx->builder,
src,
LLVMConstInt(ctx->i32, chan, false), "");
if (indir_index) {
unsigned count = glsl_count_attribute_slots(
instr->variables[0]->var->type, false);
- LLVMValueRef tmp_vec =
build_gather_values_extended(
- ctx, ctx->locals + idx + chan, count,
+ LLVMValueRef tmp_vec =
ac_build_gather_values_extended(
+ &ctx->ac, ctx->locals + idx + chan,
count,
4, true);
tmp_vec = LLVMBuildInsertElement(ctx->builder,
tmp_vec,
value,
indir_index, "");
build_store_values_extended(ctx, ctx->locals +
idx + chan,
count, 4, tmp_vec);
} else {
temp_ptr = ctx->locals[idx + chan + const_index
* 4];
LLVMBuildStore(ctx->builder, value, temp_ptr);
@@ -2320,21 +2279,21 @@ static LLVMValueRef get_image_coords(struct
nir_to_llvm_context *ctx,
}
if (add_frag_pos) {
for (chan = 0; chan < count; ++chan)
coords[chan] = LLVMBuildAdd(ctx->builder,
coords[chan], LLVMBuildFPToUI(ctx->builder, ctx->frag_pos[chan], ctx->i32, ""),
"");
}
if (count == 3) {
coords[3] = LLVMGetUndef(ctx->i32);
count = 4;
}
- res = build_gather_values(ctx, coords, count);
+ res = ac_build_gather_values(&ctx->ac, coords, count);
}
return res;
}
static void build_type_name_for_intr(
LLVMTypeRef type,
char *buf, unsigned bufsize)
{
LLVMTypeRef elem_type = type;
@@ -2763,30 +2722,30 @@ static LLVMValueRef load_sample_position(struct
nir_to_llvm_context *ctx,
LLVMValueRef sample_id)
{
/* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
LLVMValueRef offset0 = LLVMBuildMul(ctx->builder, sample_id,
LLVMConstInt(ctx->i32, 8, false), "");
LLVMValueRef offset1 = LLVMBuildAdd(ctx->builder, offset0,
LLVMConstInt(ctx->i32, 4, false), "");
LLVMValueRef result[2];
result[0] = build_indexed_load_const(ctx, ctx->sample_positions,
offset0);
result[1] = build_indexed_load_const(ctx, ctx->sample_positions,
offset1);
- return build_gather_values(ctx, result, 2);
+ return ac_build_gather_values(&ctx->ac, result, 2);
}
static LLVMValueRef load_sample_pos(struct nir_to_llvm_context *ctx)
{
LLVMValueRef values[2];
values[0] = emit_ffract(ctx, ctx->frag_pos[0]);
values[1] = emit_ffract(ctx, ctx->frag_pos[1]);
- return build_gather_values(ctx, values, 2);
+ return ac_build_gather_values(&ctx->ac, values, 2);
}
static LLVMValueRef visit_interp(struct nir_to_llvm_context *ctx,
nir_intrinsic_instr *instr)
{
LLVMValueRef result[2];
LLVMValueRef interp_param, attr_number;
unsigned location;
unsigned chan;
LLVMValueRef src_c0, src_c1;
@@ -2852,37 +2811,37 @@ static LLVMValueRef visit_interp(struct
nir_to_llvm_context *ctx,
temp1 = LLVMBuildFMul(ctx->builder, ddx_el, src_c0, "");
temp1 = LLVMBuildFAdd(ctx->builder, temp1, interp_el,
"");
temp2 = LLVMBuildFMul(ctx->builder, ddy_el, src_c1, "");
temp2 = LLVMBuildFAdd(ctx->builder, temp2, temp1, "");
ij_out[i] = LLVMBuildBitCast(ctx->builder,
temp2, ctx->i32, "");
}
- interp_param = build_gather_values(ctx, ij_out, 2);
+ interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
}
intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
for (chan = 0; chan < 2; chan++) {
LLVMValueRef args[4];
LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, false);
args[0] = llvm_chan;
args[1] = attr_number;
args[2] = ctx->prim_mask;
args[3] = interp_param;
result[chan] = ac_emit_llvm_intrinsic(&ctx->ac, intr_name,
ctx->f32, args, args[3] ? 4
: 3,
AC_FUNC_ATTR_READNONE);
}
- return build_gather_values(ctx, result, 2);
+ return ac_build_gather_values(&ctx->ac, result, 2);
}
static void visit_intrinsic(struct nir_to_llvm_context *ctx,
nir_intrinsic_instr *instr)
{
LLVMValueRef result = NULL;
switch (instr->intrinsic) {
case nir_intrinsic_load_work_group_id: {
result = ctx->workgroup_ids;
@@ -3100,21 +3059,21 @@ static void set_tex_fetch_args(struct
nir_to_llvm_context *ctx,
unsigned is_rect = 0;
bool da = instr->is_array || instr->sampler_dim ==
GLSL_SAMPLER_DIM_CUBE;
if (op == nir_texop_lod)
da = false;
/* Pad to power of two vector */
while (count < util_next_power_of_two(count))
param[count++] = LLVMGetUndef(ctx->i32);
if (count > 1)
- tinfo->args[0] = build_gather_values(ctx, param, count);
+ tinfo->args[0] = ac_build_gather_values(&ctx->ac, param, count);
else
tinfo->args[0] = param[0];
tinfo->args[1] = res_ptr;
num_args = 2;
if (op == nir_texop_txf ||
op == nir_texop_txf_ms ||
op == nir_texop_query_levels ||
op == nir_texop_texture_samples ||
--
2.7.4
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
