static void si_dump_shader_key(unsigned shader, struct si_shader_key *key,
FILE *f);
+static unsigned llvm_get_type_size(LLVMTypeRef type);
+
static void si_build_vs_prolog_function(struct si_shader_context *ctx,
union si_shader_part_key *key);
static void si_build_vs_epilog_function(struct si_shader_context *ctx,
if (rshift)
value = LLVMBuildLShr(gallivm->builder, value,
- lp_build_const_int32(gallivm, rshift), "");
+ LLVMConstInt(ctx->i32, rshift, 0), "");
if (rshift + bitwidth < 32) {
unsigned mask = (1 << bitwidth) - 1;
value = LLVMBuildAnd(gallivm->builder, value,
- lp_build_const_int32(gallivm, mask), "");
+ LLVMConstInt(ctx->i32, mask, 0), "");
}
return value;
get_tcs_in_patch_stride(struct si_shader_context *ctx)
{
if (ctx->type == PIPE_SHADER_VERTEX)
- return unpack_param(ctx, SI_PARAM_LS_OUT_LAYOUT, 0, 13);
+ return unpack_param(ctx, SI_PARAM_LS_OUT_LAYOUT, 8, 13);
else if (ctx->type == PIPE_SHADER_TESS_CTRL)
- return unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 0, 13);
+ return unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 8, 13);
else {
assert(0);
return NULL;
}
static LLVMValueRef get_instance_index_for_fetch(
- struct si_shader_context *radeon_bld,
+ struct si_shader_context *ctx,
unsigned param_start_instance, unsigned divisor)
{
- struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->bld_base);
- struct gallivm_state *gallivm = radeon_bld->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
- LLVMValueRef result = LLVMGetParam(radeon_bld->main_fn,
+ LLVMValueRef result = LLVMGetParam(ctx->main_fn,
ctx->param_instance_id);
/* The division must be done before START_INSTANCE is added. */
if (divisor > 1)
result = LLVMBuildUDiv(gallivm->builder, result,
- lp_build_const_int32(gallivm, divisor), "");
+ LLVMConstInt(ctx->i32, divisor, 0), "");
return LLVMBuildAdd(gallivm->builder, result,
- LLVMGetParam(radeon_bld->main_fn, param_start_instance), "");
+ LLVMGetParam(ctx->main_fn, param_start_instance), "");
}
/* Bitcast <4 x float> to <2 x double>, extract the component, and convert
const struct tgsi_full_declaration *decl,
LLVMValueRef out[4])
{
- struct lp_build_context *base = &ctx->bld_base.base;
- struct gallivm_state *gallivm = base->gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
unsigned chan;
unsigned fix_fetch;
LLVMValueRef t_offset;
LLVMValueRef t_list;
LLVMValueRef vertex_index;
- LLVMValueRef args[3];
LLVMValueRef input[3];
/* Load the T list */
t_list_ptr = LLVMGetParam(ctx->main_fn, SI_PARAM_VERTEX_BUFFERS);
- t_offset = lp_build_const_int32(gallivm, input_index);
+ t_offset = LLVMConstInt(ctx->i32, input_index, 0);
t_list = ac_build_indexed_load_const(&ctx->ac, t_list_ptr, t_offset);
fetch_stride = 0;
}
- args[0] = t_list;
- args[2] = vertex_index;
-
for (unsigned i = 0; i < num_fetches; i++) {
- args[1] = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
+ LLVMValueRef voffset = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
- input[i] = lp_build_intrinsic(gallivm->builder,
- "llvm.SI.vs.load.input", ctx->v4f32, args, 3,
- LP_FUNC_ATTR_READNONE |
- LP_FUNC_ATTR_LEGACY);
+ input[i] = ac_build_buffer_load_format(&ctx->ac, t_list,
+ vertex_index, voffset,
+ true);
}
/* Break up the vec4 into individual components */
for (chan = 0; chan < 4; chan++) {
- LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
+ LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, 0);
out[chan] = LLVMBuildExtractElement(gallivm->builder,
input[0], llvm_chan, "");
}
struct si_shader_context *ctx = si_shader_context(bld_base);
if (swizzle > 0)
- return bld_base->uint_bld.zero;
+ return ctx->i32_0;
switch (ctx->type) {
case PIPE_SHADER_VERTEX:
SI_PARAM_PRIMITIVE_ID);
default:
assert(0);
- return bld_base->uint_bld.zero;
+ return ctx->i32_0;
}
}
const struct tgsi_ind_register *ind,
int rel_index)
{
- struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef result;
result = ctx->addrs[ind->Index][ind->Swizzle];
result = LLVMBuildLoad(gallivm->builder, result, "");
result = LLVMBuildAdd(gallivm->builder, result,
- lp_build_const_int32(gallivm, rel_index), "");
+ LLVMConstInt(ctx->i32, rel_index, 0), "");
return result;
}
* - SI & CIK hang
* - VI crashes
*/
- if (HAVE_LLVM <= 0x0308)
+ if (HAVE_LLVM == 0x0308)
return LLVMGetUndef(ctx->i32);
return si_llvm_bound_index(ctx, result, num);
LLVMValueRef vertex_dw_stride,
LLVMValueRef base_addr)
{
- struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct tgsi_shader_info *info = &ctx->shader->selector->info;
ubyte *name, *index, *array_first;
int first, param;
index = get_indirect_index(ctx, ®.DimIndirect,
reg.Dimension.Index);
else
- index = lp_build_const_int32(gallivm, reg.Dimension.Index);
+ index = LLVMConstInt(ctx->i32, reg.Dimension.Index, 0);
base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
LLVMBuildMul(gallivm->builder, index,
base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
LLVMBuildMul(gallivm->builder, ind_index,
- lp_build_const_int32(gallivm, 4), ""), "");
+ LLVMConstInt(ctx->i32, 4, 0), ""), "");
param = si_shader_io_get_unique_index(name[first], index[first]);
} else {
/* Add the base address of the element. */
return LLVMBuildAdd(gallivm->builder, base_addr,
- lp_build_const_int32(gallivm, param * 4), "");
+ LLVMConstInt(ctx->i32, param * 4, 0), "");
}
/* The offchip buffer layout for TCS->TES is
LLVMValueRef vertex_index,
LLVMValueRef param_index)
{
- struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef base_addr, vertices_per_patch, num_patches, total_vertices;
LLVMValueRef param_stride, constant16;
total_vertices = LLVMBuildMul(gallivm->builder, vertices_per_patch,
num_patches, "");
- constant16 = lp_build_const_int32(gallivm, 16);
+ constant16 = LLVMConstInt(ctx->i32, 16, 0);
if (vertex_index) {
base_addr = LLVMBuildMul(gallivm->builder, rel_patch_id,
vertices_per_patch, "");
const struct tgsi_full_dst_register *dst,
const struct tgsi_full_src_register *src)
{
- struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct tgsi_shader_info *info = &ctx->shader->selector->info;
ubyte *name, *index, *array_first;
struct tgsi_full_src_register reg;
vertex_index = get_indirect_index(ctx, ®.DimIndirect,
reg.Dimension.Index);
else
- vertex_index = lp_build_const_int32(gallivm,
- reg.Dimension.Index);
+ vertex_index = LLVMConstInt(ctx->i32, reg.Dimension.Index, 0);
}
/* Get information about the register. */
} else {
param_base = reg.Register.Index;
- param_index = lp_build_const_int32(gallivm, 0);
+ param_index = ctx->i32_0;
}
param_index_base = si_shader_io_get_unique_index(name[param_base],
index[param_base]);
param_index = LLVMBuildAdd(gallivm->builder, param_index,
- lp_build_const_int32(gallivm, param_index_base),
+ LLVMConstInt(ctx->i32, param_index_base, 0),
"");
return get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx),
LLVMValueRef base, bool readonly_memory)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef value, value2;
LLVMTypeRef llvm_type = tgsi2llvmtype(bld_base, type);
LLVMTypeRef vec_type = LLVMVectorType(llvm_type, 4);
value = LLVMBuildBitCast(gallivm->builder, value, vec_type, "");
return LLVMBuildExtractElement(gallivm->builder, value,
- lp_build_const_int32(gallivm, swizzle), "");
+ LLVMConstInt(ctx->i32, swizzle, 0), "");
}
value = ac_build_buffer_load(&ctx->ac, buffer, 1, NULL, base, offset,
LLVMValueRef dw_addr)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef value;
if (swizzle == ~0) {
for (unsigned chan = 0; chan < TGSI_NUM_CHANNELS; chan++)
values[chan] = lds_load(bld_base, type, chan, dw_addr);
- return lp_build_gather_values(bld_base->base.gallivm, values,
+ return lp_build_gather_values(gallivm, values,
TGSI_NUM_CHANNELS);
}
dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
- lp_build_const_int32(gallivm, swizzle));
+ LLVMConstInt(ctx->i32, swizzle, 0));
value = ac_build_indexed_load(&ctx->ac, ctx->lds, dw_addr, false);
if (tgsi_type_is_64bit(type)) {
LLVMValueRef value2;
dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
- lp_build_const_int32(gallivm, 1));
+ ctx->i32_1);
value2 = ac_build_indexed_load(&ctx->ac, ctx->lds, dw_addr, false);
return si_llvm_emit_fetch_64bit(bld_base, type, value, value2);
}
LLVMValueRef value)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
- lp_build_const_int32(gallivm, swizzle));
+ LLVMConstInt(ctx->i32, swizzle, 0));
value = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, "");
ac_build_indexed_store(&ctx->ac, ctx->lds,
struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef dw_addr, stride;
- stride = unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 13, 8);
+ stride = unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 24, 8);
dw_addr = get_tcs_in_current_patch_offset(ctx);
dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr);
enum tgsi_opcode_type type, unsigned swizzle)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMValueRef rw_buffers, buffer, base, addr;
rw_buffers = LLVMGetParam(ctx->main_fn,
SI_PARAM_RW_BUFFERS);
buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
- lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
+ LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
base = LLVMGetParam(ctx->main_fn, ctx->param_oc_lds);
addr = get_tcs_tes_buffer_address_from_reg(ctx, NULL, reg);
LLVMValueRef dst[4])
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
const struct tgsi_full_dst_register *reg = &inst->Dst[0];
const struct tgsi_shader_info *sh_info = &ctx->shader->selector->info;
unsigned chan_index;
rw_buffers = LLVMGetParam(ctx->main_fn,
SI_PARAM_RW_BUFFERS);
buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
- lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
+ LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
base = LLVMGetParam(ctx->main_fn, ctx->param_oc_lds);
buf_addr = get_tcs_tes_buffer_address_from_reg(ctx, reg, NULL);
LLVMValueRef value = dst[chan_index];
if (inst->Instruction.Saturate)
- value = ac_emit_clamp(&ctx->ac, value);
+ value = ac_build_clamp(&ctx->ac, value);
/* Skip LDS stores if there is no LDS read of this output. */
if (!skip_lds_store)
if (inst->Dst[0].Register.WriteMask != 0xF && !is_tess_factor) {
ac_build_buffer_store_dword(&ctx->ac, buffer, value, 1,
buf_addr, base,
- 4 * chan_index, 1, 0);
+ 4 * chan_index, 1, 0, true, false);
}
}
if (inst->Dst[0].Register.WriteMask == 0xF && !is_tess_factor) {
- LLVMValueRef value = lp_build_gather_values(bld_base->base.gallivm,
+ LLVMValueRef value = lp_build_gather_values(gallivm,
values, 4);
ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buf_addr,
- base, 0, 1, 0);
+ base, 0, 1, 0, true, false);
}
}
enum tgsi_opcode_type type,
unsigned swizzle)
{
- struct lp_build_context *base = &bld_base->base;
struct si_shader_context *ctx = si_shader_context(bld_base);
struct si_shader *shader = ctx->shader;
struct lp_build_context *uint = &ctx->bld_base.uint_bld;
- struct gallivm_state *gallivm = base->gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef vtx_offset, soffset;
unsigned vtx_offset_param;
struct tgsi_shader_info *info = &shader->selector->info;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
values[chan] = fetch_input_gs(bld_base, reg, type, chan);
}
- return lp_build_gather_values(bld_base->base.gallivm, values,
+ return lp_build_gather_values(gallivm, values,
TGSI_NUM_CHANNELS);
}
param = si_shader_io_get_unique_index(semantic_name, semantic_index);
soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle) * 256, 0);
- value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, uint->zero,
+ value = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1, ctx->i32_0,
vtx_offset, soffset, 0, 1, 0, true);
if (tgsi_type_is_64bit(type)) {
LLVMValueRef value2;
soffset = LLVMConstInt(ctx->i32, (param * 4 + swizzle + 1) * 256, 0);
value2 = ac_build_buffer_load(&ctx->ac, ctx->esgs_ring, 1,
- uint->zero, vtx_offset, soffset,
+ ctx->i32_0, vtx_offset, soffset,
0, 1, 0, true);
return si_llvm_emit_fetch_64bit(bld_base, type,
value, value2);
LLVMValueRef face,
LLVMValueRef result[4])
{
- struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
- struct lp_build_context *base = &bld_base->base;
- struct lp_build_context *uint = &bld_base->uint_bld;
- struct gallivm_state *gallivm = base->gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef attr_number;
LLVMValueRef i, j;
*/
bool interp = interp_param != NULL;
- attr_number = lp_build_const_int32(gallivm, input_index);
+ attr_number = LLVMConstInt(ctx->i32, input_index, 0);
if (interp) {
interp_param = LLVMBuildBitCast(gallivm->builder, interp_param,
LLVMVectorType(ctx->f32, 2), "");
i = LLVMBuildExtractElement(gallivm->builder, interp_param,
- uint->zero, "");
+ ctx->i32_0, "");
j = LLVMBuildExtractElement(gallivm->builder, interp_param,
- uint->one, "");
+ ctx->i32_1, "");
}
if (semantic_name == TGSI_SEMANTIC_COLOR &&
if (semantic_index == 1 && colors_read_mask & 0xf)
back_attr_offset += 1;
- back_attr_number = lp_build_const_int32(gallivm, back_attr_offset);
+ back_attr_number = LLVMConstInt(ctx->i32, back_attr_offset, 0);
is_face_positive = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
- face, uint->zero, "");
+ face, ctx->i32_0, "");
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
+ LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, 0);
LLVMValueRef front, back;
if (interp) {
i, j);
} else {
front = ac_build_fs_interp_mov(&ctx->ac,
- lp_build_const_int32(gallivm, 2), /* P0 */
+ LLVMConstInt(ctx->i32, 2, 0), /* P0 */
llvm_chan, attr_number, prim_mask);
back = ac_build_fs_interp_mov(&ctx->ac,
- lp_build_const_int32(gallivm, 2), /* P0 */
+ LLVMConstInt(ctx->i32, 2, 0), /* P0 */
llvm_chan, back_attr_number, prim_mask);
}
}
} else if (semantic_name == TGSI_SEMANTIC_FOG) {
if (interp) {
- result[0] = ac_build_fs_interp(&ctx->ac, uint->zero,
+ result[0] = ac_build_fs_interp(&ctx->ac, ctx->i32_0,
attr_number, prim_mask, i, j);
} else {
- result[0] = ac_build_fs_interp_mov(&ctx->ac, uint->zero,
- lp_build_const_int32(gallivm, 2), /* P0 */
+ result[0] = ac_build_fs_interp_mov(&ctx->ac, ctx->i32_0,
+ LLVMConstInt(ctx->i32, 2, 0), /* P0 */
attr_number, prim_mask);
}
result[1] =
- result[2] = lp_build_const_float(gallivm, 0.0f);
- result[3] = lp_build_const_float(gallivm, 1.0f);
+ result[2] = LLVMConstReal(ctx->f32, 0.0f);
+ result[3] = LLVMConstReal(ctx->f32, 1.0f);
} else {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
+ LLVMValueRef llvm_chan = LLVMConstInt(ctx->i32, chan, 0);
if (interp) {
result[chan] = ac_build_fs_interp(&ctx->ac,
llvm_chan, attr_number, prim_mask, i, j);
} else {
result[chan] = ac_build_fs_interp_mov(&ctx->ac,
- lp_build_const_int32(gallivm, 2), /* P0 */
+ LLVMConstInt(ctx->i32, 2, 0), /* P0 */
llvm_chan, attr_number, prim_mask);
}
}
}
static void declare_input_fs(
- struct si_shader_context *radeon_bld,
+ struct si_shader_context *ctx,
unsigned input_index,
const struct tgsi_full_declaration *decl,
LLVMValueRef out[4])
{
- struct lp_build_context *base = &radeon_bld->bld_base.base;
- struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->bld_base);
+ struct lp_build_context *base = &ctx->bld_base.base;
struct si_shader *shader = ctx->shader;
- LLVMValueRef main_fn = radeon_bld->main_fn;
+ LLVMValueRef main_fn = ctx->main_fn;
LLVMValueRef interp_param = NULL;
int interp_param_idx;
&out[0]);
}
-static LLVMValueRef get_sample_id(struct si_shader_context *radeon_bld)
+static LLVMValueRef get_sample_id(struct si_shader_context *ctx)
{
- return unpack_param(si_shader_context(&radeon_bld->bld_base),
- SI_PARAM_ANCILLARY, 8, 4);
+ return unpack_param(ctx, SI_PARAM_ANCILLARY, 8, 4);
}
LP_FUNC_ATTR_LEGACY);
}
-static LLVMValueRef load_sample_position(struct si_shader_context *radeon_bld, LLVMValueRef sample_id)
+static LLVMValueRef load_sample_position(struct si_shader_context *ctx, LLVMValueRef sample_id)
{
- struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->bld_base);
- struct lp_build_context *uint_bld = &radeon_bld->bld_base.uint_bld;
- struct gallivm_state *gallivm = &radeon_bld->gallivm;
+ struct lp_build_context *uint_bld = &ctx->bld_base.uint_bld;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef desc = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
- LLVMValueRef buf_index = lp_build_const_int32(gallivm, SI_PS_CONST_SAMPLE_POSITIONS);
+ LLVMValueRef buf_index = LLVMConstInt(ctx->i32, SI_PS_CONST_SAMPLE_POSITIONS, 0);
LLVMValueRef resource = ac_build_indexed_load_const(&ctx->ac, desc, buf_index);
/* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
LLVMValueRef offset0 = lp_build_mul_imm(uint_bld, sample_id, 8);
- LLVMValueRef offset1 = LLVMBuildAdd(builder, offset0, lp_build_const_int32(gallivm, 4), "");
+ LLVMValueRef offset1 = LLVMBuildAdd(builder, offset0, LLVMConstInt(ctx->i32, 4, 0), "");
LLVMValueRef pos[4] = {
buffer_load_const(ctx, resource, offset0),
buffer_load_const(ctx, resource, offset1),
- lp_build_const_float(gallivm, 0),
- lp_build_const_float(gallivm, 0)
+ LLVMConstReal(ctx->f32, 0),
+ LLVMConstReal(ctx->f32, 0)
};
return lp_build_gather_values(gallivm, pos, 4);
}
-static void declare_system_value(
- struct si_shader_context *radeon_bld,
- unsigned index,
- const struct tgsi_full_declaration *decl)
+static void declare_system_value(struct si_shader_context *ctx,
+ unsigned index,
+ const struct tgsi_full_declaration *decl)
{
- struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->bld_base);
- struct lp_build_context *bld = &radeon_bld->bld_base.base;
- struct gallivm_state *gallivm = &radeon_bld->gallivm;
+ struct lp_build_context *bld = &ctx->bld_base.base;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef value = 0;
+ assert(index < RADEON_LLVM_MAX_SYSTEM_VALUES);
+
switch (decl->Semantic.Name) {
case TGSI_SEMANTIC_INSTANCEID:
- value = LLVMGetParam(radeon_bld->main_fn,
+ value = LLVMGetParam(ctx->main_fn,
ctx->param_instance_id);
break;
case TGSI_SEMANTIC_VERTEXID:
value = LLVMBuildAdd(gallivm->builder,
- LLVMGetParam(radeon_bld->main_fn,
+ LLVMGetParam(ctx->main_fn,
ctx->param_vertex_id),
- LLVMGetParam(radeon_bld->main_fn,
+ LLVMGetParam(ctx->main_fn,
SI_PARAM_BASE_VERTEX), "");
break;
case TGSI_SEMANTIC_VERTEXID_NOBASE:
- value = LLVMGetParam(radeon_bld->main_fn,
- ctx->param_vertex_id);
+ /* Unused. Clarify the meaning in indexed vs. non-indexed
+ * draws if this is ever used again. */
+ assert(false);
break;
case TGSI_SEMANTIC_BASEVERTEX:
- value = LLVMGetParam(radeon_bld->main_fn,
+ value = LLVMGetParam(ctx->main_fn,
SI_PARAM_BASE_VERTEX);
break;
case TGSI_SEMANTIC_BASEINSTANCE:
- value = LLVMGetParam(radeon_bld->main_fn,
+ value = LLVMGetParam(ctx->main_fn,
SI_PARAM_START_INSTANCE);
break;
case TGSI_SEMANTIC_DRAWID:
- value = LLVMGetParam(radeon_bld->main_fn,
+ value = LLVMGetParam(ctx->main_fn,
SI_PARAM_DRAWID);
break;
if (ctx->type == PIPE_SHADER_TESS_CTRL)
value = unpack_param(ctx, SI_PARAM_REL_IDS, 8, 5);
else if (ctx->type == PIPE_SHADER_GEOMETRY)
- value = LLVMGetParam(radeon_bld->main_fn,
+ value = LLVMGetParam(ctx->main_fn,
SI_PARAM_GS_INSTANCE_ID);
else
assert(!"INVOCATIONID not implemented");
case TGSI_SEMANTIC_POSITION:
{
LLVMValueRef pos[4] = {
- LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_X_FLOAT),
- LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_Y_FLOAT),
- LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_Z_FLOAT),
- lp_build_emit_llvm_unary(&radeon_bld->bld_base, TGSI_OPCODE_RCP,
- LLVMGetParam(radeon_bld->main_fn,
+ LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT),
+ LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT),
+ LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Z_FLOAT),
+ lp_build_emit_llvm_unary(&ctx->bld_base, TGSI_OPCODE_RCP,
+ LLVMGetParam(ctx->main_fn,
SI_PARAM_POS_W_FLOAT)),
};
value = lp_build_gather_values(gallivm, pos, 4);
}
case TGSI_SEMANTIC_FACE:
- value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_FRONT_FACE);
+ value = LLVMGetParam(ctx->main_fn, SI_PARAM_FRONT_FACE);
break;
case TGSI_SEMANTIC_SAMPLEID:
- value = get_sample_id(radeon_bld);
+ value = get_sample_id(ctx);
break;
case TGSI_SEMANTIC_SAMPLEPOS: {
LLVMValueRef pos[4] = {
- LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_X_FLOAT),
- LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_Y_FLOAT),
- lp_build_const_float(gallivm, 0),
- lp_build_const_float(gallivm, 0)
+ LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT),
+ LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT),
+ LLVMConstReal(ctx->f32, 0),
+ LLVMConstReal(ctx->f32, 0)
};
- pos[0] = lp_build_emit_llvm_unary(&radeon_bld->bld_base,
+ pos[0] = lp_build_emit_llvm_unary(&ctx->bld_base,
TGSI_OPCODE_FRC, pos[0]);
- pos[1] = lp_build_emit_llvm_unary(&radeon_bld->bld_base,
+ pos[1] = lp_build_emit_llvm_unary(&ctx->bld_base,
TGSI_OPCODE_FRC, pos[1]);
value = lp_build_gather_values(gallivm, pos, 4);
break;
/* This can only occur with the OpenGL Core profile, which
* doesn't support smoothing.
*/
- value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_SAMPLE_COVERAGE);
+ value = LLVMGetParam(ctx->main_fn, SI_PARAM_SAMPLE_COVERAGE);
break;
case TGSI_SEMANTIC_TESSCOORD:
{
LLVMValueRef coord[4] = {
- LLVMGetParam(radeon_bld->main_fn, ctx->param_tes_u),
- LLVMGetParam(radeon_bld->main_fn, ctx->param_tes_v),
+ LLVMGetParam(ctx->main_fn, ctx->param_tes_u),
+ LLVMGetParam(ctx->main_fn, ctx->param_tes_v),
bld->zero,
bld->zero
};
rw_buffers = LLVMGetParam(ctx->main_fn,
SI_PARAM_RW_BUFFERS);
buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
- lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
+ LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
base = LLVMGetParam(ctx->main_fn, ctx->param_oc_lds);
addr = get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx), NULL,
- lp_build_const_int32(gallivm, param));
+ LLVMConstInt(ctx->i32, param, 0));
- value = buffer_load(&radeon_bld->bld_base, TGSI_TYPE_FLOAT,
+ value = buffer_load(&ctx->bld_base, TGSI_TYPE_FLOAT,
~0, buffer, base, addr, true);
break;
LLVMValueRef buf, slot, val[4];
int i, offset;
- slot = lp_build_const_int32(gallivm, SI_HS_CONST_DEFAULT_TESS_LEVELS);
+ slot = LLVMConstInt(ctx->i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
buf = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
buf = ac_build_indexed_load_const(&ctx->ac, buf, slot);
offset = decl->Semantic.Name == TGSI_SEMANTIC_DEFAULT_TESSINNER_SI ? 4 : 0;
for (i = 0; i < 4; i++)
val[i] = buffer_load_const(ctx, buf,
- lp_build_const_int32(gallivm, (offset + i) * 4));
+ LLVMConstInt(ctx->i32, (offset + i) * 4, 0));
value = lp_build_gather_values(gallivm, val, 4);
break;
}
case TGSI_SEMANTIC_PRIMID:
- value = get_primitive_id(&radeon_bld->bld_base, 0);
+ value = get_primitive_id(&ctx->bld_base, 0);
break;
case TGSI_SEMANTIC_GRID_SIZE:
- value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_GRID_SIZE);
+ value = LLVMGetParam(ctx->main_fn, SI_PARAM_GRID_SIZE);
break;
case TGSI_SEMANTIC_BLOCK_SIZE:
};
for (i = 0; i < 3; ++i)
- values[i] = lp_build_const_int32(gallivm, sizes[i]);
+ values[i] = LLVMConstInt(ctx->i32, sizes[i], 0);
value = lp_build_gather_values(gallivm, values, 3);
} else {
- value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_BLOCK_SIZE);
+ value = LLVMGetParam(ctx->main_fn, SI_PARAM_BLOCK_SIZE);
}
break;
}
case TGSI_SEMANTIC_BLOCK_ID:
- value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_BLOCK_ID);
+ value = LLVMGetParam(ctx->main_fn, SI_PARAM_BLOCK_ID);
break;
case TGSI_SEMANTIC_THREAD_ID:
- value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_THREAD_ID);
+ value = LLVMGetParam(ctx->main_fn, SI_PARAM_THREAD_ID);
break;
case TGSI_SEMANTIC_HELPER_INVOCATION:
}
break;
+ case TGSI_SEMANTIC_SUBGROUP_SIZE:
+ value = LLVMConstInt(ctx->i32, 64, 0);
+ break;
+
+ case TGSI_SEMANTIC_SUBGROUP_INVOCATION:
+ value = ac_get_thread_id(&ctx->ac);
+ break;
+
+ case TGSI_SEMANTIC_SUBGROUP_EQ_MASK:
+ {
+ LLVMValueRef id = ac_get_thread_id(&ctx->ac);
+ id = LLVMBuildZExt(gallivm->builder, id, ctx->i64, "");
+ value = LLVMBuildShl(gallivm->builder, LLVMConstInt(ctx->i64, 1, 0), id, "");
+ value = LLVMBuildBitCast(gallivm->builder, value, ctx->v2i32, "");
+ break;
+ }
+
+ case TGSI_SEMANTIC_SUBGROUP_GE_MASK:
+ case TGSI_SEMANTIC_SUBGROUP_GT_MASK:
+ case TGSI_SEMANTIC_SUBGROUP_LE_MASK:
+ case TGSI_SEMANTIC_SUBGROUP_LT_MASK:
+ {
+ LLVMValueRef id = ac_get_thread_id(&ctx->ac);
+ if (decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_GT_MASK ||
+ decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LE_MASK) {
+ /* All bits set except LSB */
+ value = LLVMConstInt(ctx->i64, -2, 0);
+ } else {
+ /* All bits set */
+ value = LLVMConstInt(ctx->i64, -1, 0);
+ }
+ id = LLVMBuildZExt(gallivm->builder, id, ctx->i64, "");
+ value = LLVMBuildShl(gallivm->builder, value, id, "");
+ if (decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LE_MASK ||
+ decl->Semantic.Name == TGSI_SEMANTIC_SUBGROUP_LT_MASK)
+ value = LLVMBuildNot(gallivm->builder, value, "");
+ value = LLVMBuildBitCast(gallivm->builder, value, ctx->v2i32, "");
+ break;
+ }
+
default:
assert(!"unknown system value");
return;
}
- radeon_bld->system_values[index] = value;
+ ctx->system_values[index] = value;
}
-static void declare_compute_memory(struct si_shader_context *radeon_bld,
+static void declare_compute_memory(struct si_shader_context *ctx,
const struct tgsi_full_declaration *decl)
{
- struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->bld_base);
struct si_shader_selector *sel = ctx->shader->selector;
- struct gallivm_state *gallivm = &radeon_bld->gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMTypeRef i8p = LLVMPointerType(ctx->i8, LOCAL_ADDR_SPACE);
LLVMValueRef var;
for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan)
values[chan] = fetch_constant(bld_base, reg, type, chan);
- return lp_build_gather_values(bld_base->base.gallivm, values, 4);
+ return lp_build_gather_values(&ctx->gallivm, values, 4);
}
buf = reg->Register.Dimension ? reg->Dimension.Index : 0;
addr = LLVMBuildLoad(base->gallivm->builder, addr, "load addr reg");
addr = lp_build_mul_imm(&bld_base->uint_bld, addr, 16);
addr = lp_build_add(&bld_base->uint_bld, addr,
- lp_build_const_int32(base->gallivm, idx * 4));
+ LLVMConstInt(ctx->i32, idx * 4, 0));
} else {
addr = LLVMConstInt(ctx->i32, idx * 4, 0);
}
}
/* Upper 16 bits must be zero. */
-static LLVMValueRef si_llvm_pack_two_int16(struct gallivm_state *gallivm,
+static LLVMValueRef si_llvm_pack_two_int16(struct si_shader_context *ctx,
LLVMValueRef val[2])
{
- return LLVMBuildOr(gallivm->builder, val[0],
- LLVMBuildShl(gallivm->builder, val[1],
- lp_build_const_int32(gallivm, 16),
+ return LLVMBuildOr(ctx->gallivm.builder, val[0],
+ LLVMBuildShl(ctx->gallivm.builder, val[1],
+ LLVMConstInt(ctx->i32, 16, 0),
""), "");
}
/* Upper 16 bits are ignored and will be dropped. */
-static LLVMValueRef si_llvm_pack_two_int32_as_int16(struct gallivm_state *gallivm,
+static LLVMValueRef si_llvm_pack_two_int32_as_int16(struct si_shader_context *ctx,
LLVMValueRef val[2])
{
LLVMValueRef v[2] = {
- LLVMBuildAnd(gallivm->builder, val[0],
- lp_build_const_int32(gallivm, 0xffff), ""),
+ LLVMBuildAnd(ctx->gallivm.builder, val[0],
+ LLVMConstInt(ctx->i32, 0xffff, 0), ""),
val[1],
};
- return si_llvm_pack_two_int16(gallivm, v);
+ return si_llvm_pack_two_int16(ctx, v);
}
/* Initialize arguments for the shader export intrinsic */
{
struct si_shader_context *ctx = si_shader_context(bld_base);
struct lp_build_context *base = &bld_base->base;
- struct gallivm_state *gallivm = base->gallivm;
- LLVMBuilderRef builder = base->gallivm->builder;
+ LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef val[4];
unsigned spi_shader_col_format = V_028714_SPI_SHADER_32_ABGR;
unsigned chan;
};
LLVMValueRef packed;
- packed = ac_emit_cvt_pkrtz_f16(&ctx->ac, pack_args);
+ packed = ac_build_cvt_pkrtz_f16(&ctx->ac, pack_args);
args->out[chan] =
- LLVMBuildBitCast(base->gallivm->builder,
+ LLVMBuildBitCast(ctx->gallivm.builder,
packed, ctx->f32, "");
}
break;
case V_028714_SPI_SHADER_UNORM16_ABGR:
for (chan = 0; chan < 4; chan++) {
- val[chan] = ac_emit_clamp(&ctx->ac, values[chan]);
+ val[chan] = ac_build_clamp(&ctx->ac, values[chan]);
val[chan] = LLVMBuildFMul(builder, val[chan],
- lp_build_const_float(gallivm, 65535), "");
+ LLVMConstReal(ctx->f32, 65535), "");
val[chan] = LLVMBuildFAdd(builder, val[chan],
- lp_build_const_float(gallivm, 0.5), "");
+ LLVMConstReal(ctx->f32, 0.5), "");
val[chan] = LLVMBuildFPToUI(builder, val[chan],
ctx->i32, "");
}
args->compr = 1; /* COMPR flag */
args->out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- si_llvm_pack_two_int16(gallivm, val));
+ si_llvm_pack_two_int16(ctx, val));
args->out[1] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- si_llvm_pack_two_int16(gallivm, val+2));
+ si_llvm_pack_two_int16(ctx, val+2));
break;
case V_028714_SPI_SHADER_SNORM16_ABGR:
/* Clamp between [-1, 1]. */
val[chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_MIN,
values[chan],
- lp_build_const_float(gallivm, 1));
+ LLVMConstReal(ctx->f32, 1));
val[chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_MAX,
val[chan],
- lp_build_const_float(gallivm, -1));
+ LLVMConstReal(ctx->f32, -1));
/* Convert to a signed integer in [-32767, 32767]. */
val[chan] = LLVMBuildFMul(builder, val[chan],
- lp_build_const_float(gallivm, 32767), "");
+ LLVMConstReal(ctx->f32, 32767), "");
/* If positive, add 0.5, else add -0.5. */
val[chan] = LLVMBuildFAdd(builder, val[chan],
LLVMBuildSelect(builder,
LLVMBuildFCmp(builder, LLVMRealOGE,
val[chan], base->zero, ""),
- lp_build_const_float(gallivm, 0.5),
- lp_build_const_float(gallivm, -0.5), ""), "");
+ LLVMConstReal(ctx->f32, 0.5),
+ LLVMConstReal(ctx->f32, -0.5), ""), "");
val[chan] = LLVMBuildFPToSI(builder, val[chan], ctx->i32, "");
}
args->compr = 1; /* COMPR flag */
args->out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- si_llvm_pack_two_int32_as_int16(gallivm, val));
+ si_llvm_pack_two_int32_as_int16(ctx, val));
args->out[1] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- si_llvm_pack_two_int32_as_int16(gallivm, val+2));
+ si_llvm_pack_two_int32_as_int16(ctx, val+2));
break;
case V_028714_SPI_SHADER_UINT16_ABGR: {
- LLVMValueRef max_rgb = lp_build_const_int32(gallivm,
- is_int8 ? 255 : is_int10 ? 1023 : 65535);
+ LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
+ is_int8 ? 255 : is_int10 ? 1023 : 65535, 0);
LLVMValueRef max_alpha =
- !is_int10 ? max_rgb : lp_build_const_int32(gallivm, 3);
+ !is_int10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
/* Clamp. */
for (chan = 0; chan < 4; chan++) {
args->compr = 1; /* COMPR flag */
args->out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- si_llvm_pack_two_int16(gallivm, val));
+ si_llvm_pack_two_int16(ctx, val));
args->out[1] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- si_llvm_pack_two_int16(gallivm, val+2));
+ si_llvm_pack_two_int16(ctx, val+2));
break;
}
case V_028714_SPI_SHADER_SINT16_ABGR: {
- LLVMValueRef max_rgb = lp_build_const_int32(gallivm,
- is_int8 ? 127 : is_int10 ? 511 : 32767);
- LLVMValueRef min_rgb = lp_build_const_int32(gallivm,
- is_int8 ? -128 : is_int10 ? -512 : -32768);
+ LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
+ is_int8 ? 127 : is_int10 ? 511 : 32767, 0);
+ LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
+ is_int8 ? -128 : is_int10 ? -512 : -32768, 0);
LLVMValueRef max_alpha =
- !is_int10 ? max_rgb : lp_build_const_int32(gallivm, 1);
+ !is_int10 ? max_rgb : ctx->i32_1;
LLVMValueRef min_alpha =
- !is_int10 ? min_rgb : lp_build_const_int32(gallivm, -2);
+ !is_int10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
/* Clamp. */
for (chan = 0; chan < 4; chan++) {
args->compr = 1; /* COMPR flag */
args->out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- si_llvm_pack_two_int32_as_int16(gallivm, val));
+ si_llvm_pack_two_int32_as_int16(ctx, val));
args->out[1] = bitcast(bld_base, TGSI_TYPE_FLOAT,
- si_llvm_pack_two_int32_as_int16(gallivm, val+2));
+ si_llvm_pack_two_int32_as_int16(ctx, val+2));
break;
}
LLVMValueRef alpha)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
LLVMValueRef alpha_ref = LLVMGetParam(ctx->main_fn,
LLVMValueRef arg =
lp_build_select(&bld_base->base,
alpha_pass,
- lp_build_const_float(gallivm, 1.0f),
- lp_build_const_float(gallivm, -1.0f));
+ LLVMConstReal(ctx->f32, 1.0f),
+ LLVMConstReal(ctx->f32, -1.0f));
- ac_emit_kill(&ctx->ac, arg);
+ ac_build_kill(&ctx->ac, arg);
} else {
- ac_emit_kill(&ctx->ac, NULL);
+ ac_build_kill(&ctx->ac, NULL);
}
}
unsigned samplemask_param)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef coverage;
/* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
ctx->f32, "");
coverage = LLVMBuildFMul(gallivm->builder, coverage,
- lp_build_const_float(gallivm,
+ LLVMConstReal(ctx->f32,
1.0 / SI_NUM_SMOOTH_AA_SAMPLES), "");
return LLVMBuildFMul(gallivm->builder, alpha, coverage, "");
unsigned const_chan;
LLVMValueRef base_elt;
LLVMValueRef ptr = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
- LLVMValueRef constbuf_index = lp_build_const_int32(base->gallivm,
- SI_VS_CONST_CLIP_PLANES);
+ LLVMValueRef constbuf_index = LLVMConstInt(ctx->i32,
+ SI_VS_CONST_CLIP_PLANES, 0);
LLVMValueRef const_resource = ac_build_indexed_load_const(&ctx->ac, ptr, constbuf_index);
for (reg_index = 0; reg_index < 2; reg_index ++) {
args->out[0] =
args->out[1] =
args->out[2] =
- args->out[3] = lp_build_const_float(base->gallivm, 0.0f);
+ args->out[3] = LLVMConstReal(ctx->f32, 0.0f);
/* Compute dot products of position and user clip plane vectors */
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf_idx],
vdata, num_comps,
so_write_offsets[buf_idx],
- LLVMConstInt(ctx->i32, 0, 0),
- stream_out->dst_offset * 4, 1, 1);
+ ctx->i32_0,
+ stream_out->dst_offset * 4, 1, 1, true, false);
}
/**
if (!so->stride[i])
continue;
- LLVMValueRef offset = lp_build_const_int32(gallivm,
- SI_VS_STREAMOUT_BUF0 + i);
+ LLVMValueRef offset = LLVMConstInt(ctx->i32,
+ SI_VS_STREAMOUT_BUF0 + i, 0);
so_buffers[i] = ac_build_indexed_load_const(&ctx->ac, buf_ptr, offset);
memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
&args, sizeof(args));
} else {
- ac_emit_export(&ctx->ac, &args);
+ ac_build_export(&ctx->ac, &args);
}
if (semantic_name == TGSI_SEMANTIC_CLIPDIST) {
if (shader->selector->info.writes_edgeflag) {
/* The output is a float, but the hw expects an integer
* with the first bit containing the edge flag. */
- edgeflag_value = LLVMBuildFPToUI(base->gallivm->builder,
+ edgeflag_value = LLVMBuildFPToUI(ctx->gallivm.builder,
edgeflag_value,
ctx->i32, "");
edgeflag_value = lp_build_min(&bld_base->int_bld,
edgeflag_value,
- bld_base->int_bld.one);
+ ctx->i32_1);
/* The LLVM intrinsic expects a float. */
- pos_args[1].out[1] = LLVMBuildBitCast(base->gallivm->builder,
+ pos_args[1].out[1] = LLVMBuildBitCast(ctx->gallivm.builder,
edgeflag_value,
ctx->f32, "");
}
/* Specify that this is the last export */
pos_args[i].done = 1;
- ac_emit_export(&ctx->ac, &pos_args[i]);
+ ac_build_export(&ctx->ac, &pos_args[i]);
}
}
static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef invocation_id, rw_buffers, buffer, buffer_offset;
LLVMValueRef lds_vertex_stride, lds_vertex_offset, lds_base;
uint64_t inputs;
rw_buffers = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
- lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
+ LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
buffer_offset = LLVMGetParam(ctx->main_fn, ctx->param_oc_lds);
- lds_vertex_stride = unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 13, 8);
+ lds_vertex_stride = unpack_param(ctx, SI_PARAM_TCS_IN_LAYOUT, 24, 8);
lds_vertex_offset = LLVMBuildMul(gallivm->builder, invocation_id,
lds_vertex_stride, "");
lds_base = get_tcs_in_current_patch_offset(ctx);
unsigned i = u_bit_scan64(&inputs);
LLVMValueRef lds_ptr = LLVMBuildAdd(gallivm->builder, lds_base,
- lp_build_const_int32(gallivm, 4 * i),
+ LLVMConstInt(ctx->i32, 4 * i, 0),
"");
LLVMValueRef buffer_addr = get_tcs_tes_buffer_address(ctx,
get_rel_patch_id(ctx),
invocation_id,
- lp_build_const_int32(gallivm, i));
+ LLVMConstInt(ctx->i32, i, 0));
LLVMValueRef value = lds_load(bld_base, TGSI_TYPE_SIGNED, ~0,
lds_ptr);
ac_build_buffer_store_dword(&ctx->ac, buffer, value, 4, buffer_addr,
- buffer_offset, 0, 1, 0);
+ buffer_offset, 0, 1, 0, true, false);
}
}
LLVMValueRef tcs_out_current_patch_data_offset)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_shader *shader = ctx->shader;
unsigned tess_inner_index, tess_outer_index;
LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
*/
lp_build_if(&if_ctx, gallivm,
LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
- invocation_id, bld_base->uint_bld.zero, ""));
+ invocation_id, ctx->i32_0, ""));
/* Determine the layout of one tess factor element in the buffer. */
switch (shader->key.part.tcs.epilog.prim_mode) {
lds_base = tcs_out_current_patch_data_offset;
lds_inner = LLVMBuildAdd(gallivm->builder, lds_base,
- lp_build_const_int32(gallivm,
- tess_inner_index * 4), "");
+ LLVMConstInt(ctx->i32,
+ tess_inner_index * 4, 0), "");
lds_outer = LLVMBuildAdd(gallivm->builder, lds_base,
- lp_build_const_int32(gallivm,
- tess_outer_index * 4), "");
+ LLVMConstInt(ctx->i32,
+ tess_outer_index * 4, 0), "");
for (i = 0; i < 4; i++) {
inner[i] = LLVMGetUndef(ctx->i32);
rw_buffers = LLVMGetParam(ctx->main_fn,
SI_PARAM_RW_BUFFERS);
buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
- lp_build_const_int32(gallivm, SI_HS_RING_TESS_FACTOR));
+ LLVMConstInt(ctx->i32, SI_HS_RING_TESS_FACTOR, 0));
/* Get the offset. */
tf_base = LLVMGetParam(ctx->main_fn,
SI_PARAM_TESS_FACTOR_OFFSET);
byteoffset = LLVMBuildMul(gallivm->builder, rel_patch_id,
- lp_build_const_int32(gallivm, 4 * stride), "");
+ LLVMConstInt(ctx->i32, 4 * stride, 0), "");
lp_build_if(&inner_if_ctx, gallivm,
LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
- rel_patch_id, bld_base->uint_bld.zero, ""));
+ rel_patch_id, ctx->i32_0, ""));
/* Store the dynamic HS control word. */
ac_build_buffer_store_dword(&ctx->ac, buffer,
- lp_build_const_int32(gallivm, 0x80000000),
- 1, lp_build_const_int32(gallivm, 0), tf_base,
- 0, 1, 0);
+ LLVMConstInt(ctx->i32, 0x80000000, 0),
+ 1, ctx->i32_0, tf_base,
+ 0, 1, 0, true, false);
lp_build_endif(&inner_if_ctx);
/* Store the tessellation factors. */
ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
MIN2(stride, 4), byteoffset, tf_base,
- 4, 1, 0);
+ 4, 1, 0, true, false);
if (vec1)
ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
stride - 4, byteoffset, tf_base,
- 20, 1, 0);
+ 20, 1, 0, true, false);
/* Store the tess factors into the offchip buffer if TES reads them. */
if (shader->key.part.tcs.epilog.tes_reads_tess_factors) {
ac_build_buffer_store_dword(&ctx->ac, buf, outer_vec,
outer_comps, tf_outer_offset,
- base, 0, 1, 0);
+ base, 0, 1, 0, true, false);
if (inner_comps) {
param_inner = si_shader_io_get_unique_index(
TGSI_SEMANTIC_TESSINNER, 0);
lp_build_gather_values(gallivm, inner, inner_comps);
ac_build_buffer_store_dword(&ctx->ac, buf, inner_vec,
inner_comps, tf_inner_offset,
- base, 0, 1, 0);
+ base, 0, 1, 0, true, false);
}
}
tf_lds_offset = get_tcs_out_current_patch_data_offset(ctx);
/* Return epilog parameters from this function. */
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMBuilderRef builder = ctx->gallivm.builder;
LLVMValueRef ret = ctx->return_value;
LLVMValueRef rw_buffers, rw0, rw1, tf_soffset;
unsigned vgpr;
rw_buffers = LLVMBuildPtrToInt(builder, rw_buffers, ctx->i64, "");
rw_buffers = LLVMBuildBitCast(builder, rw_buffers, ctx->v2i32, "");
rw0 = LLVMBuildExtractElement(builder, rw_buffers,
- bld_base->uint_bld.zero, "");
+ ctx->i32_0, "");
rw1 = LLVMBuildExtractElement(builder, rw_buffers,
- bld_base->uint_bld.one, "");
+ ctx->i32_1, "");
ret = LLVMBuildInsertValue(builder, ret, rw0, 0, "");
ret = LLVMBuildInsertValue(builder, ret, rw1, 1, "");
struct si_shader_context *ctx = si_shader_context(bld_base);
struct si_shader *shader = ctx->shader;
struct tgsi_shader_info *info = &shader->selector->info;
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
unsigned i, chan;
LLVMValueRef vertex_id = LLVMGetParam(ctx->main_fn,
ctx->param_rel_auto_id);
LLVMValueRef vertex_dw_stride =
- unpack_param(ctx, SI_PARAM_LS_OUT_LAYOUT, 13, 8);
+ unpack_param(ctx, SI_PARAM_LS_OUT_LAYOUT, 24, 8);
LLVMValueRef base_dw_addr = LLVMBuildMul(gallivm->builder, vertex_id,
vertex_dw_stride, "");
unsigned index = info->output_semantic_index[i];
int param = si_shader_io_get_unique_index(name, index);
LLVMValueRef dw_addr = LLVMBuildAdd(gallivm->builder, base_dw_addr,
- lp_build_const_int32(gallivm, param * 4), "");
+ LLVMConstInt(ctx->i32, param * 4, 0), "");
for (chan = 0; chan < 4; chan++) {
lds_store(bld_base, chan, dw_addr,
static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context *bld_base)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_shader *es = ctx->shader;
struct tgsi_shader_info *info = &es->selector->info;
LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
ctx->esgs_ring,
out_val, 1, NULL, soffset,
(4 * param_index + chan) * 4,
- 1, 1);
+ 1, 1, true, true);
}
}
}
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- ac_emit_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE,
- LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
+ ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE,
+ LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
}
static void si_llvm_emit_vs_epilogue(struct lp_build_tgsi_context *bld_base)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct tgsi_shader_info *info = &ctx->shader->selector->info;
struct si_shader_output_values *outputs = NULL;
int i,j;
for (j = 0; j < 4; j++) {
addr = ctx->outputs[i][j];
val = LLVMBuildLoad(gallivm->builder, addr, "");
- val = ac_emit_clamp(&ctx->ac, val);
+ val = ac_build_clamp(&ctx->ac, val);
LLVMBuildStore(gallivm->builder, val, addr);
}
}
if (stencil) {
/* Stencil should be in X[23:16]. */
stencil = bitcast(bld_base, TGSI_TYPE_UNSIGNED, stencil);
- stencil = LLVMBuildShl(base->gallivm->builder, stencil,
+ stencil = LLVMBuildShl(ctx->gallivm.builder, stencil,
LLVMConstInt(ctx->i32, 16, 0), "");
args.out[0] = bitcast(bld_base, TGSI_TYPE_FLOAT, stencil);
mask |= 0x3;
/* Clamp color */
if (ctx->shader->key.part.ps.epilog.clamp_color)
for (i = 0; i < 4; i++)
- color[i] = ac_emit_clamp(&ctx->ac, color[i]);
+ color[i] = ac_build_clamp(&ctx->ac, color[i]);
/* Alpha to one */
if (ctx->shader->key.part.ps.epilog.alpha_to_one)
struct si_ps_exports *exp)
{
for (unsigned i = 0; i < exp->num; i++)
- ac_emit_export(&ctx->ac, &exp->args[i]);
+ ac_build_export(&ctx->ac, &exp->args[i]);
}
static void si_export_null(struct lp_build_tgsi_context *bld_base)
args.out[2] = base->undef; /* B */
args.out[3] = base->undef; /* A */
- ac_emit_export(&ctx->ac, &args);
+ ac_build_export(&ctx->ac, &args);
}
/**
{
struct si_shader_context *ctx = si_shader_context(bld_base);
struct si_shader *shader = ctx->shader;
- struct lp_build_context *base = &bld_base->base;
struct tgsi_shader_info *info = &shader->selector->info;
- LLVMBuilderRef builder = base->gallivm->builder;
+ LLVMBuilderRef builder = ctx->gallivm.builder;
unsigned i, j, first_vgpr, vgpr;
LLVMValueRef color[8][4] = {};
LLVMValueRef descriptor)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef size =
LLVMBuildExtractElement(builder, descriptor,
- lp_build_const_int32(gallivm, 2), "");
+ LLVMConstInt(ctx->i32, 2, 0), "");
- if (ctx->screen->b.chip_class >= VI) {
+ if (ctx->screen->b.chip_class == VI) {
/* On VI, the descriptor contains the size in bytes,
* but TXQ must return the size in elements.
* The stride is always non-zero for resources using TXQ.
*/
LLVMValueRef stride =
LLVMBuildExtractElement(builder, descriptor,
- lp_build_const_int32(gallivm, 1), "");
+ ctx->i32_1, "");
stride = LLVMBuildLShr(builder, stride,
- lp_build_const_int32(gallivm, 16), "");
+ LLVMConstInt(ctx->i32, 16, 0), "");
stride = LLVMBuildAnd(builder, stride,
- lp_build_const_int32(gallivm, 0x3FFF), "");
+ LLVMConstInt(ctx->i32, 0x3FFF, 0), "");
size = LLVMBuildUDiv(builder, size, stride, "");
}
/* Prevent optimizations (at least of memory accesses) across the current
* point in the program by emitting empty inline assembly that is marked as
* having side effects.
+ *
+ * Optionally, a value can be passed through the inline assembly to prevent
+ * LLVM from hoisting calls to ReadNone functions.
*/
-#if 0 /* unused currently */
-static void emit_optimization_barrier(struct si_shader_context *ctx)
+static void emit_optimization_barrier(struct si_shader_context *ctx,
+ LLVMValueRef *pvgpr)
{
+ static int counter = 0;
+
LLVMBuilderRef builder = ctx->gallivm.builder;
- LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
- LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, "", "", true, false);
- LLVMBuildCall(builder, inlineasm, NULL, 0, "");
+ char code[16];
+
+ snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
+
+ if (!pvgpr) {
+ LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
+ LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
+ LLVMBuildCall(builder, inlineasm, NULL, 0, "");
+ } else {
+ LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
+ LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
+ LLVMValueRef vgpr = *pvgpr;
+ LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
+ unsigned vgpr_size = llvm_get_type_size(vgpr_type);
+ LLVMValueRef vgpr0;
+
+ assert(vgpr_size % 4 == 0);
+
+ vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
+ vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
+ vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
+ vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
+ vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
+
+ *pvgpr = vgpr;
+ }
}
-#endif
/* Combine these with & instead of |. */
#define NOOP_WAITCNT 0xf7f
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef args[1] = {
- lp_build_const_int32(gallivm, simm16)
+ LLVMConstInt(ctx->i32, simm16, 0)
};
lp_build_intrinsic(builder, "llvm.amdgcn.s.waitcnt",
ctx->voidt, args, 1, 0);
emit_waitcnt(ctx, waitcnt);
}
+static void clock_emit(
+ const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMValueRef tmp;
+
+ tmp = lp_build_intrinsic(gallivm->builder, "llvm.readcyclecounter",
+ ctx->i64, NULL, 0, 0);
+ tmp = LLVMBuildBitCast(gallivm->builder, tmp, ctx->v2i32, "");
+
+ emit_data->output[0] =
+ LLVMBuildExtractElement(gallivm->builder, tmp, ctx->i32_0, "");
+ emit_data->output[1] =
+ LLVMBuildExtractElement(gallivm->builder, tmp, ctx->i32_1, "");
+}
+
static LLVMValueRef
shader_buffer_fetch_rsrc(struct si_shader_context *ctx,
const struct tgsi_full_src_register *reg)
CONST_ADDR_SPACE);
}
+static LLVMValueRef load_image_desc(struct si_shader_context *ctx,
+ LLVMValueRef list, LLVMValueRef index,
+ unsigned target)
+{
+ LLVMBuilderRef builder = ctx->gallivm.builder;
+
+ if (target == TGSI_TEXTURE_BUFFER) {
+ index = LLVMBuildMul(builder, index,
+ LLVMConstInt(ctx->i32, 2, 0), "");
+ index = LLVMBuildAdd(builder, index,
+ ctx->i32_1, "");
+ list = LLVMBuildPointerCast(builder, list,
+ const_array(ctx->v4i32, 0), "");
+ }
+
+ return ac_build_indexed_load_const(&ctx->ac, list, index);
+}
+
/**
* Load the resource descriptor for \p image.
*/
struct si_shader_context *ctx = si_shader_context(bld_base);
LLVMValueRef rsrc_ptr = LLVMGetParam(ctx->main_fn,
SI_PARAM_IMAGES);
- LLVMValueRef index, tmp;
- bool dcc_off = target != TGSI_TEXTURE_BUFFER && is_store;
+ LLVMValueRef index;
+ bool dcc_off = is_store;
assert(image->Register.File == TGSI_FILE_IMAGE);
index = LLVMConstInt(ctx->i32, image->Register.Index, 0);
- if (images_writemask & (1 << image->Register.Index) &&
- target != TGSI_TEXTURE_BUFFER)
+ if (images_writemask & (1 << image->Register.Index))
dcc_off = true;
} else {
/* From the GL_ARB_shader_image_load_store extension spec:
SI_NUM_IMAGES);
}
- if (target == TGSI_TEXTURE_BUFFER) {
- LLVMBuilderRef builder = ctx->gallivm.builder;
-
- rsrc_ptr = LLVMBuildPointerCast(builder, rsrc_ptr,
- const_array(ctx->v4i32, 0), "");
- index = LLVMBuildMul(builder, index,
- LLVMConstInt(ctx->i32, 2, 0), "");
- index = LLVMBuildAdd(builder, index,
- LLVMConstInt(ctx->i32, 1, 0), "");
- *rsrc = ac_build_indexed_load_const(&ctx->ac, rsrc_ptr, index);
- return;
- }
-
- tmp = ac_build_indexed_load_const(&ctx->ac, rsrc_ptr, index);
- if (dcc_off)
- tmp = force_dcc_off(ctx, tmp);
- *rsrc = tmp;
+ *rsrc = load_image_desc(ctx, rsrc_ptr, index, target);
+ if (dcc_off && target != TGSI_TEXTURE_BUFFER)
+ *rsrc = force_dcc_off(ctx, *rsrc);
}
static LLVMValueRef image_fetch_coords(
const struct tgsi_full_instruction *inst,
unsigned src)
{
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
unsigned target = inst->Memory.Texture;
unsigned num_coords = tgsi_util_get_texture_coord_dim(target);
for (chan = 0; chan < num_coords; ++chan) {
tmp = lp_build_emit_fetch(bld_base, inst, src, chan);
- tmp = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
+ tmp = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
coords[chan] = tmp;
}
+ /* 1D textures are allocated and used as 2D on GFX9. */
+ if (ctx->screen->b.chip_class >= GFX9) {
+ if (target == TGSI_TEXTURE_1D) {
+ coords[1] = ctx->i32_0;
+ num_coords++;
+ } else if (target == TGSI_TEXTURE_1D_ARRAY) {
+ coords[2] = coords[1];
+ coords[1] = ctx->i32_0;
+ }
+ }
+
if (num_coords == 1)
return coords[0];
struct lp_build_emit_data * emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
const struct tgsi_full_instruction * inst = emit_data->inst;
unsigned target = inst->Memory.Texture;
LLVMValueRef rsrc;
- emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
+ emit_data->dst_type = ctx->v4f32;
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
LLVMBuilderRef builder = gallivm->builder;
rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]);
tmp = lp_build_emit_fetch(bld_base, inst, 1, 0);
- offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
+ offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
- buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
+ buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
offset, false, false);
} else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
LLVMValueRef coords;
if (target == TGSI_TEXTURE_BUFFER) {
buffer_append_args(ctx, emit_data, rsrc, coords,
- bld_base->uint_bld.zero, false, false);
+ ctx->i32_0, false, false);
} else {
emit_data->args[0] = coords;
emit_data->args[1] = rsrc;
- emit_data->args[2] = lp_build_const_int32(gallivm, 15); /* dmask */
+ emit_data->args[2] = LLVMConstInt(ctx->i32, 15, 0); /* dmask */
emit_data->arg_count = 3;
image_append_args(ctx, emit_data, target, false, false);
}
}
+static unsigned get_load_intr_attribs(bool readonly_memory)
+{
+ /* READNONE means writes can't affect it, while READONLY means that
+ * writes can affect it. */
+ return readonly_memory && HAVE_LLVM >= 0x0400 ?
+ LP_FUNC_ATTR_READNONE :
+ LP_FUNC_ATTR_READONLY;
+}
+
+static unsigned get_store_intr_attribs(bool writeonly_memory)
+{
+ return writeonly_memory && HAVE_LLVM >= 0x0400 ?
+ LP_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
+ LP_FUNC_ATTR_WRITEONLY;
+}
+
static void load_emit_buffer(struct si_shader_context *ctx,
- struct lp_build_emit_data *emit_data)
+ struct lp_build_emit_data *emit_data,
+ bool readonly_memory)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
struct gallivm_state *gallivm = &ctx->gallivm;
emit_data->output[emit_data->chan] = lp_build_intrinsic(
builder, intrinsic_name, dst_type,
emit_data->args, emit_data->arg_count,
- LP_FUNC_ATTR_READONLY);
+ get_load_intr_attribs(readonly_memory));
}
static LLVMValueRef get_memory_ptr(struct si_shader_context *ctx,
struct lp_build_emit_data *emit_data)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
- struct lp_build_context *base = &ctx->bld_base.base;
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
unsigned writemask = inst->Dst[0].Register.WriteMask;
LLVMValueRef channels[4], ptr, derived_ptr, index;
int chan;
- ptr = get_memory_ptr(ctx, inst, base->elem_type, 1);
+ ptr = get_memory_ptr(ctx, inst, ctx->f32, 1);
for (chan = 0; chan < 4; ++chan) {
if (!(writemask & (1 << chan))) {
- channels[chan] = LLVMGetUndef(base->elem_type);
+ channels[chan] = LLVMGetUndef(ctx->f32);
continue;
}
- index = lp_build_const_int32(gallivm, chan);
+ index = LLVMConstInt(ctx->i32, chan, 0);
derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
channels[chan] = LLVMBuildLoad(builder, derived_ptr, "");
}
emit_data->output[emit_data->chan] = lp_build_gather_values(gallivm, channels, 4);
}
-static void get_image_intr_name(const char *base_name,
- LLVMTypeRef data_type,
- LLVMTypeRef coords_type,
- LLVMTypeRef rsrc_type,
- char *out_name, unsigned out_len)
-{
- char coords_type_name[8];
+/**
+ * Return true if the memory accessed by a LOAD or STORE instruction is
+ * read-only or write-only, respectively.
+ *
+ * \param shader_buffers_reverse_access_mask
+ * For LOAD, set this to (store | atomic) slot usage in the shader.
+ * For STORE, set this to (load | atomic) slot usage in the shader.
+ * \param images_reverse_access_mask Same as above, but for images.
+ */
+static bool is_oneway_access_only(const struct tgsi_full_instruction *inst,
+ const struct tgsi_shader_info *info,
+ unsigned shader_buffers_reverse_access_mask,
+ unsigned images_reverse_access_mask)
+{
+ /* RESTRICT means NOALIAS.
+ * If there are no writes, we can assume the accessed memory is read-only.
+ * If there are no reads, we can assume the accessed memory is write-only.
+ */
+ if (inst->Memory.Qualifier & TGSI_MEMORY_RESTRICT) {
+ unsigned reverse_access_mask;
+
+ if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
+ reverse_access_mask = shader_buffers_reverse_access_mask;
+ } else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
+ reverse_access_mask = info->images_buffers &
+ images_reverse_access_mask;
+ } else {
+ reverse_access_mask = ~info->images_buffers &
+ images_reverse_access_mask;
+ }
- ac_build_type_name_for_intr(coords_type, coords_type_name,
- sizeof(coords_type_name));
+ if (inst->Src[0].Register.Indirect) {
+ if (!reverse_access_mask)
+ return true;
+ } else {
+ if (!(reverse_access_mask &
+ (1u << inst->Src[0].Register.Index)))
+ return true;
+ }
+ }
- if (HAVE_LLVM <= 0x0309) {
- snprintf(out_name, out_len, "%s.%s", base_name, coords_type_name);
+ /* If there are no buffer writes (for both shader buffers & image
+ * buffers), it implies that buffer memory is read-only.
+ * If there are no buffer reads (for both shader buffers & image
+ * buffers), it implies that buffer memory is write-only.
+ *
+ * Same for the case when there are no writes/reads for non-buffer
+ * images.
+ */
+ if (inst->Src[0].Register.File == TGSI_FILE_BUFFER ||
+ (inst->Src[0].Register.File == TGSI_FILE_IMAGE &&
+ inst->Memory.Texture == TGSI_TEXTURE_BUFFER)) {
+ if (!shader_buffers_reverse_access_mask &&
+ !(info->images_buffers & images_reverse_access_mask))
+ return true;
} else {
- char data_type_name[8];
- char rsrc_type_name[8];
-
- ac_build_type_name_for_intr(data_type, data_type_name,
- sizeof(data_type_name));
- ac_build_type_name_for_intr(rsrc_type, rsrc_type_name,
- sizeof(rsrc_type_name));
- snprintf(out_name, out_len, "%s.%s.%s.%s", base_name,
- data_type_name, coords_type_name, rsrc_type_name);
+ if (!(~info->images_buffers & images_reverse_access_mask))
+ return true;
}
+ return false;
}
static void load_emit(
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
+ const struct tgsi_shader_info *info = &ctx->shader->selector->info;
char intrinsic_name[64];
+ bool readonly_memory = false;
if (inst->Src[0].Register.File == TGSI_FILE_MEMORY) {
load_emit_memory(ctx, emit_data);
if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
emit_waitcnt(ctx, VM_CNT);
+ readonly_memory = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) &&
+ is_oneway_access_only(inst, info,
+ info->shader_buffers_store |
+ info->shader_buffers_atomic,
+ info->images_store |
+ info->images_atomic);
+
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
- load_emit_buffer(ctx, emit_data);
+ load_emit_buffer(ctx, emit_data, readonly_memory);
return;
}
lp_build_intrinsic(
builder, "llvm.amdgcn.buffer.load.format.v4f32", emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LP_FUNC_ATTR_READONLY);
+ get_load_intr_attribs(readonly_memory));
} else {
- get_image_intr_name("llvm.amdgcn.image.load",
- emit_data->dst_type, /* vdata */
- LLVMTypeOf(emit_data->args[0]), /* coords */
- LLVMTypeOf(emit_data->args[1]), /* rsrc */
- intrinsic_name, sizeof(intrinsic_name));
+ ac_get_image_intr_name("llvm.amdgcn.image.load",
+ emit_data->dst_type, /* vdata */
+ LLVMTypeOf(emit_data->args[0]), /* coords */
+ LLVMTypeOf(emit_data->args[1]), /* rsrc */
+ intrinsic_name, sizeof(intrinsic_name));
emit_data->output[emit_data->chan] =
lp_build_intrinsic(
builder, intrinsic_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LP_FUNC_ATTR_READONLY);
+ get_load_intr_attribs(readonly_memory));
}
}
struct lp_build_emit_data * emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
struct tgsi_full_src_register memory;
rsrc = shader_buffer_fetch_rsrc(ctx, &memory);
tmp = lp_build_emit_fetch(bld_base, inst, 0, 0);
- offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
+ offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
- buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
+ buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
offset, false, false);
} else if (inst->Dst[0].Register.File == TGSI_FILE_IMAGE) {
unsigned target = inst->Memory.Texture;
if (target == TGSI_TEXTURE_BUFFER) {
image_fetch_rsrc(bld_base, &memory, true, target, &rsrc);
buffer_append_args(ctx, emit_data, rsrc, coords,
- bld_base->uint_bld.zero, false, force_glc);
+ ctx->i32_0, false, force_glc);
} else {
emit_data->args[1] = coords;
image_fetch_rsrc(bld_base, &memory, true, target,
&emit_data->args[2]);
- emit_data->args[3] = lp_build_const_int32(gallivm, 15); /* dmask */
+ emit_data->args[3] = LLVMConstInt(ctx->i32, 15, 0); /* dmask */
emit_data->arg_count = 4;
image_append_args(ctx, emit_data, target, false, force_glc);
static void store_emit_buffer(
struct si_shader_context *ctx,
- struct lp_build_emit_data *emit_data)
+ struct lp_build_emit_data *emit_data,
+ bool writeonly_memory)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
- struct lp_build_context *uint_bld = &ctx->bld_base.uint_bld;
LLVMValueRef base_data = emit_data->args[0];
LLVMValueRef base_offset = emit_data->args[3];
unsigned writemask = inst->Dst[0].Register.WriteMask;
tmp = LLVMBuildExtractElement(
builder, base_data,
- lp_build_const_int32(gallivm, start), "");
+ LLVMConstInt(ctx->i32, start, 0), "");
data = LLVMBuildInsertElement(
builder, LLVMGetUndef(v2f32), tmp,
- uint_bld->zero, "");
+ ctx->i32_0, "");
tmp = LLVMBuildExtractElement(
builder, base_data,
- lp_build_const_int32(gallivm, start + 1), "");
+ LLVMConstInt(ctx->i32, start + 1, 0), "");
data = LLVMBuildInsertElement(
- builder, data, tmp, uint_bld->one, "");
+ builder, data, tmp, ctx->i32_1, "");
intrinsic_name = "llvm.amdgcn.buffer.store.v2f32";
} else {
assert(count == 1);
data = LLVMBuildExtractElement(
builder, base_data,
- lp_build_const_int32(gallivm, start), "");
+ LLVMConstInt(ctx->i32, start, 0), "");
intrinsic_name = "llvm.amdgcn.buffer.store.f32";
}
if (start != 0) {
offset = LLVMBuildAdd(
builder, offset,
- lp_build_const_int32(gallivm, start * 4), "");
+ LLVMConstInt(ctx->i32, start * 4, 0), "");
}
emit_data->args[0] = data;
lp_build_intrinsic(
builder, intrinsic_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LP_FUNC_ATTR_WRITEONLY);
+ get_store_intr_attribs(writeonly_memory));
}
}
{
const struct tgsi_full_instruction *inst = emit_data->inst;
struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_context *base = &ctx->bld_base.base;
LLVMBuilderRef builder = gallivm->builder;
unsigned writemask = inst->Dst[0].Register.WriteMask;
LLVMValueRef ptr, derived_ptr, data, index;
int chan;
- ptr = get_memory_ptr(ctx, inst, base->elem_type, 0);
+ ptr = get_memory_ptr(ctx, inst, ctx->f32, 0);
for (chan = 0; chan < 4; ++chan) {
if (!(writemask & (1 << chan))) {
continue;
}
data = lp_build_emit_fetch(&ctx->bld_base, inst, 1, chan);
- index = lp_build_const_int32(gallivm, chan);
+ index = LLVMConstInt(ctx->i32, chan, 0);
derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
LLVMBuildStore(builder, data, derived_ptr);
}
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
+ const struct tgsi_shader_info *info = &ctx->shader->selector->info;
unsigned target = inst->Memory.Texture;
char intrinsic_name[64];
+ bool writeonly_memory = false;
if (inst->Dst[0].Register.File == TGSI_FILE_MEMORY) {
store_emit_memory(ctx, emit_data);
if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
emit_waitcnt(ctx, VM_CNT);
+ writeonly_memory = is_oneway_access_only(inst, info,
+ info->shader_buffers_load |
+ info->shader_buffers_atomic,
+ info->images_load |
+ info->images_atomic);
+
if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
- store_emit_buffer(ctx, emit_data);
+ store_emit_buffer(ctx, emit_data, writeonly_memory);
return;
}
builder, "llvm.amdgcn.buffer.store.format.v4f32",
emit_data->dst_type, emit_data->args,
emit_data->arg_count,
- LP_FUNC_ATTR_WRITEONLY);
+ get_store_intr_attribs(writeonly_memory));
} else {
- get_image_intr_name("llvm.amdgcn.image.store",
- LLVMTypeOf(emit_data->args[0]), /* vdata */
- LLVMTypeOf(emit_data->args[1]), /* coords */
- LLVMTypeOf(emit_data->args[2]), /* rsrc */
- intrinsic_name, sizeof(intrinsic_name));
+ ac_get_image_intr_name("llvm.amdgcn.image.store",
+ LLVMTypeOf(emit_data->args[0]), /* vdata */
+ LLVMTypeOf(emit_data->args[1]), /* coords */
+ LLVMTypeOf(emit_data->args[2]), /* rsrc */
+ intrinsic_name, sizeof(intrinsic_name));
emit_data->output[emit_data->chan] =
lp_build_intrinsic(
builder, intrinsic_name, emit_data->dst_type,
emit_data->args, emit_data->arg_count,
- LP_FUNC_ATTR_WRITEONLY);
+ get_store_intr_attribs(writeonly_memory));
}
}
struct lp_build_emit_data * emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
LLVMValueRef data1, data2;
LLVMValueRef rsrc;
LLVMValueRef tmp;
- emit_data->dst_type = bld_base->base.elem_type;
+ emit_data->dst_type = ctx->f32;
tmp = lp_build_emit_fetch(bld_base, inst, 2, 0);
- data1 = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
+ data1 = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
tmp = lp_build_emit_fetch(bld_base, inst, 3, 0);
- data2 = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
+ data2 = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
}
/* llvm.amdgcn.image/buffer.atomic.cmpswap reflect the hardware order
rsrc = shader_buffer_fetch_rsrc(ctx, &inst->Src[0]);
tmp = lp_build_emit_fetch(bld_base, inst, 1, 0);
- offset = LLVMBuildBitCast(builder, tmp, bld_base->uint_bld.elem_type, "");
+ offset = LLVMBuildBitCast(builder, tmp, ctx->i32, "");
- buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
+ buffer_append_args(ctx, emit_data, rsrc, ctx->i32_0,
offset, true, false);
} else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
unsigned target = inst->Memory.Texture;
if (target == TGSI_TEXTURE_BUFFER) {
buffer_append_args(ctx, emit_data, rsrc, coords,
- bld_base->uint_bld.zero, true, false);
+ ctx->i32_0, true, false);
} else {
emit_data->args[emit_data->arg_count++] = coords;
emit_data->args[emit_data->arg_count++] = rsrc;
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
char intrinsic_name[40];
}
tmp = lp_build_intrinsic(
- builder, intrinsic_name, bld_base->uint_bld.elem_type,
+ builder, intrinsic_name, ctx->i32,
emit_data->args, emit_data->arg_count, 0);
emit_data->output[emit_data->chan] =
- LLVMBuildBitCast(builder, tmp, bld_base->base.elem_type, "");
+ LLVMBuildBitCast(builder, tmp, ctx->f32, "");
}
static void set_tex_fetch_args(struct si_shader_context *ctx,
memcpy(emit_data->args, &args, sizeof(args));
}
+static LLVMValueRef fix_resinfo(struct si_shader_context *ctx,
+ unsigned target, LLVMValueRef out)
+{
+ LLVMBuilderRef builder = ctx->gallivm.builder;
+
+ /* 1D textures are allocated and used as 2D on GFX9. */
+ if (ctx->screen->b.chip_class >= GFX9 &&
+ (target == TGSI_TEXTURE_1D_ARRAY ||
+ target == TGSI_TEXTURE_SHADOW1D_ARRAY)) {
+ LLVMValueRef layers =
+ LLVMBuildExtractElement(builder, out,
+ LLVMConstInt(ctx->i32, 2, 0), "");
+ out = LLVMBuildInsertElement(builder, out, layers,
+ ctx->i32_1, "");
+ }
+
+ /* Divide the number of layers by 6 to get the number of cubes. */
+ if (target == TGSI_TEXTURE_CUBE_ARRAY ||
+ target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
+ LLVMValueRef imm2 = LLVMConstInt(ctx->i32, 2, 0);
+
+ LLVMValueRef z = LLVMBuildExtractElement(builder, out, imm2, "");
+ z = LLVMBuildSDiv(builder, z, LLVMConstInt(ctx->i32, 6, 0), "");
+
+ out = LLVMBuildInsertElement(builder, out, z, imm2, "");
+ }
+ return out;
+}
+
static void resq_fetch_args(
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
image_fetch_rsrc(bld_base, reg, false, inst->Memory.Texture,
&res_ptr);
set_tex_fetch_args(ctx, emit_data, image_target,
- res_ptr, NULL, &bld_base->uint_bld.zero, 1,
+ res_ptr, NULL, &ctx->i32_0, 1,
0xf);
}
}
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction *inst = emit_data->inst;
LLVMValueRef out;
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
out = LLVMBuildExtractElement(builder, emit_data->args[0],
- lp_build_const_int32(gallivm, 2), "");
+ LLVMConstInt(ctx->i32, 2, 0), "");
} else if (inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
out = get_buffer_size(bld_base, emit_data->args[0]);
} else {
memcpy(&args, emit_data->args, sizeof(args)); /* ugly */
args.opcode = ac_image_get_resinfo;
- out = ac_emit_image_opcode(&ctx->ac, &args);
+ out = ac_build_image_opcode(&ctx->ac, &args);
- /* Divide the number of layers by 6 to get the number of cubes. */
- if (inst->Memory.Texture == TGSI_TEXTURE_CUBE_ARRAY) {
- LLVMValueRef imm2 = lp_build_const_int32(gallivm, 2);
- LLVMValueRef imm6 = lp_build_const_int32(gallivm, 6);
-
- LLVMValueRef z = LLVMBuildExtractElement(builder, out, imm2, "");
- z = LLVMBuildSDiv(builder, z, imm6, "");
- out = LLVMBuildInsertElement(builder, out, z, imm2, "");
- }
+ out = fix_resinfo(ctx, inst->Memory.Texture, out);
}
emit_data->output[emit_data->chan] = out;
/**
* Load an image view, fmask view. or sampler state descriptor.
*/
-static LLVMValueRef load_sampler_desc_custom(struct si_shader_context *ctx,
- LLVMValueRef list, LLVMValueRef index,
- enum desc_type type)
+static LLVMValueRef load_sampler_desc(struct si_shader_context *ctx,
+ LLVMValueRef list, LLVMValueRef index,
+ enum desc_type type)
{
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
case DESC_BUFFER:
/* The buffer is in [4:7]. */
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
- index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 1, 0), "");
+ index = LLVMBuildAdd(builder, index, ctx->i32_1, "");
list = LLVMBuildPointerCast(builder, list,
const_array(ctx->v4i32, 0), "");
break;
case DESC_FMASK:
/* The FMASK is at [8:15]. */
index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
- index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 1, 0), "");
+ index = LLVMBuildAdd(builder, index, ctx->i32_1, "");
break;
case DESC_SAMPLER:
/* The sampler state is at [12:15]. */
return ac_build_indexed_load_const(&ctx->ac, list, index);
}
-static LLVMValueRef load_sampler_desc(struct si_shader_context *ctx,
- LLVMValueRef index, enum desc_type type)
-{
- LLVMValueRef list = LLVMGetParam(ctx->main_fn,
- SI_PARAM_SAMPLERS);
-
- return load_sampler_desc_custom(ctx, list, index, type);
-}
-
/* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
*
* SI-CI:
img7 = LLVMBuildExtractElement(builder, res,
LLVMConstInt(ctx->i32, 7, 0), "");
samp0 = LLVMBuildExtractElement(builder, samp,
- LLVMConstInt(ctx->i32, 0, 0), "");
+ ctx->i32_0, "");
samp0 = LLVMBuildAnd(builder, samp0, img7, "");
return LLVMBuildInsertElement(builder, samp, samp0,
- LLVMConstInt(ctx->i32, 0, 0), "");
+ ctx->i32_0, "");
}
static void tex_fetch_ptrs(
LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr, LLVMValueRef *fmask_ptr)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMValueRef list = LLVMGetParam(ctx->main_fn, SI_PARAM_SAMPLERS);
const struct tgsi_full_instruction *inst = emit_data->inst;
+ const struct tgsi_full_src_register *reg;
unsigned target = inst->Texture.Texture;
unsigned sampler_src;
- unsigned sampler_index;
LLVMValueRef index;
sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1;
- sampler_index = emit_data->inst->Src[sampler_src].Register.Index;
-
- if (emit_data->inst->Src[sampler_src].Register.Indirect) {
- const struct tgsi_full_src_register *reg = &emit_data->inst->Src[sampler_src];
+ reg = &emit_data->inst->Src[sampler_src];
+ if (reg->Register.Indirect) {
index = get_bounded_indirect_index(ctx,
®->Indirect,
reg->Register.Index,
SI_NUM_SAMPLERS);
} else {
- index = LLVMConstInt(ctx->i32, sampler_index, 0);
+ index = LLVMConstInt(ctx->i32, reg->Register.Index, 0);
}
if (target == TGSI_TEXTURE_BUFFER)
- *res_ptr = load_sampler_desc(ctx, index, DESC_BUFFER);
+ *res_ptr = load_sampler_desc(ctx, list, index, DESC_BUFFER);
else
- *res_ptr = load_sampler_desc(ctx, index, DESC_IMAGE);
+ *res_ptr = load_sampler_desc(ctx, list, index, DESC_IMAGE);
if (samp_ptr)
*samp_ptr = NULL;
if (target == TGSI_TEXTURE_2D_MSAA ||
target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
if (fmask_ptr)
- *fmask_ptr = load_sampler_desc(ctx, index, DESC_FMASK);
+ *fmask_ptr = load_sampler_desc(ctx, list, index,
+ DESC_FMASK);
} else if (target != TGSI_TEXTURE_BUFFER) {
if (samp_ptr) {
- *samp_ptr = load_sampler_desc(ctx, index, DESC_SAMPLER);
+ *samp_ptr = load_sampler_desc(ctx, list, index,
+ DESC_SAMPLER);
*samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
}
}
memcpy(&args, emit_data->args, sizeof(args)); /* ugly */
args.opcode = ac_image_get_resinfo;
- emit_data->output[emit_data->chan] =
- ac_emit_image_opcode(&ctx->ac, &args);
+ LLVMValueRef result = ac_build_image_opcode(&ctx->ac, &args);
- /* Divide the number of layers by 6 to get the number of cubes. */
- if (target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- LLVMValueRef two = lp_build_const_int32(bld_base->base.gallivm, 2);
- LLVMValueRef six = lp_build_const_int32(bld_base->base.gallivm, 6);
-
- LLVMValueRef v4 = emit_data->output[emit_data->chan];
- LLVMValueRef z = LLVMBuildExtractElement(builder, v4, two, "");
- z = LLVMBuildSDiv(builder, z, six, "");
-
- emit_data->output[emit_data->chan] =
- LLVMBuildInsertElement(builder, v4, z, two, "");
- }
+ emit_data->output[emit_data->chan] = fix_resinfo(ctx, target, result);
}
static void tex_fetch_args(
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
const struct tgsi_full_instruction *inst = emit_data->inst;
unsigned opcode = inst->Instruction.Opcode;
unsigned target = inst->Texture.Texture;
emit_data->dst_type = ctx->v4f32;
emit_data->args[0] = LLVMBuildBitCast(gallivm->builder, res_ptr,
ctx->v16i8, "");
- emit_data->args[1] = bld_base->uint_bld.zero;
+ emit_data->args[1] = ctx->i32_0;
emit_data->args[2] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
emit_data->arg_count = 3;
return;
coords[3] = bld_base->base.one;
/* Pack offsets. */
- if (has_offset && opcode != TGSI_OPCODE_TXF) {
+ if (has_offset &&
+ opcode != TGSI_OPCODE_TXF &&
+ opcode != TGSI_OPCODE_TXF_LZ) {
/* The offsets are six-bit signed integers packed like this:
* X=[5:0], Y=[13:8], and Z=[21:16].
*/
offset[chan] = lp_build_emit_fetch_texoffset(bld_base,
emit_data->inst, 0, chan);
offset[chan] = LLVMBuildAnd(gallivm->builder, offset[chan],
- lp_build_const_int32(gallivm, 0x3f), "");
+ LLVMConstInt(ctx->i32, 0x3f, 0), "");
if (chan)
offset[chan] = LLVMBuildShl(gallivm->builder, offset[chan],
- lp_build_const_int32(gallivm, chan*8), "");
+ LLVMConstInt(ctx->i32, chan*8, 0), "");
}
pack = LLVMBuildOr(gallivm->builder, offset[0], offset[1], "");
* Z32_FLOAT, but we don't know that here.
*/
if (ctx->screen->b.chip_class == VI)
- z = ac_emit_clamp(&ctx->ac, z);
+ z = ac_build_clamp(&ctx->ac, z);
address[count++] = z;
}
/* Pack user derivatives */
if (opcode == TGSI_OPCODE_TXD) {
- int param, num_src_deriv_channels;
+ int param, num_src_deriv_channels, num_dst_deriv_channels;
switch (target) {
case TGSI_TEXTURE_3D:
num_src_deriv_channels = 3;
+ num_dst_deriv_channels = 3;
num_deriv_channels = 3;
break;
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_2D_ARRAY:
case TGSI_TEXTURE_SHADOW2D_ARRAY:
num_src_deriv_channels = 2;
+ num_dst_deriv_channels = 2;
num_deriv_channels = 2;
break;
case TGSI_TEXTURE_CUBE:
case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
/* Cube derivatives will be converted to 2D. */
num_src_deriv_channels = 3;
+ num_dst_deriv_channels = 3;
num_deriv_channels = 2;
break;
case TGSI_TEXTURE_1D:
case TGSI_TEXTURE_1D_ARRAY:
case TGSI_TEXTURE_SHADOW1D_ARRAY:
num_src_deriv_channels = 1;
- num_deriv_channels = 1;
+
+ /* 1D textures are allocated and used as 2D on GFX9. */
+ if (ctx->screen->b.chip_class >= GFX9) {
+ num_dst_deriv_channels = 2;
+ num_deriv_channels = 2;
+ } else {
+ num_dst_deriv_channels = 1;
+ num_deriv_channels = 1;
+ }
break;
default:
unreachable("invalid target");
}
- for (param = 0; param < 2; param++)
+ for (param = 0; param < 2; param++) {
for (chan = 0; chan < num_src_deriv_channels; chan++)
- derivs[param * num_src_deriv_channels + chan] =
+ derivs[param * num_dst_deriv_channels + chan] =
lp_build_emit_fetch(bld_base, inst, param+1, chan);
+
+ /* Fill in the rest with zeros. */
+ for (chan = num_src_deriv_channels;
+ chan < num_dst_deriv_channels; chan++)
+ derivs[param * num_dst_deriv_channels + chan] =
+ bld_base->base.zero;
+ }
}
if (target == TGSI_TEXTURE_CUBE ||
if (num_coords > 2)
address[count++] = coords[2];
+ /* 1D textures are allocated and used as 2D on GFX9. */
+ if (ctx->screen->b.chip_class >= GFX9) {
+ LLVMValueRef filler;
+
+ /* Use 0.5, so that we don't sample the border color. */
+ if (opcode == TGSI_OPCODE_TXF)
+ filler = ctx->i32_0;
+ else
+ filler = LLVMConstReal(ctx->f32, 0.5);
+
+ if (target == TGSI_TEXTURE_1D ||
+ target == TGSI_TEXTURE_SHADOW1D) {
+ address[count++] = filler;
+ } else if (target == TGSI_TEXTURE_1D_ARRAY ||
+ target == TGSI_TEXTURE_SHADOW1D_ARRAY) {
+ address[count] = address[count - 1];
+ address[count - 1] = filler;
+ count++;
+ }
+ }
+
/* Pack LOD or sample index */
if (opcode == TGSI_OPCODE_TXL || opcode == TGSI_OPCODE_TXF)
address[count++] = coords[3];
*/
if (target == TGSI_TEXTURE_2D_MSAA ||
target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
- struct lp_build_context *uint_bld = &bld_base->uint_bld;
struct lp_build_emit_data txf_emit_data = *emit_data;
LLVMValueRef txf_address[4];
/* We only need .xy for non-arrays, and .xyz for arrays. */
memcpy(txf_address, address, sizeof(txf_address));
- /* Read FMASK using TXF. */
- inst.Instruction.Opcode = TGSI_OPCODE_TXF;
+ /* Read FMASK using TXF_LZ. */
+ inst.Instruction.Opcode = TGSI_OPCODE_TXF_LZ;
inst.Texture.Texture = target;
txf_emit_data.inst = &inst;
txf_emit_data.chan = 0;
LLVMValueRef fmask =
LLVMBuildExtractElement(gallivm->builder,
txf_emit_data.output[0],
- uint_bld->zero, "");
+ ctx->i32_0, "");
unsigned sample_chan = txf_count; /* the sample index is last */
LLVMValueRef fmask_word1 =
LLVMBuildExtractElement(gallivm->builder, fmask_desc,
- uint_bld->one, "");
+ ctx->i32_1, "");
LLVMValueRef word1_is_nonzero =
LLVMBuildICmp(gallivm->builder, LLVMIntNE,
- fmask_word1, uint_bld->zero, "");
+ fmask_word1, ctx->i32_0, "");
/* Replace the MSAA sample index. */
address[sample_chan] =
final_sample, address[sample_chan], "");
}
- if (opcode == TGSI_OPCODE_TXF) {
+ if (opcode == TGSI_OPCODE_TXF ||
+ opcode == TGSI_OPCODE_TXF_LZ) {
/* add tex offsets */
if (inst->Texture.NumOffsets) {
struct lp_build_context *uint_bld = &bld_base->uint_bld;
txq_emit_data.inst = &txq_inst;
txq_emit_data.dst_type = ctx->v4i32;
set_tex_fetch_args(ctx, &txq_emit_data, target,
- args->resource, NULL,
- &ctx->bld_base.uint_bld.zero,
+ args->resource, NULL, &ctx->i32_0,
1, 0xf);
txq_emit(NULL, &ctx->bld_base, &txq_emit_data);
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct lp_build_context *base = &bld_base->base;
const struct tgsi_full_instruction *inst = emit_data->inst;
struct ac_image_args args;
unsigned opcode = inst->Instruction.Opcode;
unsigned target = inst->Texture.Texture;
if (target == TGSI_TEXTURE_BUFFER) {
- emit_data->output[emit_data->chan] = lp_build_intrinsic(
- base->gallivm->builder,
- "llvm.SI.vs.load.input", emit_data->dst_type,
- emit_data->args, emit_data->arg_count,
- LP_FUNC_ATTR_READNONE | LP_FUNC_ATTR_LEGACY);
+ emit_data->output[emit_data->chan] =
+ ac_build_buffer_load_format(&ctx->ac,
+ emit_data->args[0],
+ emit_data->args[2],
+ emit_data->args[1],
+ true);
return;
}
switch (opcode) {
case TGSI_OPCODE_TXF:
- args.opcode = target == TGSI_TEXTURE_2D_MSAA ||
+ case TGSI_OPCODE_TXF_LZ:
+ args.opcode = opcode == TGSI_OPCODE_TXF_LZ ||
+ target == TGSI_TEXTURE_2D_MSAA ||
target == TGSI_TEXTURE_2D_ARRAY_MSAA ?
ac_image_load : ac_image_load_mip;
args.compare = false;
if (ctx->type != PIPE_SHADER_FRAGMENT)
args.level_zero = true;
break;
+ case TGSI_OPCODE_TEX_LZ:
+ args.level_zero = true;
+ break;
case TGSI_OPCODE_TXB:
case TGSI_OPCODE_TXB2:
assert(ctx->type == PIPE_SHADER_FRAGMENT);
}
/* The hardware needs special lowering for Gather4 with integer formats. */
- if (opcode == TGSI_OPCODE_TG4) {
+ if (ctx->screen->b.chip_class <= VI &&
+ opcode == TGSI_OPCODE_TG4) {
struct tgsi_shader_info *info = &ctx->shader->selector->info;
/* This will also work with non-constant indexing because of how
* glsl_to_tgsi works and we intent to preserve that behavior.
}
emit_data->output[emit_data->chan] =
- ac_emit_image_opcode(&ctx->ac, &args);
+ ac_build_image_opcode(&ctx->ac, &args);
}
static void si_llvm_emit_txqs(
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef res, samples;
LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
res = LLVMBuildBitCast(builder, res_ptr, ctx->v8i32, "");
samples = LLVMBuildExtractElement(
builder, res,
- lp_build_const_int32(gallivm, 3), "");
+ LLVMConstInt(ctx->i32, 3, 0), "");
samples = LLVMBuildLShr(builder, samples,
- lp_build_const_int32(gallivm, 16), "");
+ LLVMConstInt(ctx->i32, 16, 0), "");
samples = LLVMBuildAnd(builder, samples,
- lp_build_const_int32(gallivm, 0xf), "");
- samples = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1),
+ LLVMConstInt(ctx->i32, 0xf, 0), "");
+ samples = LLVMBuildShl(builder, ctx->i32_1,
samples, "");
emit_data->output[emit_data->chan] = samples;
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
unsigned opcode = emit_data->info->opcode;
LLVMValueRef val;
int idx;
idx = (opcode == TGSI_OPCODE_DDX || opcode == TGSI_OPCODE_DDX_FINE) ? 1 : 2;
val = LLVMBuildBitCast(gallivm->builder, emit_data->args[0], ctx->i32, "");
- val = ac_emit_ddxy(&ctx->ac, ctx->screen->has_ds_bpermute,
- mask, idx, ctx->lds, val);
+ val = ac_build_ddxy(&ctx->ac, ctx->screen->has_ds_bpermute,
+ mask, idx, ctx->lds, val);
emit_data->output[emit_data->chan] = val;
}
LLVMValueRef interp_ij)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef result[4], a;
unsigned i;
struct lp_build_emit_data *emit_data)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
const struct tgsi_full_instruction *inst = emit_data->inst;
if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET) {
} else if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
LLVMValueRef sample_position;
LLVMValueRef sample_id;
- LLVMValueRef halfval = lp_build_const_float(gallivm, 0.5f);
+ LLVMValueRef halfval = LLVMConstReal(ctx->f32, 0.5f);
/* fetch sample ID, then fetch its sample position,
* and place into first two channels.
emit_data->args[0] = LLVMBuildExtractElement(gallivm->builder,
sample_position,
- lp_build_const_int32(gallivm, 0), "");
+ ctx->i32_0, "");
emit_data->args[0] = LLVMBuildFSub(gallivm->builder, emit_data->args[0], halfval, "");
emit_data->args[1] = LLVMBuildExtractElement(gallivm->builder,
sample_position,
- lp_build_const_int32(gallivm, 1), "");
+ ctx->i32_1, "");
emit_data->args[1] = LLVMBuildFSub(gallivm->builder, emit_data->args[1], halfval, "");
emit_data->arg_count = 2;
}
{
struct si_shader_context *ctx = si_shader_context(bld_base);
struct si_shader *shader = ctx->shader;
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- struct lp_build_context *uint = &bld_base->uint_bld;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef interp_param;
const struct tgsi_full_instruction *inst = emit_data->inst;
int input_index = inst->Src[0].Register.Index;
else
interp_param = NULL;
- attr_number = lp_build_const_int32(gallivm, input_index);
+ attr_number = LLVMConstInt(ctx->i32, input_index, 0);
if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
* interp_param.J = ddy * offset/sample.y + temp1;
*/
for (i = 0; i < 2; i++) {
- LLVMValueRef ix_ll = lp_build_const_int32(gallivm, i);
- LLVMValueRef iy_ll = lp_build_const_int32(gallivm, i + 2);
+ LLVMValueRef ix_ll = LLVMConstInt(ctx->i32, i, 0);
+ LLVMValueRef iy_ll = LLVMConstInt(ctx->i32, i + 2, 0);
LLVMValueRef ddx_el = LLVMBuildExtractElement(gallivm->builder,
ddxy_out, ix_ll, "");
LLVMValueRef ddy_el = LLVMBuildExtractElement(gallivm->builder,
ij_out[i] = LLVMBuildFAdd(gallivm->builder, temp2, temp1, "");
}
- interp_param = lp_build_gather_values(bld_base->base.gallivm, ij_out, 2);
+ interp_param = lp_build_gather_values(gallivm, ij_out, 2);
}
for (chan = 0; chan < 4; chan++) {
unsigned schan;
schan = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], chan);
- llvm_chan = lp_build_const_int32(gallivm, schan);
+ llvm_chan = LLVMConstInt(ctx->i32, schan, 0);
if (interp_param) {
interp_param = LLVMBuildBitCast(gallivm->builder,
interp_param, LLVMVectorType(ctx->f32, 2), "");
LLVMValueRef i = LLVMBuildExtractElement(
- gallivm->builder, interp_param, uint->zero, "");
+ gallivm->builder, interp_param, ctx->i32_0, "");
LLVMValueRef j = LLVMBuildExtractElement(
- gallivm->builder, interp_param, uint->one, "");
+ gallivm->builder, interp_param, ctx->i32_1, "");
emit_data->output[chan] = ac_build_fs_interp(&ctx->ac,
llvm_chan, attr_number, params,
i, j);
} else {
emit_data->output[chan] = ac_build_fs_interp_mov(&ctx->ac,
- lp_build_const_int32(gallivm, 2), /* P0 */
+ LLVMConstInt(ctx->i32, 2, 0), /* P0 */
llvm_chan, attr_number, params);
}
}
}
+static LLVMValueRef si_emit_ballot(struct si_shader_context *ctx,
+ LLVMValueRef value)
+{
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMValueRef args[3] = {
+ value,
+ ctx->i32_0,
+ LLVMConstInt(ctx->i32, LLVMIntNE, 0)
+ };
+
+ /* We currently have no other way to prevent LLVM from lifting the icmp
+ * calls to a dominating basic block.
+ */
+ emit_optimization_barrier(ctx, &args[0]);
+
+ if (LLVMTypeOf(args[0]) != ctx->i32)
+ args[0] = LLVMBuildBitCast(gallivm->builder, args[0], ctx->i32, "");
+
+ return lp_build_intrinsic(gallivm->builder,
+ "llvm.amdgcn.icmp.i32",
+ ctx->i64, args, 3,
+ LP_FUNC_ATTR_NOUNWIND |
+ LP_FUNC_ATTR_READNONE |
+ LP_FUNC_ATTR_CONVERGENT);
+}
+
+static void vote_all_emit(
+ const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMValueRef active_set, vote_set;
+ LLVMValueRef tmp;
+
+ active_set = si_emit_ballot(ctx, ctx->i32_1);
+ vote_set = si_emit_ballot(ctx, emit_data->args[0]);
+
+ tmp = LLVMBuildICmp(gallivm->builder, LLVMIntEQ, vote_set, active_set, "");
+ emit_data->output[emit_data->chan] =
+ LLVMBuildSExt(gallivm->builder, tmp, ctx->i32, "");
+}
+
+static void vote_any_emit(
+ const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMValueRef vote_set;
+ LLVMValueRef tmp;
+
+ vote_set = si_emit_ballot(ctx, emit_data->args[0]);
+
+ tmp = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
+ vote_set, LLVMConstInt(ctx->i64, 0, 0), "");
+ emit_data->output[emit_data->chan] =
+ LLVMBuildSExt(gallivm->builder, tmp, ctx->i32, "");
+}
+
+static void vote_eq_emit(
+ const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMValueRef active_set, vote_set;
+ LLVMValueRef all, none, tmp;
+
+ active_set = si_emit_ballot(ctx, ctx->i32_1);
+ vote_set = si_emit_ballot(ctx, emit_data->args[0]);
+
+ all = LLVMBuildICmp(gallivm->builder, LLVMIntEQ, vote_set, active_set, "");
+ none = LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
+ vote_set, LLVMConstInt(ctx->i64, 0, 0), "");
+ tmp = LLVMBuildOr(gallivm->builder, all, none, "");
+ emit_data->output[emit_data->chan] =
+ LLVMBuildSExt(gallivm->builder, tmp, ctx->i32, "");
+}
+
+static void ballot_emit(
+ const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMBuilderRef builder = ctx->gallivm.builder;
+ LLVMValueRef tmp;
+
+ tmp = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
+ tmp = si_emit_ballot(ctx, tmp);
+ tmp = LLVMBuildBitCast(builder, tmp, ctx->v2i32, "");
+
+ emit_data->output[0] = LLVMBuildExtractElement(builder, tmp, ctx->i32_0, "");
+ emit_data->output[1] = LLVMBuildExtractElement(builder, tmp, ctx->i32_1, "");
+}
+
+static void read_invoc_fetch_args(
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
+ 0, emit_data->src_chan);
+
+ /* Always read the source invocation (= lane) from the X channel. */
+ emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
+ 1, TGSI_CHAN_X);
+ emit_data->arg_count = 2;
+}
+
+static void read_lane_emit(
+ const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *ctx = si_shader_context(bld_base);
+ LLVMBuilderRef builder = ctx->gallivm.builder;
+
+ /* We currently have no other way to prevent LLVM from lifting the icmp
+ * calls to a dominating basic block.
+ */
+ emit_optimization_barrier(ctx, &emit_data->args[0]);
+
+ for (unsigned i = 0; i < emit_data->arg_count; ++i) {
+ emit_data->args[i] = LLVMBuildBitCast(builder, emit_data->args[i],
+ ctx->i32, "");
+ }
+
+ emit_data->output[emit_data->chan] =
+ ac_build_intrinsic(&ctx->ac, action->intr_name,
+ ctx->i32, emit_data->args, emit_data->arg_count,
+ AC_FUNC_ATTR_READNONE |
+ AC_FUNC_ATTR_CONVERGENT);
+}
+
static unsigned si_llvm_get_stream(struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
struct lp_build_context *uint = &bld_base->uint_bld;
struct si_shader *shader = ctx->shader;
struct tgsi_shader_info *info = &shader->selector->info;
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct lp_build_if_state if_state;
LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
SI_PARAM_GS2VS_OFFSET);
* altogether.
*/
can_emit = LLVMBuildICmp(gallivm->builder, LLVMIntULT, gs_next_vertex,
- lp_build_const_int32(gallivm,
- shader->selector->gs_max_out_vertices), "");
+ LLVMConstInt(ctx->i32,
+ shader->selector->gs_max_out_vertices, 0), "");
bool use_kill = !info->writes_memory;
if (use_kill) {
kill = lp_build_select(&bld_base->base, can_emit,
- lp_build_const_float(gallivm, 1.0f),
- lp_build_const_float(gallivm, -1.0f));
+ LLVMConstReal(ctx->f32, 1.0f),
+ LLVMConstReal(ctx->f32, -1.0f));
- ac_emit_kill(&ctx->ac, kill);
+ ac_build_kill(&ctx->ac, kill);
} else {
lp_build_if(&if_state, gallivm, can_emit);
}
LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
LLVMValueRef voffset =
- lp_build_const_int32(gallivm, offset *
- shader->selector->gs_max_out_vertices);
+ LLVMConstInt(ctx->i32, offset *
+ shader->selector->gs_max_out_vertices, 0);
offset++;
voffset = lp_build_add(uint, voffset, gs_next_vertex);
ctx->gsvs_ring[stream],
out_val, 1,
voffset, soffset, 0,
- 1, 1);
+ 1, 1, true, true);
}
}
gs_next_vertex = lp_build_add(uint, gs_next_vertex,
- lp_build_const_int32(gallivm, 1));
+ ctx->i32_1);
LLVMBuildStore(gallivm->builder, gs_next_vertex, ctx->gs_next_vertex[stream]);
/* Signal vertex emission */
- ac_emit_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
- LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
+ ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (stream << 8),
+ LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
if (!use_kill)
lp_build_endif(&if_state);
}
/* Signal primitive cut */
stream = si_llvm_get_stream(bld_base, emit_data);
- ac_emit_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
- LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
+ ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8),
+ LLVMGetParam(ctx->main_fn, SI_PARAM_GS_WAVE_ID));
}
static void si_llvm_emit_barrier(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 = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
/* SI only (thanks to a hw bug workaround):
* The real barrier instruction isn’t needed, because an entire patch
lp_build_intrinsic(gallivm->builder,
HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.s.barrier"
: "llvm.AMDGPU.barrier.local",
- ctx->voidt, NULL, 0, 0);
+ ctx->voidt, NULL, 0, LP_FUNC_ATTR_CONVERGENT);
}
static const struct lp_build_tgsi_action tex_action = {
if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_BYVAL);
lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_NOALIAS);
- lp_add_attr_dereferenceable(P, UINT64_MAX);
+ ac_add_attr_dereferenceable(P, UINT64_MAX);
} else
lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_INREG);
}
static void declare_tess_lds(struct si_shader_context *ctx)
{
struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
- struct lp_build_context *uint = &bld_base->uint_bld;
unsigned lds_size = ctx->screen->b.chip_class >= CIK ? 65536 : 32768;
- ctx->lds = LLVMBuildIntToPtr(gallivm->builder, uint->zero,
+ ctx->lds = LLVMBuildIntToPtr(gallivm->builder, ctx->i32_0,
LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), LOCAL_ADDR_SPACE),
"tess_lds");
}
static void create_function(struct si_shader_context *ctx)
{
struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
struct si_shader *shader = ctx->shader;
LLVMTypeRef params[SI_NUM_PARAMS + SI_MAX_ATTRIBS], v3i32;
LLVMTypeRef returns[16+32*4];
*/
static void preload_ring_buffers(struct si_shader_context *ctx)
{
- struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
unsigned ring =
ctx->type == PIPE_SHADER_GEOMETRY ? SI_GS_RING_ESGS
: SI_ES_RING_ESGS;
- LLVMValueRef offset = lp_build_const_int32(gallivm, ring);
+ LLVMValueRef offset = LLVMConstInt(ctx->i32, ring, 0);
ctx->esgs_ring =
ac_build_indexed_load_const(&ctx->ac, buf_ptr, offset);
}
if (ctx->shader->is_gs_copy_shader) {
- LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS);
+ LLVMValueRef offset = LLVMConstInt(ctx->i32, SI_RING_GSVS, 0);
ctx->gsvs_ring[0] =
ac_build_indexed_load_const(&ctx->ac, buf_ptr, offset);
} else if (ctx->type == PIPE_SHADER_GEOMETRY) {
const struct si_shader_selector *sel = ctx->shader->selector;
- struct lp_build_context *uint = &ctx->bld_base.uint_bld;
- LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS);
+ LLVMValueRef offset = LLVMConstInt(ctx->i32, SI_RING_GSVS, 0);
LLVMValueRef base_ring;
base_ring = ac_build_indexed_load_const(&ctx->ac, buf_ptr, offset);
num_records = 64;
ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
- tmp = LLVMBuildExtractElement(builder, ring, uint->zero, "");
+ tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_0, "");
tmp = LLVMBuildAdd(builder, tmp,
LLVMConstInt(ctx->i64,
stream_offset, 0), "");
stream_offset += stride * 64;
- ring = LLVMBuildInsertElement(builder, ring, tmp, uint->zero, "");
+ ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_0, "");
ring = LLVMBuildBitCast(builder, ring, ctx->v4i32, "");
- tmp = LLVMBuildExtractElement(builder, ring, uint->one, "");
+ tmp = LLVMBuildExtractElement(builder, ring, ctx->i32_1, "");
tmp = LLVMBuildOr(builder, tmp,
LLVMConstInt(ctx->i32,
S_008F04_STRIDE(stride) |
S_008F04_SWIZZLE_ENABLE(1), 0), "");
- ring = LLVMBuildInsertElement(builder, ring, tmp, uint->one, "");
+ ring = LLVMBuildInsertElement(builder, ring, tmp, ctx->i32_1, "");
ring = LLVMBuildInsertElement(builder, ring,
LLVMConstInt(ctx->i32, num_records, 0),
LLVMConstInt(ctx->i32, 2, 0), "");
LLVMValueRef param_rw_buffers,
unsigned param_pos_fixed_pt)
{
- struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
- struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef slot, desc, offset, row, bit, address[2];
address[1] = unpack_param(ctx, param_pos_fixed_pt, 16, 5);
/* Load the buffer descriptor. */
- slot = lp_build_const_int32(gallivm, SI_PS_CONST_POLY_STIPPLE);
+ slot = LLVMConstInt(ctx->i32, SI_PS_CONST_POLY_STIPPLE, 0);
desc = ac_build_indexed_load_const(&ctx->ac, param_rw_buffers, slot);
/* The stipple pattern is 32x32, each row has 32 bits. */
/* The intrinsic kills the thread if arg < 0. */
bit = LLVMBuildSelect(builder, bit, LLVMConstReal(ctx->f32, 0),
LLVMConstReal(ctx->f32, -1), "");
- ac_emit_kill(&ctx->ac, bit);
+ ac_build_kill(&ctx->ac, bit);
}
void si_shader_binary_read_config(struct ac_shader_binary *binary,
assert((!prolog && !epilog) || !mainb->rodata_size);
assert(!epilog || !epilog->rodata_size);
+ /* GFX9 can fetch at most 128 bytes past the end of the shader.
+ * Prevent VM faults.
+ */
+ if (sscreen->b.chip_class >= GFX9)
+ bo_size += 128;
+
r600_resource_reference(&shader->bo, NULL);
shader->bo = (struct r600_resource*)
pipe_buffer_create(&sscreen->b.b, 0,
if (gs_selector->so.num_outputs)
stream_id = unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
else
- stream_id = uint->zero;
+ stream_id = ctx.i32_0;
/* Fill in output information. */
for (i = 0; i < gsinfo->num_outputs; ++i) {
continue;
bb = LLVMInsertBasicBlockInContext(gallivm->context, end_bb, "out");
- LLVMAddCase(switch_inst, lp_build_const_int32(gallivm, stream), bb);
+ LLVMAddCase(switch_inst, LLVMConstInt(ctx.i32, stream, 0), bb);
LLVMPositionBuilderAtEnd(builder, bb);
/* Fetch vertex data from GSVS ring */
outputs[i].values[chan] =
ac_build_buffer_load(&ctx.ac,
ctx.gsvs_ring[0], 1,
- uint->zero, voffset,
+ ctx.i32_0, voffset,
soffset, 0, 1, 1, true);
}
}
/* Dump LLVM IR before any optimization passes */
if (sscreen->b.debug_flags & DBG_PREOPT_IR &&
r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
- ac_dump_module(bld_base->base.gallivm->module);
+ ac_dump_module(ctx.gallivm.module);
si_llvm_finalize_module(&ctx,
r600_extra_shader_checks(&sscreen->b, PIPE_SHADER_GEOMETRY));
r = si_compile_llvm(sscreen, &ctx.shader->binary,
&ctx.shader->config, ctx.tm,
- bld_base->base.gallivm->module,
+ ctx.gallivm.module,
debug, PIPE_SHADER_GEOMETRY,
"GS Copy Shader");
if (!r) {
bld_base->op_actions[TGSI_OPCODE_INTERP_OFFSET] = interp_action;
bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action;
+ bld_base->op_actions[TGSI_OPCODE_TEX_LZ] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TEX2] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXB] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXB2] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXD] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXF] = tex_action;
+ bld_base->op_actions[TGSI_OPCODE_TXF_LZ] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXL] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXL2] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
bld_base->op_actions[TGSI_OPCODE_MEMBAR].emit = membar_emit;
+ bld_base->op_actions[TGSI_OPCODE_CLOCK].emit = clock_emit;
+
bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy;
bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy;
bld_base->op_actions[TGSI_OPCODE_DDX_FINE].emit = si_llvm_emit_ddxy;
bld_base->op_actions[TGSI_OPCODE_DDY_FINE].emit = si_llvm_emit_ddxy;
+ bld_base->op_actions[TGSI_OPCODE_VOTE_ALL].emit = vote_all_emit;
+ bld_base->op_actions[TGSI_OPCODE_VOTE_ANY].emit = vote_any_emit;
+ bld_base->op_actions[TGSI_OPCODE_VOTE_EQ].emit = vote_eq_emit;
+ bld_base->op_actions[TGSI_OPCODE_BALLOT].emit = ballot_emit;
+ bld_base->op_actions[TGSI_OPCODE_READ_FIRST].intr_name = "llvm.amdgcn.readfirstlane";
+ bld_base->op_actions[TGSI_OPCODE_READ_FIRST].emit = read_lane_emit;
+ bld_base->op_actions[TGSI_OPCODE_READ_INVOC].intr_name = "llvm.amdgcn.readlane";
+ bld_base->op_actions[TGSI_OPCODE_READ_INVOC].fetch_args = read_invoc_fetch_args;
+ bld_base->op_actions[TGSI_OPCODE_READ_INVOC].emit = read_lane_emit;
+
bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive;
bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
}
+#define EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
+#define EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
+
/* Return true if the PARAM export has been eliminated. */
static bool si_eliminate_const_output(struct si_shader_context *ctx,
LLVMValueRef inst, unsigned offset)
for (i = 0; i < 4; i++) {
LLVMBool loses_info;
- LLVMValueRef p = LLVMGetOperand(inst, (HAVE_LLVM >= 0x0500 ? 2 : 5) + i);
+ LLVMValueRef p = LLVMGetOperand(inst, EXP_OUT0 + i);
/* It's a constant expression. Undef outputs are eliminated too. */
if (LLVMIsUndef(p)) {
/* Check if this is an export instruction. */
if ((num_args != 9 && num_args != 8) ||
(strcmp(name, "llvm.SI.export") &&
- strcmp(name, "llvm.amdgcn.exp.")))
+ strcmp(name, "llvm.amdgcn.exp.f32")))
continue;
- LLVMValueRef arg = LLVMGetOperand(cur, HAVE_LLVM >= 0x0500 ? 0 : 3);
+ LLVMValueRef arg = LLVMGetOperand(cur, EXP_TARGET);
unsigned target = LLVMConstIntGetZExtValue(arg);
if (target < V_008DFC_SQ_EXP_PARAM)
if (current_offset[out] != offset)
continue;
- LLVMSetOperand(exports.inst[i], 3,
+ LLVMSetOperand(exports.inst[i], EXP_TARGET,
LLVMConstInt(ctx->i32,
V_008DFC_SQ_EXP_PARAM + new_count, 0));
shader->info.vs_output_param_offset[out] = new_count;
int i;
for (i = 0; i < 4; i++) {
ctx->gs_next_vertex[i] =
- lp_build_alloca(bld_base->base.gallivm,
+ lp_build_alloca(&ctx->gallivm,
ctx->i32, "");
}
}
{
struct si_shader_selector *sel = shader->selector;
struct si_shader_context ctx;
- struct lp_build_tgsi_context *bld_base;
LLVMModuleRef mod;
int r = -1;
shader->info.uses_instanceid = sel->info.uses_instanceid;
- bld_base = &ctx.bld_base;
ctx.load_system_value = declare_system_value;
if (!si_compile_tgsi_main(&ctx, shader)) {
si_build_wrapper_function(&ctx, parts, need_prolog ? 3 : 2, need_prolog ? 1 : 0);
}
- mod = bld_base->base.gallivm->module;
+ mod = ctx.gallivm.module;
/* Dump LLVM IR before any optimization passes */
if (sscreen->b.debug_flags & DBG_PREOPT_IR &&
{
struct si_shader_part *result;
- pipe_mutex_lock(sscreen->shader_parts_mutex);
+ mtx_lock(&sscreen->shader_parts_mutex);
/* Find existing. */
for (result = *list; result; result = result->next) {
if (memcmp(&result->key, key, sizeof(*key)) == 0) {
- pipe_mutex_unlock(sscreen->shader_parts_mutex);
+ mtx_unlock(&sscreen->shader_parts_mutex);
return result;
}
}
out:
si_llvm_dispose(&ctx);
- pipe_mutex_unlock(sscreen->shader_parts_mutex);
+ mtx_unlock(&sscreen->shader_parts_mutex);
return result;
}
args.out[2] = base->undef; /* Z */
args.out[3] = base->undef; /* W */
- ac_emit_export(&ctx->ac, &args);
+ ac_build_export(&ctx->ac, &args);
}
LLVMBuildRetVoid(gallivm->builder);