OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ac/llvm: pass ac_llvm_ptr instead of separate type/ptr

Browse Source 
This simplifies the API.

Reviewed-by: Qiang Yu <yuq825@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Acked-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19035>
This commit is contained in:
Pierre-Eric Pelloux-Prayer
2022-10-13 11:40:52 +02:00
committed by Marge Bot
parent efa5c52822
commit f5abb7ce4c
10 changed files with 138 additions and 175 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
src/amd/llvm/ac_llvm_build.c
View File

@@ -1112,14 +1112,14 @@ LLVMTypeRef ac_build_gep0_type(LLVMTypeRef pointee_type, LLVMValueRef index)
return NULL;
}
LLVMValueRef ac_build_gep0(struct ac_llvm_context *ctx, LLVMTypeRef pointee_type, LLVMValueRef value, LLVMValueRef index)
LLVMValueRef ac_build_gep0(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index)
{
LLVMValueRef indices[2] = {
ctx->i32_0,
index,
};
return LLVMBuildGEP2(ctx->builder, pointee_type, value, indices, 2, "");
return LLVMBuildGEP2(ctx->builder, ptr.t, ptr.v, indices, 2, "");
}
LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef ptr, LLVMValueRef index)
@@ -1127,10 +1127,10 @@ LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMTypeRef type,
return LLVMBuildGEP2(ctx->builder, type, ptr, &index, 1, "");
}
void ac_build_indexed_store(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef base_ptr, LLVMValueRef index,
void ac_build_indexed_store(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index,
LLVMValueRef value)
{
LLVMBuildStore(ctx->builder, value, ac_build_gep0(ctx, type, base_ptr, index));
LLVMBuildStore(ctx->builder, value, ac_build_gep0(ctx, ptr, index));
}
/**
@@ -1182,30 +1182,29 @@ static LLVMValueRef ac_build_load_custom(struct ac_llvm_context *ctx, LLVMTypeRe
return result;
}
LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef base_ptr, LLVMValueRef index)
LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index)
{
return ac_build_load_custom(ctx, type, base_ptr, index, false, false, false);
return ac_build_load_custom(ctx, ptr.t, ptr.v, index, false, false, false);
}
LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef base_ptr,
LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr,
LLVMValueRef index)
{
return ac_build_load_custom(ctx, type, base_ptr, index, false, true, false);
return ac_build_load_custom(ctx, ptr.t, ptr.v, index, false, true, false);
}
/* This assumes that there is no unsigned integer wraparound during the address
* computation, excluding all GEPs within base_ptr. */
LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef base_ptr,
LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr,
LLVMValueRef index)
{
return ac_build_load_custom(ctx, type, base_ptr, index, true, true, true);
return ac_build_load_custom(ctx, ptr.t, ptr.v, index, true, true, true);
}
/* See ac_build_load_custom() documentation. */
LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx, LLVMTypeRef type,
LLVMValueRef base_ptr, LLVMValueRef index)
LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index)
{
return ac_build_load_custom(ctx, type, base_ptr, index, true, true, false);
return ac_build_load_custom(ctx, ptr.t, ptr.v, index, true, true, false);
}
static unsigned get_load_cache_policy(struct ac_llvm_context *ctx, unsigned cache_policy)
@@ -2783,14 +2782,14 @@ void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx, LLVMValueRef dw_addr)
{
LLVMValueRef v = ac_build_gep0(ctx, ctx->lds.t, ctx->lds.v, dw_addr);
LLVMValueRef v = ac_build_gep0(ctx, ctx->lds, dw_addr);
return LLVMBuildLoad2(ctx->builder, ctx->i32, v, "");
}
void ac_lds_store(struct ac_llvm_context *ctx, LLVMValueRef dw_addr, LLVMValueRef value)
{
value = ac_to_integer(ctx, value);
ac_build_indexed_store(ctx, ctx->lds.t, ctx->lds.v, dw_addr, value);
ac_build_indexed_store(ctx, ctx->lds, dw_addr, value);
}
LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx, LLVMTypeRef dst_type, LLVMValueRef src0)

14
src/amd/llvm/ac_llvm_build.h
View File

@@ -258,19 +258,19 @@ LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMTypeRef type,
LLVMValueRef index);
LLVMTypeRef ac_build_gep0_type(LLVMTypeRef pointee_type, LLVMValueRef index);
LLVMValueRef ac_build_gep0(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef value, LLVMValueRef index);
LLVMValueRef ac_build_gep0(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index);
void ac_build_indexed_store(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef base_ptr, LLVMValueRef index,
void ac_build_indexed_store(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index,
LLVMValueRef value);
LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef base_ptr, LLVMValueRef index);
LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef base_ptr,
LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr, LLVMValueRef index);
LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr,
LLVMValueRef index);
LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx, LLVMTypeRef type, LLVMValueRef base_ptr,
LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr,
LLVMValueRef index);
LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx, LLVMTypeRef type,
LLVMValueRef base_ptr, LLVMValueRef index);
LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx, struct ac_llvm_pointer ptr,
LLVMValueRef index);
void ac_build_buffer_store_dword(struct ac_llvm_context *ctx, LLVMValueRef rsrc, LLVMValueRef vdata,
LLVMValueRef vindex, LLVMValueRef voffset, LLVMValueRef soffset,

14
src/amd/llvm/ac_nir_to_llvm.c
View File

@@ -1667,8 +1667,8 @@ static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx, nir_int
}
}
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, ctx->args, ctx->args->push_constants);
ptr = LLVMBuildGEP2(ctx->ac.builder, type, ac_get_arg(&ctx->ac, ctx->args->push_constants), &addr, 1, "");
struct ac_llvm_pointer pc = ac_get_ptr_arg(&ctx->ac, ctx->args, ctx->args->push_constants);
ptr = LLVMBuildGEP2(ctx->ac.builder, pc.t, pc.v, &addr, 1, "");
if (instr->dest.ssa.bit_size == 8) {
unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
@@ -3717,8 +3717,8 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
if (ctx->abi->load_grid_size_from_user_sgpr) {
result = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
} else {
LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
result = ac_build_load_invariant(&ctx->ac, ctx->ac.v3i32, ptr, ctx->ac.i32_0);
result = ac_build_load_invariant(&ctx->ac,
ac_get_ptr_arg(&ctx->ac, ctx->args, ctx->args->num_work_groups), ctx->ac.i32_0);
}
if (nir_dest_bit_size(instr->dest) == 64)
result = LLVMBuildZExt(ctx->ac.builder, result, LLVMVectorType(ctx->ac.i64, 3), "");
@@ -4115,7 +4115,7 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
break;
case nir_intrinsic_load_scratch: {
LLVMValueRef offset = get_src(ctx, instr->src[0]);
LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch.t, ctx->scratch.v, offset);
LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch, offset);
LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
LLVMTypeRef vec_type = instr->dest.ssa.num_components == 1
? comp_type
@@ -4125,7 +4125,7 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
}
case nir_intrinsic_store_scratch: {
LLVMValueRef offset = get_src(ctx, instr->src[1]);
LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch.t, ctx->scratch.v, offset);
LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch, offset);
LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
LLVMValueRef src = get_src(ctx, instr->src[0]);
unsigned wrmask = nir_intrinsic_write_mask(instr);
@@ -4154,7 +4154,7 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
offset = LLVMBuildSelect(ctx->ac.builder, cond, offset, size, "");
LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->constant_data.t, ctx->constant_data.v, offset);
LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->constant_data, offset);
LLVMTypeRef comp_type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
LLVMTypeRef vec_type = instr->dest.ssa.num_components == 1
? comp_type

28
src/amd/vulkan/radv_nir_to_llvm.c
View File

@@ -108,13 +108,12 @@ load_descriptor_sets(struct radv_shader_context *ctx)
uint32_t mask = ctx->shader_info->desc_set_used_mask;
if (user_sgprs_locs->shader_data[AC_UD_INDIRECT_DESCRIPTOR_SETS].sgpr_idx != -1) {
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args->ac, ctx->args->descriptor_sets[0]);
LLVMValueRef desc_sets = ac_get_arg(&ctx->ac, ctx->args->descriptor_sets[0]);
struct ac_llvm_pointer desc_sets = ac_get_ptr_arg(&ctx->ac, &ctx->args->ac, ctx->args->descriptor_sets[0]);
while (mask) {
int i = u_bit_scan(&mask);
ctx->descriptor_sets[i] =
ac_build_load_to_sgpr(&ctx->ac, type, desc_sets, LLVMConstInt(ctx->ac.i32, i, false));
ac_build_load_to_sgpr(&ctx->ac, desc_sets, LLVMConstInt(ctx->ac.i32, i, false));
LLVMSetAlignment(ctx->descriptor_sets[i], 4);
}
} else {
@@ -350,8 +349,7 @@ static void
load_vs_input(struct radv_shader_context *ctx, unsigned driver_location, LLVMTypeRef dest_type,
LLVMValueRef out[4])
{
LLVMTypeRef t_list_type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args->ac, ctx->args->ac.vertex_buffers);
LLVMValueRef t_list_ptr = ac_get_arg(&ctx->ac, ctx->args->ac.vertex_buffers);
struct ac_llvm_pointer t_list_ptr = ac_get_ptr_arg(&ctx->ac, &ctx->args->ac, ctx->args->ac.vertex_buffers);
LLVMValueRef t_offset;
LLVMValueRef t_list;
LLVMValueRef input;
@@ -402,7 +400,7 @@ load_vs_input(struct radv_shader_context *ctx, unsigned driver_location, LLVMTyp
desc_index = util_bitcount(ctx->shader_info->vs.vb_desc_usage_mask &
u_bit_consecutive(0, desc_index));
t_offset = LLVMConstInt(ctx->ac.i32, desc_index, false);
t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_type, t_list_ptr, t_offset);
t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
/* Always split typed vertex buffer loads on GFX6 and GFX10+ to avoid any alignment issues that
* triggers memory violations and eventually a GPU hang. This can happen if the stride (static or
@@ -800,8 +798,7 @@ radv_emit_streamout(struct radv_shader_context *ctx, unsigned stream)
*/
LLVMValueRef so_write_offset[4] = {0};
LLVMValueRef so_buffers[4] = {0};
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args->ac, ctx->args->streamout_buffers);
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->args->streamout_buffers);
struct ac_llvm_pointer buf_ptr = ac_get_ptr_arg(&ctx->ac, &ctx->args->ac, ctx->args->streamout_buffers);
for (i = 0; i < 4; i++) {
uint16_t stride = ctx->shader_info->so.strides[i];
@@ -811,7 +808,7 @@ radv_emit_streamout(struct radv_shader_context *ctx, unsigned stream)
LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i, false);
so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac, type, buf_ptr, offset);
so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
LLVMValueRef so_offset = ac_get_arg(&ctx->ac, ctx->args->ac.streamout_offset[i]);
@@ -1178,6 +1175,8 @@ radv_llvm_visit_export_vertex(struct ac_shader_abi *abi)
static void
ac_setup_rings(struct radv_shader_context *ctx)
{
struct ac_llvm_pointer ring_offsets = { .t = ctx->ac.i8, .v = ctx->ring_offsets };
if (ctx->options->gfx_level <= GFX8 &&
(ctx->stage == MESA_SHADER_GEOMETRY ||
(ctx->stage == MESA_SHADER_VERTEX && ctx->shader_info->vs.as_es) ||
@@ -1185,12 +1184,11 @@ ac_setup_rings(struct radv_shader_context *ctx)
unsigned ring = ctx->stage == MESA_SHADER_GEOMETRY ? RING_ESGS_GS : RING_ESGS_VS;
LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, ring, false);
ctx->esgs_ring = ac_build_load_to_sgpr(&ctx->ac, ctx->ac.i8, ctx->ring_offsets, offset);
ctx->esgs_ring = ac_build_load_to_sgpr(&ctx->ac, ring_offsets, offset);
}
if (ctx->args->is_gs_copy_shader) {
ctx->gsvs_ring[0] = ac_build_load_to_sgpr(&ctx->ac, ctx->ac.i8, ctx->ring_offsets,
LLVMConstInt(ctx->ac.i32, RING_GSVS_VS, false));
ctx->gsvs_ring[0] = ac_build_load_to_sgpr(&ctx->ac, ring_offsets, LLVMConstInt(ctx->ac.i32, RING_GSVS_VS, false));
}
if (ctx->stage == MESA_SHADER_GEOMETRY) {
@@ -1207,7 +1205,7 @@ ac_setup_rings(struct radv_shader_context *ctx)
unsigned num_records = ctx->ac.wave_size;
LLVMValueRef base_ring;
base_ring = ac_build_load_to_sgpr(&ctx->ac, ctx->ac.i8, ctx->ring_offsets,
base_ring = ac_build_load_to_sgpr(&ctx->ac, ring_offsets,
LLVMConstInt(ctx->ac.i32, RING_GSVS_GS, false));
for (unsigned stream = 0; stream < 4; stream++) {
@@ -1248,9 +1246,9 @@ ac_setup_rings(struct radv_shader_context *ctx)
if (ctx->stage == MESA_SHADER_TESS_CTRL || ctx->stage == MESA_SHADER_TESS_EVAL) {
ctx->hs_ring_tess_offchip = ac_build_load_to_sgpr(
&ctx->ac, ctx->ac.i8, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_HS_TESS_OFFCHIP, false));
&ctx->ac, ring_offsets, LLVMConstInt(ctx->ac.i32, RING_HS_TESS_OFFCHIP, false));
ctx->hs_ring_tess_factor = ac_build_load_to_sgpr(
&ctx->ac, ctx->ac.i8, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_HS_TESS_FACTOR, false));
&ctx->ac, ring_offsets, LLVMConstInt(ctx->ac.i32, RING_HS_TESS_FACTOR, false));
}
}

92
src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
View File

@@ -65,18 +65,13 @@ static LLVMValueRef ngg_get_ordered_id(struct si_shader_context *ctx)
static LLVMValueRef ngg_get_query_buf(struct si_shader_context *ctx)
{
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
return ac_build_load_to_sgpr(&ctx->ac, type, buf_ptr,
return ac_build_load_to_sgpr(&ctx->ac, ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings),
LLVMConstInt(ctx->ac.i32, SI_GS_QUERY_BUF, false));
}
static LLVMValueRef ngg_get_emulated_counters_buf(struct si_shader_context *ctx)
{
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
return ac_build_load_to_sgpr(&ctx->ac, type, buf_ptr,
return ac_build_load_to_sgpr(&ctx->ac, ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings),
LLVMConstInt(ctx->ac.i32, SI_GS_QUERY_EMULATED_COUNTERS_BUF, false));
}
@@ -253,7 +248,7 @@ static void build_streamout_vertex(struct si_shader_context *ctx, LLVMValueRef *
for (unsigned comp = 0; comp < 4; comp++) {
LLVMValueRef idx = LLVMConstInt(ctx->ac.i32, 4 * reg + comp, false);
LLVMValueRef v = ac_build_gep0(&ctx->ac, vertexptr.t, vertexptr.v, idx);
LLVMValueRef v = ac_build_gep0(&ctx->ac, vertexptr, idx);
out.values[comp] = LLVMBuildLoad2(builder, ac_build_gep0_type(vertexptr.t, idx), v, "");
out.vertex_streams = info->output_streams[reg];
}
@@ -287,8 +282,7 @@ static void build_streamout(struct si_shader_context *ctx, struct ngg_streamout
struct si_shader_info *info = &ctx->shader->selector->info;
struct pipe_stream_output_info *so = &ctx->so;
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
struct ac_llvm_pointer arg = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings);
LLVMValueRef tid = gfx10_get_thread_id_in_tg(ctx);
LLVMValueRef tmp, tmp2;
LLVMValueRef i32_2 = LLVMConstInt(ctx->ac.i32, 2, false);
@@ -328,7 +322,7 @@ static void build_streamout(struct si_shader_context *ctx, struct ngg_streamout
LLVMConstInt(ctx->ac.i32, buffer, false));
so_buffer[buffer] = ac_build_load_to_sgpr(
&ctx->ac, type, buf_ptr, LLVMConstInt(ctx->ac.i32, SI_VS_STREAMOUT_BUF0 + buffer, false));
&ctx->ac, arg, LLVMConstInt(ctx->ac.i32, SI_VS_STREAMOUT_BUF0 + buffer, false));
}
tmp = LLVMBuildICmp(builder, LLVMIntEQ, get_wave_id_in_tg(ctx), ctx->ac.i32_0, "");
@@ -345,7 +339,7 @@ static void build_streamout(struct si_shader_context *ctx, struct ngg_streamout
ac_build_ifcc(&ctx->ac, tmp, 5210);
{
if (isgs) {
LLVMValueRef vt = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch.t, ctx->gs_ngg_scratch.v, tid);
LLVMValueRef vt = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid);
tmp = LLVMBuildLoad2(builder, ac_build_gep0_type(ctx->gs_ngg_scratch.t, tid), vt, "");
} else {
tmp = ac_build_writelane(&ctx->ac, ctx->ac.i32_0, ngg_get_prim_cnt(ctx), ctx->ac.i32_0);
@@ -452,7 +446,7 @@ static void build_streamout(struct si_shader_context *ctx, struct ngg_streamout
LLVMBuildStore(builder, tmp, offsets_vgpr);
tmp2 = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac), scratch_offset_basev, "");
tmp2 = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch.t, ctx->gs_ngg_scratch.v, tmp2);
tmp2 = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp2);
LLVMBuildStore(builder, tmp, tmp2);
}
ac_build_endif(&ctx->ac, 5210);
@@ -546,7 +540,7 @@ static void build_streamout(struct si_shader_context *ctx, struct ngg_streamout
ac_build_ifcc(&ctx->ac, tmp, 5225);
{
tmp = LLVMBuildAdd(builder, ac_get_thread_id(&ctx->ac), scratch_emit_basev, "");
tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch.t, ctx->gs_ngg_scratch.v, tmp);
tmp = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tmp);
LLVMBuildStore(builder, emit_vgpr, tmp);
}
ac_build_endif(&ctx->ac, 5225);
@@ -567,7 +561,7 @@ static void build_streamout(struct si_shader_context *ctx, struct ngg_streamout
primemit_scan[stream].op = nir_op_iadd;
primemit_scan[stream].src = nggso->prim_enable[stream];
primemit_scan[stream].scratch = ac_build_gep0(
&ctx->ac, ctx->gs_ngg_scratch.t, ctx->gs_ngg_scratch.v,
&ctx->ac, ctx->gs_ngg_scratch,
LLVMConstInt(ctx->ac.i32, 12 + 8 * stream, false));
primemit_scan[stream].waveidx = get_wave_id_in_tg(ctx);
primemit_scan[stream].numwaves = get_tgsize(ctx);
@@ -592,7 +586,7 @@ static void build_streamout(struct si_shader_context *ctx, struct ngg_streamout
LLVMValueRef scratch_vgpr;
LLVMValueRef idx = ac_get_thread_id(&ctx->ac);
LLVMValueRef v = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch.t, ctx->gs_ngg_scratch.v, idx);
LLVMValueRef v = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, idx);
scratch_vgpr = LLVMBuildLoad2(builder, ac_build_gep0_type(ctx->gs_ngg_scratch.t, idx), v, "");
for (unsigned buffer = 0; buffer < 4; ++buffer) {
@@ -779,7 +773,7 @@ static void load_vertex_counts(struct si_shader_context *ctx, struct ac_llvm_poi
*/
ac_build_ifcc(&ctx->ac, LLVMBuildICmp(builder, LLVMIntULT, tid,
LLVMConstInt(ctx->ac.i32, num_i8vec4, 0), ""), 17771);
LLVMValueRef v = ac_build_gep0(&ctx->ac, lds.t, lds.v, tid);
LLVMValueRef v = ac_build_gep0(&ctx->ac, lds, tid);
LLVMBuildStore(builder, LLVMBuildLoad2(builder, ac_build_gep0_type(lds.t, tid), v, ""), i8vec4_lane);
ac_build_endif(&ctx->ac, 17771);
@@ -953,10 +947,9 @@ static void cull_primitive(struct si_shader_context *ctx,
/* Load the viewport state for small prim culling. */
bool prim_is_lines = shader->key.ge.opt.ngg_culling & SI_NGG_CULL_LINES;
LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->small_prim_cull_info);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->small_prim_cull_info);
struct ac_llvm_pointer small_prim_cull_info_arg = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->small_prim_cull_info);
/* Lines will always use the non-AA viewport transformation. */
LLVMValueRef vp = ac_build_load_to_sgpr(&ctx->ac, type, ptr,
LLVMValueRef vp = ac_build_load_to_sgpr(&ctx->ac, small_prim_cull_info_arg,
prim_is_lines ? ctx->ac.i32_1 : ctx->ac.i32_0);
vp = LLVMBuildBitCast(builder, vp, ctx->ac.v4f32, "");
vp_scale[0] = ac_llvm_extract_elem(&ctx->ac, vp, 0);
@@ -970,7 +963,8 @@ static void cull_primitive(struct si_shader_context *ctx,
options.cull_w = true;
if (prim_is_lines) {
LLVMValueRef terms = ac_build_load_to_sgpr(&ctx->ac, ctx->ac.v2f32, ptr, LLVMConstInt(ctx->ac.i32, 4, 0));
small_prim_cull_info_arg.t = ctx->ac.v2f32;
LLVMValueRef terms = ac_build_load_to_sgpr(&ctx->ac, small_prim_cull_info_arg, LLVMConstInt(ctx->ac.i32, 4, 0));
terms = LLVMBuildBitCast(builder, terms, ctx->ac.v2f32, "");
clip_half_line_width[0] = ac_llvm_extract_elem(&ctx->ac, terms, 0);
clip_half_line_width[1] = ac_llvm_extract_elem(&ctx->ac, terms, 1);
@@ -1062,14 +1056,14 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
/* Store Position.W into LDS. */
LLVMBuildStore(
builder, ac_to_integer(&ctx->ac, position[3]),
ac_build_gep0(&ctx->ac, es_vtxptr.t, es_vtxptr.v, LLVMConstInt(ctx->ac.i32, lds_pos_cull_w, 0)));
ac_build_gep0(&ctx->ac, es_vtxptr, LLVMConstInt(ctx->ac.i32, lds_pos_cull_w, 0)));
/* Store Position.XY / W into LDS. */
for (unsigned chan = 0; chan < 2; chan++) {
LLVMValueRef val = ac_build_fdiv(&ctx->ac, position[chan], position[3]);
LLVMBuildStore(
builder, ac_to_integer(&ctx->ac, val),
ac_build_gep0(&ctx->ac, es_vtxptr.t, es_vtxptr.v, LLVMConstInt(ctx->ac.i32, lds_pos_cull_x_div_w + chan, 0)));
ac_build_gep0(&ctx->ac, es_vtxptr, LLVMConstInt(ctx->ac.i32, lds_pos_cull_x_div_w + chan, 0)));
}
break;
@@ -1111,7 +1105,7 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
/* Initialize the packed data. */
LLVMBuildStore(
builder, packed_data,
ac_build_gep0(&ctx->ac, es_vtxptr.t, es_vtxptr.v, LLVMConstInt(ctx->ac.i32, lds_packed_data, 0)));
ac_build_gep0(&ctx->ac, es_vtxptr, LLVMConstInt(ctx->ac.i32, lds_packed_data, 0)));
ac_build_endif(&ctx->ac, ctx->merged_wrap_if_label);
ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM);
@@ -1208,7 +1202,7 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
continue;
LLVMValueRef idx = LLVMConstInt(ctx->ac.i32, index, 0);
LLVMValueRef v = ac_build_gep0(&ctx->ac, gs_vtxptr[vtx].t, gs_vtxptr[vtx].v, idx);
LLVMValueRef v = ac_build_gep0(&ctx->ac, gs_vtxptr[vtx], idx);
pos[vtx][chan] = LLVMBuildLoad2(builder, ac_build_gep0_type(gs_vtxptr[vtx].t, idx), v, "");
pos[vtx][chan] = ac_to_float(&ctx->ac, pos[vtx][chan]);
}
@@ -1305,7 +1299,7 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
LLVMBuildStore(
builder, ac_to_integer(&ctx->ac,
LLVMBuildLoad2(builder, ctx->ac.f32, addrs[4 * pos_index + chan], "")),
ac_build_gep0(&ctx->ac, new_vtx.t, new_vtx.v, LLVMConstInt(ctx->ac.i32, lds_pos_x + chan, 0)));
ac_build_gep0(&ctx->ac, new_vtx, LLVMConstInt(ctx->ac.i32, lds_pos_x + chan, 0)));
}
/* Store VertexID and InstanceID into LDS. ES threads will have to load them
@@ -1315,24 +1309,24 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
if (ctx->stage == MESA_SHADER_VERTEX) {
LLVMBuildStore(
builder, ctx->abi.vertex_id,
ac_build_gep0(&ctx->ac, new_vtx.t, new_vtx.v, LLVMConstInt(ctx->ac.i32, lds_vertex_id, 0)));
ac_build_gep0(&ctx->ac, new_vtx, LLVMConstInt(ctx->ac.i32, lds_vertex_id, 0)));
if (uses_instance_id) {
LLVMBuildStore(
builder, ctx->abi.instance_id,
ac_build_gep0(&ctx->ac, new_vtx.t, new_vtx.v, LLVMConstInt(ctx->ac.i32, lds_instance_id, 0)));
ac_build_gep0(&ctx->ac, new_vtx, LLVMConstInt(ctx->ac.i32, lds_instance_id, 0)));
}
} else {
assert(ctx->stage == MESA_SHADER_TESS_EVAL);
LLVMBuildStore(builder, ac_to_integer(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args.tes_u)),
ac_build_gep0(&ctx->ac, new_vtx.t, new_vtx.v, LLVMConstInt(ctx->ac.i32, lds_tes_u, 0)));
ac_build_gep0(&ctx->ac, new_vtx, LLVMConstInt(ctx->ac.i32, lds_tes_u, 0)));
LLVMBuildStore(builder, ac_to_integer(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args.tes_v)),
ac_build_gep0(&ctx->ac, new_vtx.t, new_vtx.v, LLVMConstInt(ctx->ac.i32, lds_tes_v, 0)));
ac_build_gep0(&ctx->ac, new_vtx, LLVMConstInt(ctx->ac.i32, lds_tes_v, 0)));
LLVMBuildStore(builder, LLVMBuildTrunc(builder, ac_get_arg(&ctx->ac, ctx->args.tes_rel_patch_id), ctx->ac.i8, ""),
si_build_gep_i8(ctx, new_vtx.value, lds_byte2_tes_rel_patch_id));
if (uses_tes_prim_id) {
LLVMBuildStore(
builder, ac_get_arg(&ctx->ac, ctx->args.tes_patch_id),
ac_build_gep0(&ctx->ac, new_vtx.t, new_vtx.v, LLVMConstInt(ctx->ac.i32, lds_tes_patch_id, 0)));
ac_build_gep0(&ctx->ac, new_vtx, LLVMConstInt(ctx->ac.i32, lds_tes_patch_id, 0)));
}
}
}
@@ -1411,13 +1405,13 @@ void gfx10_ngg_culling_build_end(struct si_shader_context *ctx)
/* Prepare LDS addresses of the new ES input VGPRs. */
LLVMValueRef input_vgpr_addresses[4] = {
ac_build_gep0(&ctx->ac, es_vtxptr.t, es_vtxptr.v, LLVMConstInt(ctx->ac.i32, lds_vertex_id, 0)),
ac_build_gep0(&ctx->ac, es_vtxptr.t, es_vtxptr.v, LLVMConstInt(ctx->ac.i32, lds_instance_id, 0)),
ac_build_gep0(&ctx->ac, es_vtxptr, LLVMConstInt(ctx->ac.i32, lds_vertex_id, 0)),
ac_build_gep0(&ctx->ac, es_vtxptr, LLVMConstInt(ctx->ac.i32, lds_instance_id, 0)),
};
if (ctx->stage == MESA_SHADER_TESS_EVAL) {
input_vgpr_addresses[2] = si_build_gep_i8(ctx, es_vtxptr.v, lds_byte2_tes_rel_patch_id);
if (uses_tes_prim_id) {
input_vgpr_addresses[3] = ac_build_gep0(&ctx->ac, es_vtxptr.t, es_vtxptr.v,
input_vgpr_addresses[3] = ac_build_gep0(&ctx->ac, es_vtxptr,
LLVMConstInt(ctx->ac.i32, lds_tes_patch_id, 0));
}
}
@@ -1545,7 +1539,7 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
*/
if (ctx->so.num_outputs) {
LLVMValueRef idx = LLVMConstInt(ctx->ac.i32, 4 * i + j, false);
tmp = ac_build_gep0(&ctx->ac, vertex_ptr.t, vertex_ptr.v, idx);
tmp = ac_build_gep0(&ctx->ac, vertex_ptr, idx);
tmp2 = LLVMBuildLoad2(builder, ac_build_gep0_type(vertex_ptr.t, idx), addrs[4 * i + j], "");
LLVMTypeRef type = ac_to_integer_type(&ctx->ac, ctx->ac.f32);
tmp2 = LLVMBuildBitCast(ctx->ac.builder, tmp2, type, "");
@@ -1561,7 +1555,7 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
edgeflag = ac_build_umin(&ctx->ac, edgeflag, ctx->ac.i32_1);
tmp = LLVMConstInt(ctx->ac.i32, ngg_nogs_vertex_size(ctx->shader) - 1, 0);
tmp = ac_build_gep0(&ctx->ac, vertex_ptr.t, vertex_ptr.v, tmp);
tmp = ac_build_gep0(&ctx->ac, vertex_ptr, tmp);
LLVMBuildStore(builder, edgeflag, tmp);
}
}
@@ -1624,7 +1618,7 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
struct ac_llvm_pointer vt = ngg_nogs_vertex_ptr(ctx, vtxindex[i]);
tmp2 = LLVMConstInt(ctx->ac.i32, ngg_nogs_vertex_size(ctx->shader) - 1, 0);
tmp = LLVMBuildLoad2(builder, ac_build_gep0_type(vt.t, tmp2),
ac_build_gep0(&ctx->ac, vt.t, vt.v, tmp2), "");
ac_build_gep0(&ctx->ac, vt, tmp2), "");
tmp = LLVMBuildTrunc(builder, tmp, ctx->ac.i1, "");
user_edgeflags[i] = ac_build_alloca_init(&ctx->ac, tmp, "");
@@ -1655,7 +1649,7 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
struct ac_llvm_pointer vertex_ptr = ngg_nogs_vertex_ptr(ctx, provoking_vtx_index);
LLVMBuildStore(builder, ac_get_arg(&ctx->ac, ctx->args.gs_prim_id),
ac_build_gep0(&ctx->ac, vertex_ptr.t, vertex_ptr.v, ctx->ac.i32_0));
ac_build_gep0(&ctx->ac, vertex_ptr, ctx->ac.i32_0));
ac_build_endif(&ctx->ac, 5400);
}
@@ -1719,7 +1713,7 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
for (unsigned j = 0; j < 4; j++) {
tmp = LLVMConstInt(ctx->ac.i32, lds_pos_x + j, 0);
LLVMValueRef v = ac_build_gep0(&ctx->ac, vertex_ptr.t, vertex_ptr.v, tmp);
LLVMValueRef v = ac_build_gep0(&ctx->ac, vertex_ptr, tmp);
tmp = LLVMBuildLoad2(builder, ac_build_gep0_type(vertex_ptr.t, tmp), v, "");
outputs[i].values[j] = LLVMBuildBitCast(ctx->ac.builder, tmp,
ac_to_float_type(&ctx->ac, ctx->ac.f32), "");
@@ -1742,7 +1736,7 @@ void gfx10_ngg_build_end(struct si_shader_context *ctx)
struct ac_llvm_pointer vt = ngg_nogs_vertex_ptr(ctx, gfx10_get_thread_id_in_tg(ctx));
outputs[i].values[0] = LLVMBuildLoad2(
builder, ac_build_gep0_type(vt.t, ctx->ac.i32_0), ac_build_gep0(&ctx->ac, vt.t, vt.v, ctx->ac.i32_0), "");
builder, ac_build_gep0_type(vt.t, ctx->ac.i32_0), ac_build_gep0(&ctx->ac, vt, ctx->ac.i32_0), "");
} else {
assert(ctx->stage == MESA_SHADER_TESS_EVAL);
outputs[i].values[0] = si_get_primitive_id(ctx, 0);
@@ -1821,7 +1815,7 @@ static struct ac_llvm_pointer ngg_gs_vertex_ptr(struct si_shader_context *ctx, L
}
return (struct ac_llvm_pointer) {
.value = ac_build_gep0(&ctx->ac, storage.t, storage.v, vertexidx),
.value = ac_build_gep0(&ctx->ac, storage, vertexidx),
.pointee_type = ac_build_gep0_type(storage.t, vertexidx)
};
}
@@ -1948,7 +1942,7 @@ void gfx10_ngg_gs_emit_begin(struct si_shader_context *ctx)
tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, LLVMConstInt(ctx->ac.i32, 4, false), "");
ac_build_ifcc(&ctx->ac, tmp, 5090);
{
LLVMValueRef ptr = ac_build_gep0(&ctx->ac, scratchptr.t, scratchptr.v, tid);
LLVMValueRef ptr = ac_build_gep0(&ctx->ac, scratchptr, tid);
LLVMBuildStore(builder, ctx->ac.i32_0, ptr);
}
ac_build_endif(&ctx->ac, 5090);
@@ -2033,7 +2027,7 @@ void gfx10_ngg_gs_build_end(struct si_shader_context *ctx)
{
LLVMBuildAtomicRMW(
builder, LLVMAtomicRMWBinOpAdd,
ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch.t, ctx->gs_ngg_scratch.v, LLVMConstInt(ctx->ac.i32, stream, false)),
ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, LLVMConstInt(ctx->ac.i32, stream, false)),
numprims, LLVMAtomicOrderingMonotonic, false);
}
ac_build_endif(&ctx->ac, 5105);
@@ -2069,7 +2063,7 @@ void gfx10_ngg_gs_build_end(struct si_shader_context *ctx)
"");
struct ac_llvm_pointer vt = ngg_gs_vertex_ptr(ctx, tmp);
nggso.vertices[i].t = ac_build_gep0_type(vt.t, ctx->ac.i32_0);
nggso.vertices[i].v = ac_build_gep0(&ctx->ac, vt.t, vt.v, ctx->ac.i32_0);
nggso.vertices[i].v = ac_build_gep0(&ctx->ac, vt, ctx->ac.i32_0);
}
build_streamout(ctx, &nggso);
@@ -2096,9 +2090,7 @@ void gfx10_ngg_gs_build_end(struct si_shader_context *ctx)
offset = LLVMBuildAdd(builder, offset, tmp, "");
}
tmp = LLVMBuildLoad2(builder, ctx->ac.i32,
ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch.t, ctx->gs_ngg_scratch.v,
tid), "");
tmp = LLVMBuildLoad2(builder, ctx->ac.i32, ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, tid), "");
LLVMValueRef args[] = {
tmp, ngg_get_query_buf(ctx),
offset, LLVMConstInt(ctx->ac.i32, 16, false), /* soffset */
@@ -2136,7 +2128,7 @@ void gfx10_ngg_gs_build_end(struct si_shader_context *ctx)
tmp = LLVMBuildSub(builder, tid, LLVMConstInt(ctx->ac.i32, verts_per_prim - i - 1, false), "");
struct ac_llvm_pointer vt = ngg_gs_vertex_ptr(ctx, tmp);
vtxptr[i].t = ac_build_gep0_type(vt.t, ctx->ac.i32_0);
vtxptr[i].v = ac_build_gep0(&ctx->ac, vt.t, vt.v, ctx->ac.i32_0);
vtxptr[i].v = ac_build_gep0(&ctx->ac, vt, ctx->ac.i32_0);
}
for (unsigned i = 0; i < info->num_outputs; i++) {
@@ -2157,7 +2149,7 @@ void gfx10_ngg_gs_build_end(struct si_shader_context *ctx)
continue;
LLVMValueRef idx = LLVMConstInt(ctx->ac.i32, 4 * i + comp, false);
tmp = ac_build_gep0(&ctx->ac, vtxptr[vert].t, vtxptr[vert].v, idx);
tmp = ac_build_gep0(&ctx->ac, vtxptr[vert], idx);
pos[vert][comp] = LLVMBuildLoad2(builder,
ac_build_gep0_type(vtxptr[vert].t, idx),
tmp, "");
@@ -2229,7 +2221,7 @@ void gfx10_ngg_gs_build_end(struct si_shader_context *ctx)
vertlive_scan.enable_reduce = true;
vertlive_scan.enable_exclusive = true;
vertlive_scan.src = vertlive;
vertlive_scan.scratch = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch.t, ctx->gs_ngg_scratch.v, ctx->ac.i32_0);
vertlive_scan.scratch = ac_build_gep0(&ctx->ac, ctx->gs_ngg_scratch, ctx->ac.i32_0);
vertlive_scan.waveidx = get_wave_id_in_tg(ctx);
vertlive_scan.numwaves = get_tgsize(ctx);
vertlive_scan.maxwaves = DIV_ROUND_UP(256, ctx->ac.wave_size);

27
src/gallium/drivers/radeonsi/si_shader_llvm.c
View File

@@ -765,9 +765,7 @@ static LLVMValueRef si_llvm_load_intrinsic(struct ac_shader_abi *abi, nir_intrin
case nir_intrinsic_load_tess_level_outer_default:
case nir_intrinsic_load_tess_level_inner_default: {
LLVMValueRef slot = LLVMConstInt(ctx->ac.i32, SI_HS_CONST_DEFAULT_TESS_LEVELS, 0);
LLVMValueRef buf = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
buf = ac_build_load_to_sgpr(&ctx->ac, type, buf, slot);
LLVMValueRef buf = ac_build_load_to_sgpr(&ctx->ac, ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings), slot);
int offset = op == nir_intrinsic_load_tess_level_inner_default ? 4 : 0;
LLVMValueRef val[4];
@@ -816,15 +814,15 @@ static LLVMValueRef si_llvm_load_intrinsic(struct ac_shader_abi *abi, nir_intrin
case nir_intrinsic_load_clip_half_line_width_amd: {
LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->small_prim_cull_info);
return ac_build_load_to_sgpr(&ctx->ac, ctx->ac.v2f32, ptr, LLVMConstInt(ctx->ac.i32, 4, 0));
return ac_build_load_to_sgpr(&ctx->ac,
(struct ac_llvm_pointer) { .t = ctx->ac.v2f32, .v = ptr }, LLVMConstInt(ctx->ac.i32, 4, 0));
}
case nir_intrinsic_load_viewport_xy_scale_and_offset: {
bool prim_is_lines = ctx->shader->key.ge.opt.ngg_culling & SI_NGG_CULL_LINES;
LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->small_prim_cull_info);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->small_prim_cull_info);
struct ac_llvm_pointer ptr = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->small_prim_cull_info);
LLVMValueRef terms =
ac_build_load_to_sgpr(&ctx->ac, type, ptr, prim_is_lines ? ctx->ac.i32_1 : ctx->ac.i32_0);
ac_build_load_to_sgpr(&ctx->ac, ptr, prim_is_lines ? ctx->ac.i32_1 : ctx->ac.i32_0);
return LLVMBuildBitCast(ctx->ac.builder, terms, ctx->ac.v4f32, "");
}
@@ -878,10 +876,9 @@ static LLVMValueRef si_llvm_load_intrinsic(struct ac_shader_abi *abi, nir_intrin
static LLVMValueRef si_llvm_load_user_clip_plane(struct ac_shader_abi *abi, unsigned ucp_id)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
struct ac_llvm_pointer ptr = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings);
LLVMValueRef constbuf_index = LLVMConstInt(ctx->ac.i32, SI_VS_CONST_CLIP_PLANES, 0);
LLVMValueRef const_resource = ac_build_load_to_sgpr(&ctx->ac, type, ptr, constbuf_index);
LLVMValueRef const_resource = ac_build_load_to_sgpr(&ctx->ac, ptr, constbuf_index);
LLVMValueRef addr = LLVMConstInt(ctx->ac.i32, ucp_id * 16, 0);
return ac_build_buffer_load(&ctx->ac, const_resource, 4, NULL, addr, NULL,
ctx->ac.f32, 0, true, true);
@@ -890,11 +887,10 @@ static LLVMValueRef si_llvm_load_user_clip_plane(struct ac_shader_abi *abi, unsi
static LLVMValueRef si_llvm_load_streamout_buffer(struct ac_shader_abi *abi, unsigned buffer)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
struct ac_llvm_pointer buf_ptr = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings);
return ac_build_load_to_sgpr(
&ctx->ac, type, buf_ptr, LLVMConstInt(ctx->ac.i32, SI_VS_STREAMOUT_BUF0 + buffer, false));
&ctx->ac, buf_ptr, LLVMConstInt(ctx->ac.i32, SI_VS_STREAMOUT_BUF0 + buffer, false));
}
bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shader,
@@ -930,11 +926,10 @@ bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shad
/* preload instance_divisor_constbuf to be used for input load after culling */
if (ctx->shader->key.ge.opt.ngg_culling &&
ctx->shader->key.ge.part.vs.prolog.instance_divisor_is_fetched) {
LLVMValueRef buf = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
struct ac_llvm_pointer buf = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings);
ctx->instance_divisor_constbuf =
ac_build_load_to_sgpr(
&ctx->ac, type, buf, LLVMConstInt(ctx->ac.i32, SI_VS_CONST_INSTANCE_DIVISORS, 0));
&ctx->ac, buf, LLVMConstInt(ctx->ac.i32, SI_VS_CONST_INSTANCE_DIVISORS, 0));
}
break;

20
src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
View File

@@ -98,10 +98,7 @@ static LLVMValueRef si_get_gs_wave_id(struct si_shader_context *ctx)
static LLVMValueRef ngg_get_emulated_counters_buf(struct si_shader_context *ctx)
{
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
return ac_build_load_to_sgpr(&ctx->ac, type, buf_ptr,
return ac_build_load_to_sgpr(&ctx->ac, ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings),
LLVMConstInt(ctx->ac.i32, SI_GS_QUERY_EMULATED_COUNTERS_BUF, false));
}
@@ -263,10 +260,9 @@ void si_preload_esgs_ring(struct si_shader_context *ctx)
if (ctx->screen->info.gfx_level <= GFX8) {
LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, SI_RING_ESGS, 0);
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
ctx->esgs_ring = ac_build_load_to_sgpr(&ctx->ac, type, buf_ptr, offset);
ctx->esgs_ring = ac_build_load_to_sgpr(&ctx->ac,
ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings), offset);
if (ctx->stage != MESA_SHADER_GEOMETRY) {
LLVMValueRef desc1 = LLVMBuildExtractElement(builder, ctx->esgs_ring, ctx->ac.i32_1, "");
@@ -307,9 +303,8 @@ void si_preload_gs_rings(struct si_shader_context *ctx)
const struct si_shader_selector *sel = ctx->shader->selector;
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, SI_RING_GSVS, 0);
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
LLVMValueRef base_ring = ac_build_load_to_sgpr(&ctx->ac, type, buf_ptr, offset);
LLVMValueRef base_ring = ac_build_load_to_sgpr(&ctx->ac,
ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings), offset);
/* The conceptual layout of the GSVS ring is
* v0c0 .. vLv0 v0c1 .. vLc1 ..
@@ -435,10 +430,9 @@ struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
/* Build the main function. */
si_llvm_create_main_func(&ctx, false);
LLVMValueRef buf_ptr = ac_get_arg(&ctx.ac, ctx.internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx.ac, &ctx.args, ctx.internal_bindings);
ctx.gsvs_ring[0] =
ac_build_load_to_sgpr(&ctx.ac, type, buf_ptr, LLVMConstInt(ctx.ac.i32, SI_RING_GSVS, 0));
ac_build_load_to_sgpr(&ctx.ac,
ac_get_ptr_arg(&ctx.ac, &ctx.args, ctx.internal_bindings), LLVMConstInt(ctx.ac.i32, SI_RING_GSVS, 0));
LLVMValueRef voffset =
LLVMBuildMul(ctx.ac.builder, ctx.abi.vertex_id, LLVMConstInt(ctx.ac.i32, 4, 0), "");

15
src/gallium/drivers/radeonsi/si_shader_llvm_ps.c
View File

@@ -34,10 +34,9 @@ LLVMValueRef si_get_sample_id(struct si_shader_context *ctx)
static LLVMValueRef load_sample_position(struct ac_shader_abi *abi, LLVMValueRef sample_id)
{
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
LLVMValueRef desc = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMValueRef buf_index = LLVMConstInt(ctx->ac.i32, SI_PS_CONST_SAMPLE_POSITIONS, 0);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
LLVMValueRef resource = ac_build_load_to_sgpr(&ctx->ac, type, desc, buf_index);
LLVMValueRef resource = ac_build_load_to_sgpr(
&ctx->ac, ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings), buf_index);
/* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
LLVMValueRef offset0 =
@@ -69,8 +68,9 @@ static LLVMValueRef si_nir_emit_fbfetch(struct ac_shader_abi *abi)
ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
ptr =
LLVMBuildPointerCast(ctx->ac.builder, ptr, ac_array_in_const32_addr_space(ctx->ac.v8i32), "");
image =
ac_build_load_to_sgpr(&ctx->ac, ctx->ac.v8i32, ptr, LLVMConstInt(ctx->ac.i32, SI_PS_IMAGE_COLORBUF0 / 2, 0));
struct ac_llvm_pointer desc = { .v = ptr, .t = ctx->ac.v8i32 };
image = ac_build_load_to_sgpr(&ctx->ac, desc, LLVMConstInt(ctx->ac.i32, SI_PS_IMAGE_COLORBUF0 / 2, 0));
unsigned chan = 0;
@@ -90,8 +90,7 @@ static LLVMValueRef si_nir_emit_fbfetch(struct ac_shader_abi *abi)
ctx->shader->key.ps.mono.fbfetch_msaa &&
!(ctx->screen->debug_flags & DBG(NO_FMASK))) {
fmask = ac_build_load_to_sgpr(&ctx->ac, ctx->ac.v8i32, ptr,
LLVMConstInt(ctx->ac.i32, SI_PS_IMAGE_COLORBUF0_FMASK / 2, 0));
fmask = ac_build_load_to_sgpr(&ctx->ac, desc, LLVMConstInt(ctx->ac.i32, SI_PS_IMAGE_COLORBUF0_FMASK / 2, 0));
ac_apply_fmask_to_sample(&ctx->ac, fmask, args.coords,
ctx->shader->key.ps.mono.fbfetch_layered);
@@ -550,7 +549,7 @@ static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
/* Load the buffer descriptor. */
slot = LLVMConstInt(ctx->ac.i32, SI_PS_CONST_POLY_STIPPLE, 0);
desc = ac_build_load_to_sgpr(&ctx->ac, ctx->ac.v4i32, param_internal_bindings, slot);
desc = ac_build_load_to_sgpr(&ctx->ac, (struct ac_llvm_pointer) { .t = ctx->ac.v4i32, .v = param_internal_bindings }, slot);
/* The stipple pattern is 32x32, each row has 32 bits. */
offset = LLVMBuildMul(builder, address[1], LLVMConstInt(ctx->ac.i32, 4, 0), "");

47
src/gallium/drivers/radeonsi/si_shader_llvm_resources.c
View File

@@ -94,8 +94,6 @@ static LLVMValueRef load_ubo(struct ac_shader_abi *abi, LLVMValueRef index)
struct si_shader_context *ctx = si_shader_context_from_abi(abi);
struct si_shader_selector *sel = ctx->shader->selector;
LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->const_and_shader_buffers);
if (sel->info.base.num_ubos == 1 && sel->info.base.num_ssbos == 0) {
return load_const_buffer_desc_fast_path(ctx);
}
@@ -105,8 +103,7 @@ static LLVMValueRef load_ubo(struct ac_shader_abi *abi, LLVMValueRef index)
LLVMBuildAdd(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i32, SI_NUM_SHADER_BUFFERS, 0), "");
return ac_build_load_to_sgpr(&ctx->ac,
ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->const_and_shader_buffers),
ptr,
ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->const_and_shader_buffers),
index);
}
@@ -119,14 +116,13 @@ static LLVMValueRef load_ssbo(struct ac_shader_abi *abi, LLVMValueRef index, boo
LLVMConstIntGetZExtValue(index) < ctx->shader->selector->cs_num_shaderbufs_in_user_sgprs)
return ac_get_arg(&ctx->ac, ctx->cs_shaderbuf[LLVMConstIntGetZExtValue(index)]);
LLVMValueRef rsrc_ptr = ac_get_arg(&ctx->ac, ctx->const_and_shader_buffers);
index = si_llvm_bound_index(ctx, index, ctx->num_shader_buffers);
index = LLVMBuildSub(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, SI_NUM_SHADER_BUFFERS - 1, 0),
index, "");
return ac_build_load_to_sgpr(&ctx->ac,
ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->const_and_shader_buffers),
rsrc_ptr, index);
ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->const_and_shader_buffers),
index);
}
/**
@@ -181,25 +177,23 @@ static LLVMValueRef fixup_image_desc(struct si_shader_context *ctx, LLVMValueRef
/* AC_DESC_FMASK is handled exactly like AC_DESC_IMAGE. The caller should
* adjust "index" to point to FMASK. */
static LLVMValueRef si_load_image_desc(struct si_shader_context *ctx, LLVMTypeRef type, LLVMValueRef list,
static LLVMValueRef si_load_image_desc(struct si_shader_context *ctx, struct ac_llvm_pointer list,
LLVMValueRef index, enum ac_descriptor_type desc_type,
bool uses_store, bool bindless)
{
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef rsrc;
if (desc_type == AC_DESC_BUFFER) {
index = ac_build_imad(&ctx->ac, index, LLVMConstInt(ctx->ac.i32, 2, 0), ctx->ac.i32_1);
list = LLVMBuildPointerCast(builder, list, ac_array_in_const32_addr_space(ctx->ac.v4i32), "");
type = ctx->ac.v4i32;
list.pointee_type = ctx->ac.v4i32;
} else {
assert(desc_type == AC_DESC_IMAGE || desc_type == AC_DESC_FMASK);
}
if (bindless)
rsrc = ac_build_load_to_sgpr_uint_wraparound(&ctx->ac, type, list, index);
rsrc = ac_build_load_to_sgpr_uint_wraparound(&ctx->ac, list, index);
else
rsrc = ac_build_load_to_sgpr(&ctx->ac, type, list, index);
rsrc = ac_build_load_to_sgpr(&ctx->ac, list, index);
if (desc_type == AC_DESC_IMAGE)
rsrc = fixup_image_desc(ctx, rsrc, uses_store);
@@ -210,8 +204,7 @@ static LLVMValueRef si_load_image_desc(struct si_shader_context *ctx, LLVMTypeRe
/**
* Load an image view, fmask view. or sampler state descriptor.
*/
static LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx, LLVMTypeRef list_type,
LLVMValueRef list,
static LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx, struct ac_llvm_pointer list,
LLVMValueRef index, enum ac_descriptor_type type)
{
LLVMBuilderRef builder = ctx->ac.builder;
@@ -224,8 +217,7 @@ static LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx, LLVMType
case AC_DESC_BUFFER:
/* The buffer is in [4:7]. */
index = ac_build_imad(&ctx->ac, index, LLVMConstInt(ctx->ac.i32, 4, 0), ctx->ac.i32_1);
list = LLVMBuildPointerCast(builder, list, ac_array_in_const32_addr_space(ctx->ac.v4i32), "");
list_type = ctx->ac.v4i32;
list.pointee_type = ctx->ac.v4i32;
break;
case AC_DESC_FMASK:
/* The FMASK is at [8:15]. */
@@ -236,8 +228,7 @@ static LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx, LLVMType
/* The sampler state is at [12:15]. */
index = ac_build_imad(&ctx->ac, index, LLVMConstInt(ctx->ac.i32, 4, 0),
LLVMConstInt(ctx->ac.i32, 3, 0));
list = LLVMBuildPointerCast(builder, list, ac_array_in_const32_addr_space(ctx->ac.v4i32), "");
list_type = ctx->ac.v4i32;
list.pointee_type = ctx->ac.v4i32;
break;
case AC_DESC_PLANE_0:
case AC_DESC_PLANE_1:
@@ -248,7 +239,7 @@ static LLVMValueRef si_load_sampler_desc(struct si_shader_context *ctx, LLVMType
unreachable("Plane descriptor requested in radeonsi.");
}
return ac_build_load_to_sgpr(&ctx->ac, list_type, list, index);
return ac_build_load_to_sgpr(&ctx->ac, list, index);
}
static LLVMValueRef si_nir_load_sampler_desc(struct ac_shader_abi *abi, unsigned descriptor_set,
@@ -265,8 +256,7 @@ static LLVMValueRef si_nir_load_sampler_desc(struct ac_shader_abi *abi, unsigned
assert(desc_type <= AC_DESC_BUFFER);
if (bindless) {
LLVMValueRef list = ac_get_arg(&ctx->ac, ctx->bindless_samplers_and_images);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->bindless_samplers_and_images);
struct ac_llvm_pointer list = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->bindless_samplers_and_images);
/* dynamic_index is the bindless handle */
if (image) {
@@ -278,7 +268,7 @@ static LLVMValueRef si_nir_load_sampler_desc(struct ac_shader_abi *abi, unsigned
dynamic_index = LLVMBuildAdd(ctx->ac.builder, dynamic_index, ctx->ac.i32_1, "");
}
return si_load_image_desc(ctx, type, list, dynamic_index, desc_type, write, true);
return si_load_image_desc(ctx, list, dynamic_index, desc_type, write, true);
}
/* Since bindless handle arithmetic can contain an unsigned integer
@@ -288,8 +278,8 @@ static LLVMValueRef si_nir_load_sampler_desc(struct ac_shader_abi *abi, unsigned
*/
dynamic_index =
LLVMBuildMul(ctx->ac.builder, dynamic_index, LLVMConstInt(ctx->ac.i64, 2, 0), "");
list = ac_build_pointer_add(&ctx->ac, ctx->ac.v8i32, list, dynamic_index);
return si_load_sampler_desc(ctx, type, list, ctx->ac.i32_0, desc_type);
list.v = ac_build_pointer_add(&ctx->ac, ctx->ac.v8i32, list.v, dynamic_index);
return si_load_sampler_desc(ctx, list, ctx->ac.i32_0, desc_type);
}
unsigned num_slots = image ? ctx->num_images : ctx->num_samplers;
@@ -298,8 +288,7 @@ static LLVMValueRef si_nir_load_sampler_desc(struct ac_shader_abi *abi, unsigned
if (const_index >= num_slots)
const_index = base_index;
LLVMValueRef list = ac_get_arg(&ctx->ac, ctx->samplers_and_images);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->samplers_and_images);
struct ac_llvm_pointer list = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->samplers_and_images);
LLVMValueRef index = LLVMConstInt(ctx->ac.i32, const_index, false);
if (dynamic_index) {
@@ -336,12 +325,12 @@ static LLVMValueRef si_nir_load_sampler_desc(struct ac_shader_abi *abi, unsigned
}
index = LLVMBuildSub(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, SI_NUM_IMAGE_SLOTS - 1, 0),
index, "");
return si_load_image_desc(ctx, type, list, index, desc_type, write, false);
return si_load_image_desc(ctx, list, index, desc_type, write, false);
}
index = LLVMBuildAdd(ctx->ac.builder, index,
LLVMConstInt(ctx->ac.i32, SI_NUM_IMAGE_SLOTS / 2, 0), "");
return si_load_sampler_desc(ctx, type, list, index, desc_type);
return si_load_sampler_desc(ctx, list, index, desc_type);
}
void si_llvm_init_resource_callbacks(struct si_shader_context *ctx)

27
src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
View File

@@ -168,9 +168,9 @@ static void load_input_vs(struct si_shader_context *ctx, unsigned input_index, L
vb_desc = ac_get_arg(&ctx->ac, ctx->vb_descriptors[input_index]);
} else {
unsigned index = input_index - num_vbos_in_user_sgprs;
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->args.vertex_buffers);
vb_desc = ac_build_load_to_sgpr(&ctx->ac, type, ac_get_arg(&ctx->ac, ctx->args.vertex_buffers),
LLVMConstInt(ctx->ac.i32, index, 0));
vb_desc = ac_build_load_to_sgpr(
&ctx->ac, ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->args.vertex_buffers),
LLVMConstInt(ctx->ac.i32, index, 0));
}
if (ctx->abi.vertex_id_replaced) {
@@ -409,8 +409,7 @@ void si_llvm_emit_streamout(struct si_shader_context *ctx, struct si_shader_outp
* enabled buffer. */
LLVMValueRef so_write_offset[4] = {};
LLVMValueRef so_buffers[4];
LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
struct ac_llvm_pointer arg = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings);
for (i = 0; i < 4; i++) {
if (!so->stride[i])
@@ -418,7 +417,7 @@ void si_llvm_emit_streamout(struct si_shader_context *ctx, struct si_shader_outp
LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, SI_VS_STREAMOUT_BUF0 + i, 0);
so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac, type, buf_ptr, offset);
so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac, arg, offset);
LLVMValueRef so_offset = ac_get_arg(&ctx->ac, ctx->args.streamout_offset[i]);
so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->ac.i32, 4, 0), "");
@@ -451,10 +450,9 @@ void si_llvm_clipvertex_to_clipdist(struct si_shader_context *ctx,
unsigned chan;
unsigned const_chan;
LLVMValueRef base_elt;
LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->internal_bindings);
LLVMTypeRef type = ac_get_arg_pointee_type(&ctx->ac, &ctx->args, ctx->internal_bindings);
LLVMValueRef constbuf_index = LLVMConstInt(ctx->ac.i32, SI_VS_CONST_CLIP_PLANES, 0);
LLVMValueRef const_resource = ac_build_load_to_sgpr(&ctx->ac, type, ptr, constbuf_index);
LLVMValueRef const_resource = ac_build_load_to_sgpr(
&ctx->ac, ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings), constbuf_index);
unsigned clipdist_mask = ctx->shader->selector->info.clipdist_mask &
~ctx->shader->key.ge.opt.kill_clip_distances;
@@ -815,11 +813,9 @@ void si_llvm_build_vs_exports(struct si_shader_context *ctx, LLVMValueRef num_ex
/* Get the attribute ring address and descriptor. */
LLVMValueRef attr_address;
if (ctx->stage == MESA_SHADER_VERTEX && shader->selector->info.base.vs.blit_sgprs_amd) {
LLVMValueRef ptr =
LLVMBuildPointerCast(ctx->ac.builder,
ac_get_arg(&ctx->ac, ctx->internal_bindings),
LLVMPointerType(ctx->ac.i32, AC_ADDR_SPACE_CONST_32BIT), "");
attr_address = ac_build_load_to_sgpr(&ctx->ac, ctx->ac.i32, ptr,
struct ac_llvm_pointer ring_ptr = ac_get_ptr_arg(&ctx->ac, &ctx->args, ctx->internal_bindings);
ring_ptr.pointee_type = ctx->ac.i32;
attr_address = ac_build_load_to_sgpr(&ctx->ac, ring_ptr,
LLVMConstInt(ctx->ac.i32, SI_GS_ATTRIBUTE_RING * 4, 0));
} else {
attr_address = ac_get_arg(&ctx->ac, ctx->gs_attr_address);
@@ -1039,7 +1035,8 @@ void si_llvm_build_vs_prolog(struct si_shader_context *ctx, union si_shader_part
if (key->vs_prolog.states.instance_divisor_is_fetched) {
LLVMValueRef list = si_prolog_get_internal_bindings(ctx);
LLVMValueRef buf_index = LLVMConstInt(ctx->ac.i32, SI_VS_CONST_INSTANCE_DIVISORS, 0);
instance_divisor_constbuf = ac_build_load_to_sgpr(&ctx->ac, ctx->ac.v4i32, list, buf_index);
instance_divisor_constbuf = ac_build_load_to_sgpr(&ctx->ac,
(struct ac_llvm_pointer) { .v = list, .t = ctx->ac.v4i32 }, buf_index);
}
for (i = 0; i < key->vs_prolog.num_inputs; i++) {