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,
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;
struct si_shader_context *ctx,
unsigned param_start_instance, unsigned divisor)
{
- struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef result = LLVMGetParam(ctx->main_fn,
ctx->param_instance_id);
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;
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];
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;
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;
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;
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);
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);
}
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,
LLVMConstInt(ctx->i32, swizzle, 0));
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);
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;
}
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, 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);
}
LLVMValueRef face,
LLVMValueRef result[4])
{
- struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
- struct lp_build_context *base = &bld_base->base;
- struct gallivm_state *gallivm = base->gallivm;
+ struct gallivm_state *gallivm = &ctx->gallivm;
LLVMValueRef attr_number;
LLVMValueRef i, j;
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(ctx->main_fn,
break;
case TGSI_SEMANTIC_VERTEXID_NOBASE:
- value = LLVMGetParam(ctx->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:
}
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;
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;
{
struct si_shader_context *ctx = si_shader_context(bld_base);
struct lp_build_context *base = &bld_base->base;
- 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;
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;
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 */
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,
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, "");
}
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;
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);
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;
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;
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, "");
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,
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;
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;
{
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,
/* 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
unsigned src)
{
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;
unsigned target = inst->Memory.Texture;
unsigned num_coords = tgsi_util_get_texture_coord_dim(target);
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;
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;
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;
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;
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;
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];
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;
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;
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;
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;
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) {
{
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 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;
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++) {
LLVMConstInt(ctx->i32, LLVMIntNE, 0)
};
- if (LLVMTypeOf(value) != ctx->i32)
- args[0] = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, "");
+ /* 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",
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);
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
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,
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];
/* 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_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;
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 &&