LLVMConstVector(swizzles, 4), "");
}
+static struct tgsi_declaration_range
+get_array_range(struct lp_build_tgsi_context *bld_base,
+ unsigned File, const struct tgsi_ind_register *reg)
+{
+ struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
+
+ if (File != TGSI_FILE_TEMPORARY || reg->ArrayID == 0 ||
+ reg->ArrayID > bld_base->info->array_max[TGSI_FILE_TEMPORARY]) {
+ struct tgsi_declaration_range range;
+ range.First = 0;
+ range.Last = bld_base->info->file_max[File];
+ return range;
+ }
+
+ return ctx->arrays[reg->ArrayID - 1];
+}
+
static LLVMValueRef
emit_array_index(
struct lp_build_tgsi_soa_context *bld,
- const struct tgsi_full_src_register *reg,
- unsigned swizzle)
+ const struct tgsi_ind_register *reg,
+ unsigned offset)
{
struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
- LLVMValueRef addr = LLVMBuildLoad(gallivm->builder,
- bld->addr[reg->Indirect.Index][swizzle], "");
- LLVMValueRef offset = lp_build_const_int32(gallivm, reg->Register.Index);
- LLVMValueRef hw_index = LLVMBuildAdd(gallivm->builder, addr, offset, "");
- LLVMValueRef soa_index = LLVMBuildMul(gallivm->builder, hw_index,
- lp_build_const_int32(gallivm, 4), "");
- LLVMValueRef array_index = LLVMBuildAdd(gallivm->builder, soa_index,
- lp_build_const_int32(gallivm, swizzle), "");
+ LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, bld->addr[reg->Index][reg->Swizzle], "");
+ return LLVMBuildAdd(gallivm->builder, addr, lp_build_const_int32(gallivm, offset), "");
+}
- return array_index;
+LLVMValueRef
+radeon_llvm_emit_fetch_double(
+ struct lp_build_tgsi_context *bld_base,
+ LLVMValueRef ptr,
+ LLVMValueRef ptr2)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMValueRef result;
+
+ result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
+
+ result = LLVMBuildInsertElement(builder,
+ result,
+ bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr),
+ bld_base->int_bld.zero, "");
+ result = LLVMBuildInsertElement(builder,
+ result,
+ bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr2),
+ bld_base->int_bld.one, "");
+ return bitcast(bld_base, TGSI_TYPE_DOUBLE, result);
}
static LLVMValueRef
-emit_fetch_immediate(
+emit_array_fetch(
struct lp_build_tgsi_context *bld_base,
- const struct tgsi_full_src_register *reg,
- enum tgsi_opcode_type type,
+ unsigned File, enum tgsi_opcode_type type,
+ struct tgsi_declaration_range range,
unsigned swizzle)
{
- LLVMTypeRef ctype;
- LLVMContextRef ctx = bld_base->base.gallivm->context;
+ struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
+ struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- switch (type) {
- case TGSI_TYPE_UNSIGNED:
- case TGSI_TYPE_SIGNED:
- ctype = LLVMInt32TypeInContext(ctx);
- break;
- case TGSI_TYPE_UNTYPED:
- case TGSI_TYPE_FLOAT:
- ctype = LLVMFloatTypeInContext(ctx);
- break;
- default:
- ctype = 0;
- break;
- }
+ unsigned i, size = range.Last - range.First + 1;
+ LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
+ LLVMValueRef result = LLVMGetUndef(vec);
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
- if (swizzle == ~0) {
- LLVMValueRef values[TGSI_NUM_CHANNELS] = {};
- unsigned chan;
- for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- values[chan] = LLVMConstBitCast(bld->immediates[reg->Register.Index][chan], ctype);
- }
- return lp_build_gather_values(bld_base->base.gallivm, values,
- TGSI_NUM_CHANNELS);
- } else {
- return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
+ struct tgsi_full_src_register tmp_reg = {};
+ tmp_reg.Register.File = File;
+
+ for (i = 0; i < size; ++i) {
+ tmp_reg.Register.Index = i + range.First;
+ LLVMValueRef temp = radeon_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
+ result = LLVMBuildInsertElement(builder, result, temp,
+ lp_build_const_int32(gallivm, i), "");
}
+ return result;
}
-static LLVMValueRef
-emit_fetch_input(
- struct lp_build_tgsi_context *bld_base,
- const struct tgsi_full_src_register *reg,
- enum tgsi_opcode_type type,
- unsigned swizzle)
+static bool uses_temp_indirect_addressing(
+ struct lp_build_tgsi_context *bld_base)
{
- struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
- if (swizzle == ~0) {
- LLVMValueRef values[TGSI_NUM_CHANNELS] = {};
- unsigned chan;
- for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- values[chan] = ctx->inputs[radeon_llvm_reg_index_soa(
- reg->Register.Index, chan)];
- }
- return lp_build_gather_values(bld_base->base.gallivm, values,
- TGSI_NUM_CHANNELS);
- } else {
- return bitcast(bld_base, type, ctx->inputs[radeon_llvm_reg_index_soa(reg->Register.Index, swizzle)]);
- }
+ struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
+ return (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY));
}
-static LLVMValueRef
-emit_fetch_temporary(
- struct lp_build_tgsi_context *bld_base,
- const struct tgsi_full_src_register *reg,
- enum tgsi_opcode_type type,
- unsigned swizzle)
+LLVMValueRef radeon_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
+ const struct tgsi_full_src_register *reg,
+ enum tgsi_opcode_type type,
+ unsigned swizzle)
{
+ struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMValueRef result = NULL, ptr, ptr2;
+
if (swizzle == ~0) {
- LLVMValueRef values[TGSI_NUM_CHANNELS] = {};
+ LLVMValueRef values[TGSI_NUM_CHANNELS];
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- values[chan] = emit_fetch_temporary(bld_base, reg, type, chan);
+ values[chan] = radeon_llvm_emit_fetch(bld_base, reg, type, chan);
}
return lp_build_gather_values(bld_base->base.gallivm, values,
- TGSI_NUM_CHANNELS);
+ TGSI_NUM_CHANNELS);
}
if (reg->Register.Indirect) {
- LLVMValueRef array_index = emit_array_index(bld, reg, swizzle);
- LLVMValueRef ptr = LLVMBuildGEP(builder, bld->temps_array, &array_index,
- 1, "");
- return LLVMBuildLoad(builder, ptr, "");
- } else {
- LLVMValueRef temp_ptr;
- temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
- return bitcast(bld_base,type,LLVMBuildLoad(builder, temp_ptr, ""));
+ struct tgsi_declaration_range range = get_array_range(bld_base,
+ reg->Register.File, ®->Indirect);
+ return LLVMBuildExtractElement(builder,
+ emit_array_fetch(bld_base, reg->Register.File, type, range, swizzle),
+ emit_array_index(bld, ®->Indirect, reg->Register.Index - range.First),
+ "");
+ }
+
+ switch(reg->Register.File) {
+ case TGSI_FILE_IMMEDIATE: {
+ LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
+ if (type == TGSI_TYPE_DOUBLE) {
+ result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
+ result = LLVMConstInsertElement(result,
+ bld->immediates[reg->Register.Index][swizzle],
+ bld_base->int_bld.zero);
+ result = LLVMConstInsertElement(result,
+ bld->immediates[reg->Register.Index][swizzle + 1],
+ bld_base->int_bld.one);
+ return LLVMConstBitCast(result, ctype);
+ } else {
+ return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
+ }
+ }
+
+ case TGSI_FILE_INPUT:
+ result = ctx->inputs[radeon_llvm_reg_index_soa(reg->Register.Index, swizzle)];
+ if (type == TGSI_TYPE_DOUBLE) {
+ ptr = result;
+ ptr2 = ctx->inputs[radeon_llvm_reg_index_soa(reg->Register.Index, swizzle + 1)];
+ return radeon_llvm_emit_fetch_double(bld_base, ptr, ptr2);
+ }
+ break;
+
+ case TGSI_FILE_TEMPORARY:
+ if (reg->Register.Index >= ctx->temps_count)
+ return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
+ if (uses_temp_indirect_addressing(bld_base)) {
+ ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
+ break;
+ }
+ ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
+ if (type == TGSI_TYPE_DOUBLE) {
+ ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1];
+ return radeon_llvm_emit_fetch_double(bld_base,
+ LLVMBuildLoad(builder, ptr, ""),
+ LLVMBuildLoad(builder, ptr2, ""));
+ }
+ result = LLVMBuildLoad(builder, ptr, "");
+ break;
+
+ case TGSI_FILE_OUTPUT:
+ ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
+ if (type == TGSI_TYPE_DOUBLE) {
+ ptr2 = lp_get_output_ptr(bld, reg->Register.Index, swizzle + 1);
+ return radeon_llvm_emit_fetch_double(bld_base,
+ LLVMBuildLoad(builder, ptr, ""),
+ LLVMBuildLoad(builder, ptr2, ""));
+ }
+ result = LLVMBuildLoad(builder, ptr, "");
+ break;
+
+ default:
+ return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
}
+
+ return bitcast(bld_base, type, result);
}
-static LLVMValueRef
-emit_fetch_output(
- struct lp_build_tgsi_context *bld_base,
+static LLVMValueRef fetch_system_value(
+ struct lp_build_tgsi_context * bld_base,
const struct tgsi_full_src_register *reg,
enum tgsi_opcode_type type,
unsigned swizzle)
{
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
- LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- if (reg->Register.Indirect) {
- LLVMValueRef array_index = emit_array_index(bld, reg, swizzle);
- LLVMValueRef ptr = LLVMBuildGEP(builder, bld->outputs_array, &array_index,
- 1, "");
- return LLVMBuildLoad(builder, ptr, "");
- } else {
- LLVMValueRef temp_ptr;
- temp_ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
- return LLVMBuildLoad(builder, temp_ptr, "");
- }
+ struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+
+ LLVMValueRef cval = ctx->system_values[reg->Register.Index];
+ if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
+ cval = LLVMBuildExtractElement(gallivm->builder, cval,
+ lp_build_const_int32(gallivm, swizzle), "");
+ }
+ return bitcast(bld_base, type, cval);
+}
+
+static LLVMValueRef si_build_alloca_undef(struct gallivm_state *gallivm,
+ LLVMTypeRef type,
+ const char *name)
+{
+ LLVMValueRef ptr = lp_build_alloca(gallivm, type, name);
+ LLVMBuildStore(gallivm->builder, LLVMGetUndef(type), ptr);
+ return ptr;
}
static void emit_declaration(
const struct tgsi_full_declaration *decl)
{
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
+ unsigned first, last, i, idx;
switch(decl->Declaration.File) {
case TGSI_FILE_ADDRESS:
{
for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- ctx->soa.addr[idx][chan] = lp_build_alloca(
+ ctx->soa.addr[idx][chan] = si_build_alloca_undef(
&ctx->gallivm,
ctx->soa.bld_base.uint_bld.elem_type, "");
}
}
case TGSI_FILE_TEMPORARY:
- lp_emit_declaration_soa(bld_base, decl);
+ if (decl->Declaration.Array) {
+ if (!ctx->arrays) {
+ int size = bld_base->info->array_max[TGSI_FILE_TEMPORARY];
+ ctx->arrays = MALLOC(sizeof(ctx->arrays[0]) * size);
+ }
+
+ ctx->arrays[decl->Array.ArrayID - 1] = decl->Range;
+ }
+ if (uses_temp_indirect_addressing(bld_base)) {
+ lp_emit_declaration_soa(bld_base, decl);
+ break;
+ }
+ first = decl->Range.First;
+ last = decl->Range.Last;
+ if (!ctx->temps_count) {
+ ctx->temps_count = bld_base->info->file_max[TGSI_FILE_TEMPORARY] + 1;
+ ctx->temps = MALLOC(TGSI_NUM_CHANNELS * ctx->temps_count * sizeof(LLVMValueRef));
+ }
+ for (idx = first; idx <= last; idx++) {
+ for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
+ ctx->temps[idx * TGSI_NUM_CHANNELS + i] =
+ si_build_alloca_undef(bld_base->base.gallivm,
+ bld_base->base.vec_type,
+ "temp");
+ }
+ }
break;
case TGSI_FILE_INPUT:
{
unsigned idx;
for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
- ctx->load_input(ctx, idx, decl);
+ if (ctx->load_input)
+ ctx->load_input(ctx, idx, decl);
}
}
break;
unsigned chan;
assert(idx < RADEON_LLVM_MAX_OUTPUTS);
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- ctx->soa.outputs[idx][chan] = lp_build_alloca(&ctx->gallivm,
+ ctx->soa.outputs[idx][chan] = si_build_alloca_undef(
+ &ctx->gallivm,
ctx->soa.bld_base.base.elem_type, "");
}
}
}
}
-static void
-emit_store(
+LLVMValueRef radeon_llvm_saturate(struct lp_build_tgsi_context *bld_base,
+ LLVMValueRef value)
+{
+ struct lp_build_emit_data clamp_emit_data;
+
+ memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
+ clamp_emit_data.arg_count = 3;
+ clamp_emit_data.args[0] = value;
+ clamp_emit_data.args[2] = bld_base->base.one;
+ clamp_emit_data.args[1] = bld_base->base.zero;
+
+ return lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
+ &clamp_emit_data);
+}
+
+void radeon_llvm_emit_store(
struct lp_build_tgsi_context * bld_base,
const struct tgsi_full_instruction * inst,
const struct tgsi_opcode_info * info,
LLVMValueRef dst[4])
{
+ struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
- struct lp_build_context base = bld->bld_base.base;
const struct tgsi_full_dst_register *reg = &inst->Dst[0];
LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
- LLVMValueRef temp_ptr;
+ LLVMValueRef temp_ptr, temp_ptr2 = NULL;
unsigned chan, chan_index;
boolean is_vec_store = FALSE;
+ enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
+
if (dst[0]) {
LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
is_vec_store = (k == LLVMVectorTypeKind);
TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
LLVMValueRef value = dst[chan_index];
- if (inst->Instruction.Saturate != TGSI_SAT_NONE) {
- struct lp_build_emit_data clamp_emit_data;
-
- memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
- clamp_emit_data.arg_count = 3;
- clamp_emit_data.args[0] = value;
- clamp_emit_data.args[2] = base.one;
-
- switch(inst->Instruction.Saturate) {
- case TGSI_SAT_ZERO_ONE:
- clamp_emit_data.args[1] = base.zero;
- break;
- case TGSI_SAT_MINUS_PLUS_ONE:
- clamp_emit_data.args[1] = LLVMConstReal(
- base.elem_type, -1.0f);
- break;
- default:
- assert(0);
- }
- value = lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
- &clamp_emit_data);
- }
+ if (dtype == TGSI_TYPE_DOUBLE && (chan_index == 1 || chan_index == 3))
+ continue;
+ if (inst->Instruction.Saturate)
+ value = radeon_llvm_saturate(bld_base, value);
- switch(reg->Register.File) {
- case TGSI_FILE_ADDRESS:
+ if (reg->Register.File == TGSI_FILE_ADDRESS) {
temp_ptr = bld->addr[reg->Register.Index][chan_index];
LLVMBuildStore(builder, value, temp_ptr);
continue;
- case TGSI_FILE_OUTPUT:
- temp_ptr = bld->outputs[reg->Register.Index][chan_index];
- break;
+ }
- case TGSI_FILE_TEMPORARY:
- temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, chan_index);
- break;
+ if (dtype != TGSI_TYPE_DOUBLE)
+ value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
+
+ if (reg->Register.Indirect) {
+ struct tgsi_declaration_range range = get_array_range(bld_base,
+ reg->Register.File, ®->Indirect);
+
+ unsigned i, size = range.Last - range.First + 1;
+ LLVMValueRef array = LLVMBuildInsertElement(builder,
+ emit_array_fetch(bld_base, reg->Register.File, TGSI_TYPE_FLOAT, range, chan_index),
+ value, emit_array_index(bld, ®->Indirect, reg->Register.Index - range.First), "");
+
+ for (i = 0; i < size; ++i) {
+ switch(reg->Register.File) {
+ case TGSI_FILE_OUTPUT:
+ temp_ptr = bld->outputs[i + range.First][chan_index];
+ break;
+
+ case TGSI_FILE_TEMPORARY:
+ if (range.First + i >= ctx->temps_count)
+ continue;
+ if (uses_temp_indirect_addressing(bld_base))
+ temp_ptr = lp_get_temp_ptr_soa(bld, i + range.First, chan_index);
+ else
+ temp_ptr = ctx->temps[(i + range.First) * TGSI_NUM_CHANNELS + chan_index];
+ break;
+
+ default:
+ return;
+ }
+ value = LLVMBuildExtractElement(builder, array,
+ lp_build_const_int32(gallivm, i), "");
+ LLVMBuildStore(builder, value, temp_ptr);
+ }
- default:
- return;
- }
+ } else {
+ switch(reg->Register.File) {
+ case TGSI_FILE_OUTPUT:
+ temp_ptr = bld->outputs[reg->Register.Index][chan_index];
+ if (dtype == TGSI_TYPE_DOUBLE)
+ temp_ptr2 = bld->outputs[reg->Register.Index][chan_index + 1];
+ break;
+
+ case TGSI_FILE_TEMPORARY:
+ if (reg->Register.Index >= ctx->temps_count)
+ continue;
+ if (uses_temp_indirect_addressing(bld_base)) {
+ temp_ptr = NULL;
+ break;
+ }
+ temp_ptr = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index];
+ if (dtype == TGSI_TYPE_DOUBLE)
+ temp_ptr2 = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index + 1];
- value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
+ break;
- LLVMBuildStore(builder, value, temp_ptr);
+ default:
+ return;
+ }
+ if (dtype != TGSI_TYPE_DOUBLE)
+ LLVMBuildStore(builder, value, temp_ptr);
+ else {
+ LLVMValueRef ptr = LLVMBuildBitCast(builder, value,
+ LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), 2), "");
+ LLVMValueRef val2;
+ value = LLVMBuildExtractElement(builder, ptr,
+ bld_base->uint_bld.zero, "");
+ val2 = LLVMBuildExtractElement(builder, ptr,
+ bld_base->uint_bld.one, "");
+
+ LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, value), temp_ptr);
+ LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, val2), temp_ptr2);
+ }
+ }
}
}
endloop_block, "LOOP");
LLVMBuildBr(gallivm->builder, loop_block);
LLVMPositionBuilderAtEnd(gallivm->builder, loop_block);
- ctx->loop_depth++;
+
+ if (++ctx->loop_depth > ctx->loop_depth_max) {
+ unsigned new_max = ctx->loop_depth_max << 1;
+
+ if (!new_max)
+ new_max = RADEON_LLVM_INITIAL_CF_DEPTH;
+
+ ctx->loop = REALLOC(ctx->loop, ctx->loop_depth_max *
+ sizeof(ctx->loop[0]),
+ new_max * sizeof(ctx->loop[0]));
+ ctx->loop_depth_max = new_max;
+ }
+
ctx->loop[ctx->loop_depth - 1].loop_block = loop_block;
ctx->loop[ctx->loop_depth - 1].endloop_block = endloop_block;
}
ctx->loop_depth--;
}
-static void if_emit(
+static void if_cond_emit(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
+ struct lp_build_emit_data * emit_data,
+ LLVMValueRef cond)
{
struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
struct gallivm_state * gallivm = bld_base->base.gallivm;
- LLVMValueRef cond;
LLVMBasicBlockRef if_block, else_block, endif_block;
- cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
- bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
- bld_base->int_bld.zero, "");
-
endif_block = LLVMAppendBasicBlockInContext(gallivm->context,
ctx->main_fn, "ENDIF");
if_block = LLVMInsertBasicBlockInContext(gallivm->context,
LLVMBuildCondBr(gallivm->builder, cond, if_block, else_block);
LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
- ctx->branch_depth++;
+ if (++ctx->branch_depth > ctx->branch_depth_max) {
+ unsigned new_max = ctx->branch_depth_max << 1;
+
+ if (!new_max)
+ new_max = RADEON_LLVM_INITIAL_CF_DEPTH;
+
+ ctx->branch = REALLOC(ctx->branch, ctx->branch_depth_max *
+ sizeof(ctx->branch[0]),
+ new_max * sizeof(ctx->branch[0]));
+ ctx->branch_depth_max = new_max;
+ }
+
ctx->branch[ctx->branch_depth - 1].endif_block = endif_block;
ctx->branch[ctx->branch_depth - 1].if_block = if_block;
ctx->branch[ctx->branch_depth - 1].else_block = else_block;
ctx->branch[ctx->branch_depth - 1].has_else = 0;
}
+static void if_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct gallivm_state * gallivm = bld_base->base.gallivm;
+ LLVMValueRef cond;
+
+ cond = LLVMBuildFCmp(gallivm->builder, LLVMRealUNE,
+ emit_data->args[0],
+ bld_base->base.zero, "");
+
+ if_cond_emit(action, bld_base, emit_data, cond);
+}
+
+static void uif_emit(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct gallivm_state * gallivm = bld_base->base.gallivm;
+ LLVMValueRef cond;
+
+ cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
+ bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
+ bld_base->int_bld.zero, "");
+
+ if_cond_emit(action, bld_base, emit_data, cond);
+}
+
+static void kill_if_fetch_args(
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ const struct tgsi_full_instruction * inst = emit_data->inst;
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ unsigned i;
+ LLVMValueRef conds[TGSI_NUM_CHANNELS];
+
+ for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
+ LLVMValueRef value = lp_build_emit_fetch(bld_base, inst, 0, i);
+ conds[i] = LLVMBuildFCmp(builder, LLVMRealOLT, value,
+ bld_base->base.zero, "");
+ }
+
+ /* Or the conditions together */
+ for (i = TGSI_NUM_CHANNELS - 1; i > 0; i--) {
+ conds[i - 1] = LLVMBuildOr(builder, conds[i], conds[i - 1], "");
+ }
+
+ emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
+ emit_data->arg_count = 1;
+ emit_data->args[0] = LLVMBuildSelect(builder, conds[0],
+ lp_build_const_float(gallivm, -1.0f),
+ bld_base->base.zero, "");
+}
+
static void kil_emit(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
}
}
-void radeon_llvm_emit_prepare_cube_coords(
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data,
- LLVMValueRef *coords_arg)
+static void radeon_llvm_cube_to_2d_coords(struct lp_build_tgsi_context *bld_base,
+ LLVMValueRef *in, LLVMValueRef *out)
{
-
- unsigned target = emit_data->inst->Texture.Texture;
- unsigned opcode = emit_data->inst->Instruction.Opcode;
struct gallivm_state * gallivm = bld_base->base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef type = bld_base->base.elem_type;
LLVMValueRef coords[4];
LLVMValueRef mad_args[3];
- LLVMValueRef idx;
- struct LLVMOpaqueValue *cube_vec;
- LLVMValueRef v;
+ LLVMValueRef v, cube_vec;
unsigned i;
- cube_vec = lp_build_gather_values(bld_base->base.gallivm, coords_arg, 4);
- v = build_intrinsic(builder, "llvm.AMDGPU.cube", LLVMVectorType(type, 4),
+ cube_vec = lp_build_gather_values(bld_base->base.gallivm, in, 4);
+ v = lp_build_intrinsic(builder, "llvm.AMDGPU.cube", LLVMVectorType(type, 4),
&cube_vec, 1, LLVMReadNoneAttribute);
- for (i = 0; i < 4; ++i) {
- idx = lp_build_const_int32(gallivm, i);
- coords[i] = LLVMBuildExtractElement(builder, v, idx, "");
- }
+ for (i = 0; i < 4; ++i)
+ coords[i] = LLVMBuildExtractElement(builder, v,
+ lp_build_const_int32(gallivm, i), "");
- coords[2] = build_intrinsic(builder, "fabs",
+ coords[2] = lp_build_intrinsic(builder, "llvm.fabs.f32",
type, &coords[2], 1, LLVMReadNoneAttribute);
coords[2] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_RCP, coords[2]);
mad_args[0], mad_args[1], mad_args[2]);
/* apply xyz = yxw swizzle to cooords */
- coords[2] = coords[3];
- coords[3] = coords[1];
- coords[1] = coords[0];
- coords[0] = coords[3];
-
- /* all cases except simple cube map sampling require special handling
- * for coord vector */
- if (target != TGSI_TEXTURE_CUBE ||
- opcode != TGSI_OPCODE_TEX) {
-
- /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
- if (target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
-
- /* coords_arg.w component - array_index for cube arrays or
- * compare value for SHADOWCUBE */
- coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
- coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
- }
-
- /* for instructions that need additional src (compare/lod/bias),
- * put it in coord.w */
- if (opcode == TGSI_OPCODE_TEX2 ||
- opcode == TGSI_OPCODE_TXB2 ||
- opcode == TGSI_OPCODE_TXL2) {
- coords[3] = coords_arg[4];
- }
- }
-
- memcpy(coords_arg, coords, sizeof(coords));
+ out[0] = coords[1];
+ out[1] = coords[0];
+ out[2] = coords[3];
}
-static void txd_fetch_args(
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
+void radeon_llvm_emit_prepare_cube_coords(
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data,
+ LLVMValueRef *coords_arg,
+ LLVMValueRef *derivs_arg)
{
- const struct tgsi_full_instruction * inst = emit_data->inst;
-
- LLVMValueRef coords[4];
- unsigned chan, src;
- for (src = 0; src < 3; src++) {
- for (chan = 0; chan < 4; chan++)
- coords[chan] = lp_build_emit_fetch(bld_base, inst, src, chan);
- emit_data->args[src] = lp_build_gather_values(bld_base->base.gallivm,
- coords, 4);
- }
- emit_data->arg_count = 3;
- emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
-}
-
-
-static void txp_fetch_args(
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
-{
- const struct tgsi_full_instruction * inst = emit_data->inst;
- LLVMValueRef src_w;
- unsigned chan;
+ unsigned target = emit_data->inst->Texture.Texture;
+ unsigned opcode = emit_data->inst->Instruction.Opcode;
+ struct gallivm_state * gallivm = bld_base->base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef coords[4];
+ unsigned i;
- emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
- src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
-
- for (chan = 0; chan < 3; chan++ ) {
- LLVMValueRef arg = lp_build_emit_fetch(bld_base,
- emit_data->inst, 0, chan);
- coords[chan] = lp_build_emit_llvm_binary(bld_base,
- TGSI_OPCODE_DIV, arg, src_w);
- }
- coords[3] = bld_base->base.one;
-
- if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
- inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY ||
- inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
- inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE_ARRAY) &&
- inst->Instruction.Opcode != TGSI_OPCODE_TXQ &&
- inst->Instruction.Opcode != TGSI_OPCODE_TXQ_LZ) {
- radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords);
- }
-
- emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
- coords, 4);
- emit_data->arg_count = 1;
-}
-
-static void tex_fetch_args(
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
-{
- /* XXX: lp_build_swizzle_aos() was failing with wrong arg types,
- * when we used CHAN_ALL. We should be able to get this to work,
- * but for now we will swizzle it ourselves
- emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
- 0, CHAN_ALL);
-
- */
-
- const struct tgsi_full_instruction * inst = emit_data->inst;
-
- LLVMValueRef coords[5];
- unsigned chan;
- for (chan = 0; chan < 4; chan++) {
- coords[chan] = lp_build_emit_fetch(bld_base, inst, 0, chan);
- }
+ radeon_llvm_cube_to_2d_coords(bld_base, coords_arg, coords);
+
+ if (opcode == TGSI_OPCODE_TXD && derivs_arg) {
+ LLVMValueRef derivs[4];
+ int axis;
+
+ /* Convert cube derivatives to 2D derivatives. */
+ for (axis = 0; axis < 2; axis++) {
+ LLVMValueRef shifted_cube_coords[4], shifted_coords[4];
+
+ /* Shift the cube coordinates by the derivatives to get
+ * the cube coordinates of the "neighboring pixel".
+ */
+ for (i = 0; i < 3; i++)
+ shifted_cube_coords[i] =
+ LLVMBuildFAdd(builder, coords_arg[i],
+ derivs_arg[axis*3+i], "");
+ shifted_cube_coords[3] = LLVMGetUndef(bld_base->base.elem_type);
+
+ /* Project the shifted cube coordinates onto the face. */
+ radeon_llvm_cube_to_2d_coords(bld_base, shifted_cube_coords,
+ shifted_coords);
+
+ /* Subtract both sets of 2D coordinates to get 2D derivatives.
+ * This won't work if the shifted coordinates ended up
+ * in a different face.
+ */
+ for (i = 0; i < 2; i++)
+ derivs[axis * 2 + i] =
+ LLVMBuildFSub(builder, shifted_coords[i],
+ coords[i], "");
+ }
- if (inst->Instruction.Opcode == TGSI_OPCODE_TEX2 ||
- inst->Instruction.Opcode == TGSI_OPCODE_TXB2 ||
- inst->Instruction.Opcode == TGSI_OPCODE_TXL2) {
- /* These instructions have additional operand that should be packed
- * into the cube coord vector by radeon_llvm_emit_prepare_cube_coords.
- * That operand should be passed as a float value in the args array
- * right after the coord vector. After packing it's not used anymore,
- * that's why arg_count is not increased */
- coords[4] = lp_build_emit_fetch(bld_base, inst, 1, 0);
+ memcpy(derivs_arg, derivs, sizeof(derivs));
}
- if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
- inst->Texture.Texture == TGSI_TEXTURE_CUBE_ARRAY ||
- inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE ||
- inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE_ARRAY) &&
- inst->Instruction.Opcode != TGSI_OPCODE_TXQ &&
- inst->Instruction.Opcode != TGSI_OPCODE_TXQ_LZ) {
- radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords);
+ if (target == TGSI_TEXTURE_CUBE_ARRAY ||
+ target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
+ /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
+ /* coords_arg.w component - array_index for cube arrays */
+ coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
+ coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
}
- emit_data->arg_count = 1;
- emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
- coords, 4);
- emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
-}
-
-static void txf_fetch_args(
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
-{
- const struct tgsi_full_instruction * inst = emit_data->inst;
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
- const struct tgsi_texture_offset * off = inst->TexOffsets;
- LLVMTypeRef offset_type = bld_base->int_bld.elem_type;
-
- /* fetch tex coords */
- tex_fetch_args(bld_base, emit_data);
-
- /* fetch tex offsets */
- if (inst->Texture.NumOffsets) {
- assert(inst->Texture.NumOffsets == 1);
-
- emit_data->args[1] = LLVMConstBitCast(
- bld->immediates[off->Index][off->SwizzleX],
- offset_type);
- emit_data->args[2] = LLVMConstBitCast(
- bld->immediates[off->Index][off->SwizzleY],
- offset_type);
- emit_data->args[3] = LLVMConstBitCast(
- bld->immediates[off->Index][off->SwizzleZ],
- offset_type);
- } else {
- emit_data->args[1] = bld_base->int_bld.zero;
- emit_data->args[2] = bld_base->int_bld.zero;
- emit_data->args[3] = bld_base->int_bld.zero;
+ /* Preserve compare/lod/bias. Put it in coords.w. */
+ if (opcode == TGSI_OPCODE_TEX2 ||
+ opcode == TGSI_OPCODE_TXB2 ||
+ opcode == TGSI_OPCODE_TXL2) {
+ coords[3] = coords_arg[4];
+ } else if (opcode == TGSI_OPCODE_TXB ||
+ opcode == TGSI_OPCODE_TXL ||
+ target == TGSI_TEXTURE_SHADOWCUBE) {
+ coords[3] = coords_arg[3];
}
- emit_data->arg_count = 4;
+ memcpy(coords_arg, coords, sizeof(coords));
}
static void emit_icmp(
{
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
- LLVMValueRef v = LLVMBuildFCmp(builder, LLVMRealUGE,
- emit_data->args[0], lp_build_const_float(bld_base->base.gallivm, 0.), "");
+ LLVMValueRef arg0 = LLVMBuildBitCast(builder, emit_data->args[0],
+ bld_base->uint_bld.elem_type, "");
- emit_data->output[emit_data->chan] = LLVMBuildSelect(builder, v, emit_data->args[2], emit_data->args[1], "");
+ LLVMValueRef v = LLVMBuildICmp(builder, LLVMIntNE, arg0,
+ bld_base->uint_bld.zero, "");
+
+ emit_data->output[emit_data->chan] =
+ LLVMBuildSelect(builder, v, emit_data->args[1], emit_data->args[2], "");
}
-static void emit_cmp(
+static void emit_cmp(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMValueRef cond, *args = emit_data->args;
+
+ cond = LLVMBuildFCmp(builder, LLVMRealOLT, args[0],
+ bld_base->base.zero, "");
+
+ emit_data->output[emit_data->chan] =
+ LLVMBuildSelect(builder, cond, args[1], args[2], "");
+}
+
+static void emit_set_cond(
const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
LLVMRealPredicate pred;
LLVMValueRef cond;
- /* XXX I'm not sure whether to do unordered or ordered comparisons,
- * but llvmpipe uses unordered comparisons, so for consistency we use
- * unordered. (The authors of llvmpipe aren't sure about using
- * unordered vs ordered comparisons either.
+ /* Use ordered for everything but NE (which is usual for
+ * float comparisons)
*/
switch (emit_data->inst->Instruction.Opcode) {
- case TGSI_OPCODE_SGE: pred = LLVMRealUGE; break;
- case TGSI_OPCODE_SEQ: pred = LLVMRealUEQ; break;
- case TGSI_OPCODE_SLE: pred = LLVMRealULE; break;
- case TGSI_OPCODE_SLT: pred = LLVMRealULT; break;
+ case TGSI_OPCODE_SGE: pred = LLVMRealOGE; break;
+ case TGSI_OPCODE_SEQ: pred = LLVMRealOEQ; break;
+ case TGSI_OPCODE_SLE: pred = LLVMRealOLE; break;
+ case TGSI_OPCODE_SLT: pred = LLVMRealOLT; break;
case TGSI_OPCODE_SNE: pred = LLVMRealUNE; break;
- case TGSI_OPCODE_SGT: pred = LLVMRealUGT; break;
+ case TGSI_OPCODE_SGT: pred = LLVMRealOGT; break;
default: assert(!"unknown instruction"); pred = 0; break;
}
cond, bld_base->base.one, bld_base->base.zero, "");
}
+static void emit_fcmp(
+ const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMContextRef context = bld_base->base.gallivm->context;
+ LLVMRealPredicate pred;
+
+ /* Use ordered for everything but NE (which is usual for
+ * float comparisons)
+ */
+ switch (emit_data->inst->Instruction.Opcode) {
+ case TGSI_OPCODE_FSEQ: pred = LLVMRealOEQ; break;
+ case TGSI_OPCODE_FSGE: pred = LLVMRealOGE; break;
+ case TGSI_OPCODE_FSLT: pred = LLVMRealOLT; break;
+ case TGSI_OPCODE_FSNE: pred = LLVMRealUNE; break;
+ default: assert(!"unknown instruction"); pred = 0; break;
+ }
+
+ LLVMValueRef v = LLVMBuildFCmp(builder, pred,
+ emit_data->args[0], emit_data->args[1],"");
+
+ v = LLVMBuildSExtOrBitCast(builder, v,
+ LLVMInt32TypeInContext(context), "");
+
+ emit_data->output[emit_data->chan] = v;
+}
+
+static void emit_dcmp(
+ const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMContextRef context = bld_base->base.gallivm->context;
+ LLVMRealPredicate pred;
+
+ /* Use ordered for everything but NE (which is usual for
+ * float comparisons)
+ */
+ switch (emit_data->inst->Instruction.Opcode) {
+ case TGSI_OPCODE_DSEQ: pred = LLVMRealOEQ; break;
+ case TGSI_OPCODE_DSGE: pred = LLVMRealOGE; break;
+ case TGSI_OPCODE_DSLT: pred = LLVMRealOLT; break;
+ case TGSI_OPCODE_DSNE: pred = LLVMRealUNE; break;
+ default: assert(!"unknown instruction"); pred = 0; break;
+ }
+
+ LLVMValueRef v = LLVMBuildFCmp(builder, pred,
+ emit_data->args[0], emit_data->args[1],"");
+
+ v = LLVMBuildSExtOrBitCast(builder, v,
+ LLVMInt32TypeInContext(context), "");
+
+ emit_data->output[emit_data->chan] = v;
+}
+
static void emit_not(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
} else { // float SSG
- cmp = LLVMBuildFCmp(builder, LLVMRealUGT, emit_data->args[0], bld_base->base.zero, "");
+ cmp = LLVMBuildFCmp(builder, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
- cmp = LLVMBuildFCmp(builder, LLVMRealUGE, val, bld_base->base.zero, "");
+ cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
}
emit_data->args[0], "");
}
+static void emit_dneg(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ emit_data->output[emit_data->chan] = LLVMBuildFNeg(builder,
+ emit_data->args[0], "");
+}
+
+static void emit_frac(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ char *intr;
+
+ if (emit_data->info->opcode == TGSI_OPCODE_FRC)
+ intr = "llvm.floor.f32";
+ else if (emit_data->info->opcode == TGSI_OPCODE_DFRAC)
+ intr = "llvm.floor.f64";
+ else {
+ assert(0);
+ return;
+ }
+
+ LLVMValueRef floor = lp_build_intrinsic(builder, intr, emit_data->dst_type,
+ &emit_data->args[0], 1,
+ LLVMReadNoneAttribute);
+ emit_data->output[emit_data->chan] = LLVMBuildFSub(builder,
+ emit_data->args[0], floor, "");
+}
+
static void emit_f2i(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
ctx->soa.num_immediates++;
}
-LLVMValueRef
-build_intrinsic(LLVMBuilderRef builder,
- const char *name,
- LLVMTypeRef ret_type,
- LLVMValueRef *args,
- unsigned num_args,
- LLVMAttribute attr)
+void
+build_tgsi_intrinsic_nomem(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
{
- LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
- LLVMValueRef function;
+ struct lp_build_context * base = &bld_base->base;
+ emit_data->output[emit_data->chan] =
+ lp_build_intrinsic(base->gallivm->builder, action->intr_name,
+ emit_data->dst_type, emit_data->args,
+ emit_data->arg_count, LLVMReadNoneAttribute);
+}
- function = LLVMGetNamedFunction(module, name);
- if(!function) {
- LLVMTypeRef arg_types[LP_MAX_FUNC_ARGS];
- unsigned i;
+static void emit_bfi(const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef bfi_args[3];
- assert(num_args <= LP_MAX_FUNC_ARGS);
+ // Calculate the bitmask: (((1 << src3) - 1) << src2
+ bfi_args[0] = LLVMBuildShl(builder,
+ LLVMBuildSub(builder,
+ LLVMBuildShl(builder,
+ bld_base->int_bld.one,
+ emit_data->args[3], ""),
+ bld_base->int_bld.one, ""),
+ emit_data->args[2], "");
- for(i = 0; i < num_args; ++i) {
- assert(args[i]);
- arg_types[i] = LLVMTypeOf(args[i]);
- }
+ bfi_args[1] = LLVMBuildShl(builder, emit_data->args[1],
+ emit_data->args[2], "");
- function = lp_declare_intrinsic(module, name, ret_type, arg_types, num_args);
+ bfi_args[2] = emit_data->args[0];
- if (attr)
- LLVMAddFunctionAttr(function, attr);
- }
+ /* Calculate:
+ * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
+ * Use the right-hand side, which the LLVM backend can convert to V_BFI.
+ */
+ emit_data->output[emit_data->chan] =
+ LLVMBuildXor(builder, bfi_args[2],
+ LLVMBuildAnd(builder, bfi_args[0],
+ LLVMBuildXor(builder, bfi_args[1], bfi_args[2],
+ ""), ""), "");
+}
- return LLVMBuildCall(builder, function, args, num_args, "");
+/* this is ffs in C */
+static void emit_lsb(const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMValueRef args[2] = {
+ emit_data->args[0],
+
+ /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
+ * add special code to check for x=0. The reason is that
+ * the LLVM behavior for x=0 is different from what we
+ * need here.
+ *
+ * The hardware already implements the correct behavior.
+ */
+ lp_build_const_int32(gallivm, 1)
+ };
+
+ emit_data->output[emit_data->chan] =
+ lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
+ emit_data->dst_type, args, Elements(args),
+ LLVMReadNoneAttribute);
}
-static void build_tgsi_intrinsic(
- const struct lp_build_tgsi_action * action,
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data,
- LLVMAttribute attr)
+/* Find the last bit set. */
+static void emit_umsb(const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
{
- struct lp_build_context * base = &bld_base->base;
- emit_data->output[emit_data->chan] = build_intrinsic(
- base->gallivm->builder, action->intr_name,
- emit_data->dst_type, emit_data->args,
- emit_data->arg_count, attr);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef args[2] = {
+ emit_data->args[0],
+ /* Don't generate code for handling zero: */
+ lp_build_const_int32(gallivm, 1)
+ };
+
+ LLVMValueRef msb =
+ lp_build_intrinsic(builder, "llvm.ctlz.i32",
+ emit_data->dst_type, args, Elements(args),
+ LLVMReadNoneAttribute);
+
+ /* The HW returns the last bit index from MSB, but TGSI wants
+ * the index from LSB. Invert it by doing "31 - msb". */
+ msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
+ msb, "");
+
+ /* Check for zero: */
+ emit_data->output[emit_data->chan] =
+ LLVMBuildSelect(builder,
+ LLVMBuildICmp(builder, LLVMIntEQ, args[0],
+ bld_base->uint_bld.zero, ""),
+ lp_build_const_int32(gallivm, -1), msb, "");
}
-void
-build_tgsi_intrinsic_nomem(
- const struct lp_build_tgsi_action * action,
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
+
+/* Find the last bit opposite of the sign bit. */
+static void emit_imsb(const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
{
- build_tgsi_intrinsic(action, bld_base, emit_data, LLVMReadNoneAttribute);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef arg = emit_data->args[0];
+
+ LLVMValueRef msb =
+ lp_build_intrinsic(builder, "llvm.AMDGPU.flbit.i32",
+ emit_data->dst_type, &arg, 1,
+ LLVMReadNoneAttribute);
+
+ /* The HW returns the last bit index from MSB, but TGSI wants
+ * the index from LSB. Invert it by doing "31 - msb". */
+ msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
+ msb, "");
+
+ /* If arg == 0 || arg == -1 (0xffffffff), return -1. */
+ LLVMValueRef all_ones = lp_build_const_int32(gallivm, -1);
+
+ LLVMValueRef cond =
+ LLVMBuildOr(builder,
+ LLVMBuildICmp(builder, LLVMIntEQ, arg,
+ bld_base->uint_bld.zero, ""),
+ LLVMBuildICmp(builder, LLVMIntEQ, arg,
+ all_ones, ""), "");
+
+ emit_data->output[emit_data->chan] =
+ LLVMBuildSelect(builder, cond, all_ones, msb, "");
}
-static void build_tgsi_intrinsic_readonly(
- const struct lp_build_tgsi_action * action,
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
+static void emit_iabs(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
{
- build_tgsi_intrinsic(action, bld_base, emit_data, LLVMReadOnlyAttribute);
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+
+ emit_data->output[emit_data->chan] =
+ lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_IMAX,
+ emit_data->args[0],
+ LLVMBuildNeg(builder,
+ emit_data->args[0], ""));
+}
+
+static void emit_minmax_int(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ LLVMBuilderRef builder = bld_base->base.gallivm->builder;
+ LLVMIntPredicate op;
+
+ switch (emit_data->info->opcode) {
+ default:
+ assert(0);
+ case TGSI_OPCODE_IMAX:
+ op = LLVMIntSGT;
+ break;
+ case TGSI_OPCODE_IMIN:
+ op = LLVMIntSLT;
+ break;
+ case TGSI_OPCODE_UMAX:
+ op = LLVMIntUGT;
+ break;
+ case TGSI_OPCODE_UMIN:
+ op = LLVMIntULT;
+ break;
+ }
+
+ emit_data->output[emit_data->chan] =
+ LLVMBuildSelect(builder,
+ LLVMBuildICmp(builder, op, emit_data->args[0],
+ emit_data->args[1], ""),
+ emit_data->args[0],
+ emit_data->args[1], "");
}
void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
ctx->gallivm.context);
ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
- ctx->store_output_intr = "llvm.AMDGPU.store.output.";
- ctx->swizzle_intr = "llvm.AMDGPU.swizzle";
struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
- /* XXX: We need to revisit this.I think the correct way to do this is
- * to use length = 4 here and use the elem_bld for everything. */
type.floating = TRUE;
+ type.fixed = FALSE;
type.sign = TRUE;
+ type.norm = FALSE;
type.width = 32;
type.length = 1;
lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
lp_build_context_init(&ctx->soa.bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
lp_build_context_init(&ctx->soa.bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
+ {
+ struct lp_type dbl_type;
+ dbl_type = type;
+ dbl_type.width *= 2;
+ lp_build_context_init(&ctx->soa.bld_base.dbl_bld, &ctx->gallivm, dbl_type);
+ }
bld_base->soa = 1;
- bld_base->emit_store = emit_store;
+ bld_base->emit_store = radeon_llvm_emit_store;
bld_base->emit_swizzle = emit_swizzle;
bld_base->emit_declaration = emit_declaration;
bld_base->emit_immediate = emit_immediate;
- bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
- bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
- bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
- bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = emit_fetch_output;
+ bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = radeon_llvm_emit_fetch;
+ bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = radeon_llvm_emit_fetch;
+ bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = radeon_llvm_emit_fetch;
+ bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = radeon_llvm_emit_fetch;
+ bld_base->emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
/* Allocate outputs */
ctx->soa.outputs = ctx->outputs;
- /* XXX: Is there a better way to initialize all this ? */
-
lp_set_default_actions(bld_base);
- bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_readonly;
- bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "fabs";
- bld_base->op_actions[TGSI_OPCODE_ARL].emit = emit_arl;
+ bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "llvm.fabs.f32";
bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
+ bld_base->op_actions[TGSI_OPCODE_ARL].emit = emit_arl;
+ bld_base->op_actions[TGSI_OPCODE_BFI].emit = emit_bfi;
bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
+ bld_base->op_actions[TGSI_OPCODE_BREV].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_BREV].intr_name = "llvm.AMDGPU.brev";
bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
- bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_readonly;
- bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "ceil";
+ bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "llvm.ceil.f32";
bld_base->op_actions[TGSI_OPCODE_CLAMP].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_CLAMP].intr_name = "llvm.AMDIL.clamp.";
- bld_base->op_actions[TGSI_OPCODE_CMP].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_CMP].intr_name = "llvm.AMDGPU.cndlt";
+ bld_base->op_actions[TGSI_OPCODE_CMP].emit = emit_cmp;
bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
- bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_readonly;
+ bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_COS].intr_name = "llvm.cos.f32";
- bld_base->op_actions[TGSI_OPCODE_DDX].intr_name = "llvm.AMDGPU.ddx";
- bld_base->op_actions[TGSI_OPCODE_DDX].fetch_args = tex_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_DDY].intr_name = "llvm.AMDGPU.ddy";
- bld_base->op_actions[TGSI_OPCODE_DDY].fetch_args = tex_fetch_args;
+ bld_base->op_actions[TGSI_OPCODE_DABS].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DABS].intr_name = "llvm.fabs.f64";
+ bld_base->op_actions[TGSI_OPCODE_DFMA].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DFMA].intr_name = "llvm.fma.f64";
+ bld_base->op_actions[TGSI_OPCODE_DFRAC].emit = emit_frac;
+ bld_base->op_actions[TGSI_OPCODE_DNEG].emit = emit_dneg;
+ bld_base->op_actions[TGSI_OPCODE_DSEQ].emit = emit_dcmp;
+ bld_base->op_actions[TGSI_OPCODE_DSGE].emit = emit_dcmp;
+ bld_base->op_actions[TGSI_OPCODE_DSLT].emit = emit_dcmp;
+ bld_base->op_actions[TGSI_OPCODE_DSNE].emit = emit_dcmp;
+ bld_base->op_actions[TGSI_OPCODE_DRSQ].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DRSQ].intr_name = "llvm.AMDGPU.rsq.f64";
+ bld_base->op_actions[TGSI_OPCODE_DSQRT].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_DSQRT].intr_name = "llvm.sqrt.f64";
bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.AMDIL.exp.";
- bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_readonly;
- bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "floor";
- bld_base->op_actions[TGSI_OPCODE_FRC].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_FRC].intr_name = "llvm.AMDIL.fraction.";
+ bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "llvm.floor.f32";
+ bld_base->op_actions[TGSI_OPCODE_FMA].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_FMA].intr_name = "llvm.fma.f32";
+ bld_base->op_actions[TGSI_OPCODE_FRC].emit = emit_frac;
bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
- bld_base->op_actions[TGSI_OPCODE_IABS].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_IABS].intr_name = "llvm.AMDIL.abs.";
+ bld_base->op_actions[TGSI_OPCODE_FSEQ].emit = emit_fcmp;
+ bld_base->op_actions[TGSI_OPCODE_FSGE].emit = emit_fcmp;
+ bld_base->op_actions[TGSI_OPCODE_FSLT].emit = emit_fcmp;
+ bld_base->op_actions[TGSI_OPCODE_FSNE].emit = emit_fcmp;
+ bld_base->op_actions[TGSI_OPCODE_IABS].emit = emit_iabs;
+ bld_base->op_actions[TGSI_OPCODE_IBFE].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_IBFE].intr_name = "llvm.AMDGPU.bfe.i32";
bld_base->op_actions[TGSI_OPCODE_IDIV].emit = emit_idiv;
bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
- bld_base->op_actions[TGSI_OPCODE_IMAX].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_IMAX].intr_name = "llvm.AMDGPU.imax";
- bld_base->op_actions[TGSI_OPCODE_IMIN].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_IMIN].intr_name = "llvm.AMDGPU.imin";
+ bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
+ bld_base->op_actions[TGSI_OPCODE_IMAX].emit = emit_minmax_int;
+ bld_base->op_actions[TGSI_OPCODE_IMIN].emit = emit_minmax_int;
+ bld_base->op_actions[TGSI_OPCODE_IMSB].emit = emit_imsb;
bld_base->op_actions[TGSI_OPCODE_INEG].emit = emit_ineg;
bld_base->op_actions[TGSI_OPCODE_ISHR].emit = emit_ishr;
bld_base->op_actions[TGSI_OPCODE_ISGE].emit = emit_icmp;
bld_base->op_actions[TGSI_OPCODE_ISLT].emit = emit_icmp;
bld_base->op_actions[TGSI_OPCODE_ISSG].emit = emit_ssg;
bld_base->op_actions[TGSI_OPCODE_I2F].emit = emit_i2f;
- bld_base->op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
- bld_base->op_actions[TGSI_OPCODE_KIL].intr_name = "llvm.AMDGPU.kill";
- bld_base->op_actions[TGSI_OPCODE_KILP].emit = lp_build_tgsi_intrinsic;
- bld_base->op_actions[TGSI_OPCODE_KILP].intr_name = "llvm.AMDGPU.kilp";
- bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_readonly;
+ bld_base->op_actions[TGSI_OPCODE_KILL_IF].fetch_args = kill_if_fetch_args;
+ bld_base->op_actions[TGSI_OPCODE_KILL_IF].emit = kil_emit;
+ bld_base->op_actions[TGSI_OPCODE_KILL_IF].intr_name = "llvm.AMDGPU.kill";
+ bld_base->op_actions[TGSI_OPCODE_KILL].emit = lp_build_tgsi_intrinsic;
+ bld_base->op_actions[TGSI_OPCODE_KILL].intr_name = "llvm.AMDGPU.kilp";
+ bld_base->op_actions[TGSI_OPCODE_LSB].emit = emit_lsb;
+ bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_LG2].intr_name = "llvm.log2.f32";
- bld_base->op_actions[TGSI_OPCODE_LRP].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_LRP].intr_name = "llvm.AMDGPU.lrp";
bld_base->op_actions[TGSI_OPCODE_MOD].emit = emit_mod;
+ bld_base->op_actions[TGSI_OPCODE_UMSB].emit = emit_umsb;
bld_base->op_actions[TGSI_OPCODE_NOT].emit = emit_not;
bld_base->op_actions[TGSI_OPCODE_OR].emit = emit_or;
- bld_base->op_actions[TGSI_OPCODE_POW].emit = build_tgsi_intrinsic_readonly;
+ bld_base->op_actions[TGSI_OPCODE_POPC].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_POPC].intr_name = "llvm.ctpop.i32";
+ bld_base->op_actions[TGSI_OPCODE_POW].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_POW].intr_name = "llvm.pow.f32";
bld_base->op_actions[TGSI_OPCODE_ROUND].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_ROUND].intr_name = "llvm.AMDIL.round.nearest.";
- bld_base->op_actions[TGSI_OPCODE_SGE].emit = emit_cmp;
- bld_base->op_actions[TGSI_OPCODE_SEQ].emit = emit_cmp;
+ bld_base->op_actions[TGSI_OPCODE_ROUND].intr_name = "llvm.rint.f32";
+ bld_base->op_actions[TGSI_OPCODE_RSQ].intr_name = "llvm.AMDGPU.rsq.clamped.f32";
+ bld_base->op_actions[TGSI_OPCODE_RSQ].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_SGE].emit = emit_set_cond;
+ bld_base->op_actions[TGSI_OPCODE_SEQ].emit = emit_set_cond;
bld_base->op_actions[TGSI_OPCODE_SHL].emit = emit_shl;
- bld_base->op_actions[TGSI_OPCODE_SLE].emit = emit_cmp;
- bld_base->op_actions[TGSI_OPCODE_SLT].emit = emit_cmp;
- bld_base->op_actions[TGSI_OPCODE_SNE].emit = emit_cmp;
- bld_base->op_actions[TGSI_OPCODE_SGT].emit = emit_cmp;
- bld_base->op_actions[TGSI_OPCODE_SIN].emit = build_tgsi_intrinsic_readonly;
+ bld_base->op_actions[TGSI_OPCODE_SLE].emit = emit_set_cond;
+ bld_base->op_actions[TGSI_OPCODE_SLT].emit = emit_set_cond;
+ bld_base->op_actions[TGSI_OPCODE_SNE].emit = emit_set_cond;
+ bld_base->op_actions[TGSI_OPCODE_SGT].emit = emit_set_cond;
+ bld_base->op_actions[TGSI_OPCODE_SIN].emit = build_tgsi_intrinsic_nomem;
bld_base->op_actions[TGSI_OPCODE_SIN].intr_name = "llvm.sin.f32";
+ bld_base->op_actions[TGSI_OPCODE_SQRT].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_SQRT].intr_name = "llvm.sqrt.f32";
bld_base->op_actions[TGSI_OPCODE_SSG].emit = emit_ssg;
- bld_base->op_actions[TGSI_OPCODE_TEX].fetch_args = tex_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TEX].intr_name = "llvm.AMDGPU.tex";
- bld_base->op_actions[TGSI_OPCODE_TEX2].fetch_args = tex_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TEX2].intr_name = "llvm.AMDGPU.tex";
- bld_base->op_actions[TGSI_OPCODE_TXB].fetch_args = tex_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TXB].intr_name = "llvm.AMDGPU.txb";
- bld_base->op_actions[TGSI_OPCODE_TXB2].fetch_args = tex_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TXB2].intr_name = "llvm.AMDGPU.txb";
- bld_base->op_actions[TGSI_OPCODE_TXD].fetch_args = txd_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TXD].intr_name = "llvm.AMDGPU.txd";
- bld_base->op_actions[TGSI_OPCODE_TXF].fetch_args = txf_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TXF].intr_name = "llvm.AMDGPU.txf";
- bld_base->op_actions[TGSI_OPCODE_TXL].fetch_args = tex_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TXL].intr_name = "llvm.AMDGPU.txl";
- bld_base->op_actions[TGSI_OPCODE_TXL2].fetch_args = tex_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TXL2].intr_name = "llvm.AMDGPU.txl";
- bld_base->op_actions[TGSI_OPCODE_TXP].fetch_args = txp_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TXP].intr_name = "llvm.AMDGPU.tex";
- bld_base->op_actions[TGSI_OPCODE_TXQ].fetch_args = tex_fetch_args;
- bld_base->op_actions[TGSI_OPCODE_TXQ].intr_name = "llvm.AMDGPU.txq";
bld_base->op_actions[TGSI_OPCODE_TRUNC].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.AMDGPU.trunc";
+ bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.trunc.f32";
bld_base->op_actions[TGSI_OPCODE_UADD].emit = emit_uadd;
+ bld_base->op_actions[TGSI_OPCODE_UBFE].emit = build_tgsi_intrinsic_nomem;
+ bld_base->op_actions[TGSI_OPCODE_UBFE].intr_name = "llvm.AMDGPU.bfe.u32";
bld_base->op_actions[TGSI_OPCODE_UDIV].emit = emit_udiv;
- bld_base->op_actions[TGSI_OPCODE_UMAX].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_UMAX].intr_name = "llvm.AMDGPU.umax";
- bld_base->op_actions[TGSI_OPCODE_UMIN].emit = build_tgsi_intrinsic_nomem;
- bld_base->op_actions[TGSI_OPCODE_UMIN].intr_name = "llvm.AMDGPU.umin";
+ bld_base->op_actions[TGSI_OPCODE_UMAX].emit = emit_minmax_int;
+ bld_base->op_actions[TGSI_OPCODE_UMIN].emit = emit_minmax_int;
bld_base->op_actions[TGSI_OPCODE_UMOD].emit = emit_umod;
bld_base->op_actions[TGSI_OPCODE_USEQ].emit = emit_icmp;
bld_base->op_actions[TGSI_OPCODE_USGE].emit = emit_icmp;
bld_base->op_actions[TGSI_OPCODE_U2F].emit = emit_u2f;
bld_base->op_actions[TGSI_OPCODE_XOR].emit = emit_xor;
bld_base->op_actions[TGSI_OPCODE_UCMP].emit = emit_ucmp;
-
- bld_base->rsq_action.emit = build_tgsi_intrinsic_nomem;
- bld_base->rsq_action.intr_name = "llvm.AMDGPU.rsq";
}
void radeon_llvm_create_func(struct radeon_llvm_context * ctx,
/* Add some optimization passes */
LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
+ LLVMAddLICMPass(gallivm->passmgr);
+ LLVMAddAggressiveDCEPass(gallivm->passmgr);
LLVMAddCFGSimplificationPass(gallivm->passmgr);
+ LLVMAddInstructionCombiningPass(gallivm->passmgr);
- /* Run the passs */
+ /* Run the pass */
LLVMRunFunctionPassManager(gallivm->passmgr, ctx->main_fn);
LLVMDisposeBuilder(gallivm->builder);
{
LLVMDisposeModule(ctx->soa.bld_base.base.gallivm->module);
LLVMContextDispose(ctx->soa.bld_base.base.gallivm->context);
+ FREE(ctx->arrays);
+ ctx->arrays = NULL;
+ FREE(ctx->temps);
+ ctx->temps = NULL;
+ ctx->temps_count = 0;
+ FREE(ctx->loop);
+ ctx->loop = NULL;
+ ctx->loop_depth_max = 0;
+ FREE(ctx->branch);
+ ctx->branch = NULL;
+ ctx->branch_depth_max = 0;
}