static LLVMValueRef
emit_array_fetch(
struct lp_build_tgsi_context *bld_base,
unsigned File, enum tgsi_opcode_type type,
struct tgsi_declaration_range range,
unsigned swizzle)
{
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;
unsigned i, size = range.Last - range.First + 1;
LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
LLVMValueRef result = LLVMGetUndef(vec);
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 = 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(
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, ptr;
if (swizzle == ~0) {
LLVMValueRef values[TGSI_NUM_CHANNELS];
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
values[chan] = emit_fetch(bld_base, reg, type, chan);
}
return lp_build_gather_values(bld_base->base.gallivm, values,
TGSI_NUM_CHANNELS);
}
if (reg->Register.Indirect) {
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);
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)];
break;
case TGSI_FILE_TEMPORARY:
ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
result = LLVMBuildLoad(builder, ptr, "");
break;
case TGSI_FILE_OUTPUT:
ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
result = LLVMBuildLoad(builder, ptr, "");
break;
default:
return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
}
return bitcast(bld_base, type, result);
}
static LLVMValueRef
emit_tex(struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_instruction *inst,
enum lp_build_tex_modifier modifier)
{
unsigned target;
unsigned unit;
LLVMValueRef coords;
LLVMValueRef ddx;
LLVMValueRef ddy;
if (!bld->sampler) {
_debug_printf("warning: found texture instruction but no sampler generator supplied\n");
return bld->base.undef;
}
target = inst->Texture.Texture;
coords = emit_fetch( bld, inst, 0 );
if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
ddx = emit_fetch( bld, inst, 1 );
ddy = emit_fetch( bld, inst, 2 );
unit = inst->Src[3].Register.Index;
} else {
#if 0
ddx = lp_build_ddx( &bld->base, coords );
ddy = lp_build_ddy( &bld->base, coords );
#else
/* TODO */
ddx = bld->base.one;
ddy = bld->base.one;
#endif
unit = inst->Src[1].Register.Index;
}
return bld->sampler->emit_fetch_texel(bld->sampler,
&bld->base,
target, unit,
coords, ddx, ddy,
modifier);
}
/**
* Emit LLVM for one TGSI instruction.
* \param return TRUE for success, FALSE otherwise
*/
static boolean
emit_instruction(
struct lp_build_tgsi_aos_context *bld,
const struct tgsi_full_instruction *inst,
const struct tgsi_opcode_info *info,
int *pc)
{
LLVMValueRef src0, src1, src2;
LLVMValueRef tmp0, tmp1;
LLVMValueRef dst0;
/*
* Stores and write masks are handled in a general fashion after the long
* instruction opcode switch statement.
*
* Although not stricitly necessary, we avoid generating instructions for
* channels which won't be stored, in cases where's that easy. For some
* complex instructions, like texture sampling, it is more convenient to
* assume a full writemask and then let LLVM optimization passes eliminate
* redundant code.
*/
(*pc)++;
assert(info->num_dst <= 1);
if (info->num_dst) {
dst0 = bld->base.undef;
}
switch (inst->Instruction.Opcode) {
case TGSI_OPCODE_ARL:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_floor(&bld->base, src0);
break;
case TGSI_OPCODE_MOV:
dst0 = emit_fetch(bld, inst, 0);
break;
case TGSI_OPCODE_LIT:
return FALSE;
case TGSI_OPCODE_RCP:
/* TGSI_OPCODE_RECIP */
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_rcp(&bld->base, src0);
break;
case TGSI_OPCODE_RSQ:
/* TGSI_OPCODE_RECIPSQRT */
src0 = emit_fetch(bld, inst, 0);
tmp0 = lp_build_abs(&bld->base, src0);
dst0 = lp_build_rsqrt(&bld->base, tmp0);
break;
case TGSI_OPCODE_EXP:
return FALSE;
case TGSI_OPCODE_LOG:
return FALSE;
case TGSI_OPCODE_MUL:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_mul(&bld->base, src0, src1);
break;
case TGSI_OPCODE_ADD:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_add(&bld->base, src0, src1);
break;
case TGSI_OPCODE_DP3:
/* TGSI_OPCODE_DOT3 */
return FALSE;
case TGSI_OPCODE_DP4:
/* TGSI_OPCODE_DOT4 */
return FALSE;
case TGSI_OPCODE_DST:
return FALSE;
case TGSI_OPCODE_MIN:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_max(&bld->base, src0, src1);
break;
case TGSI_OPCODE_MAX:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_max(&bld->base, src0, src1);
break;
case TGSI_OPCODE_SLT:
/* TGSI_OPCODE_SETLT */
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_SGE:
/* TGSI_OPCODE_SETGE */
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_MAD:
/* TGSI_OPCODE_MADD */
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp0 = lp_build_mul(&bld->base, src0, src1);
dst0 = lp_build_add(&bld->base, tmp0, src2);
break;
case TGSI_OPCODE_SUB:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
dst0 = lp_build_sub(&bld->base, src0, src1);
break;
case TGSI_OPCODE_LRP:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp0 = lp_build_sub(&bld->base, src1, src2);
tmp0 = lp_build_mul(&bld->base, src0, tmp0);
dst0 = lp_build_add(&bld->base, tmp0, src2);
break;
case TGSI_OPCODE_CND:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp1 = lp_build_const_vec(bld->base.type, 0.5);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src2, tmp1);
dst0 = lp_build_select(&bld->base, tmp0, src0, src1);
break;
case TGSI_OPCODE_DP2A:
return FALSE;
case TGSI_OPCODE_FRC:
src0 = emit_fetch(bld, inst, 0);
tmp0 = lp_build_floor(&bld->base, src0);
dst0 = lp_build_sub(&bld->base, src0, tmp0);
break;
case TGSI_OPCODE_CLAMP:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp0 = lp_build_max(&bld->base, src0, src1);
dst0 = lp_build_min(&bld->base, tmp0, src2);
break;
case TGSI_OPCODE_FLR:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_floor(&bld->base, src0);
break;
case TGSI_OPCODE_ROUND:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_round(&bld->base, src0);
break;
case TGSI_OPCODE_EX2:
src0 = emit_fetch(bld, inst, 0);
tmp0 = lp_build_swizzle_scalar_aos(&bld->base, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_exp2(&bld->base, tmp0);
break;
case TGSI_OPCODE_LG2:
src0 = emit_fetch(bld, inst, 0);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_log2(&bld->base, tmp0);
break;
case TGSI_OPCODE_POW:
src0 = emit_fetch(bld, inst, 0);
src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
src1 = emit_fetch(bld, inst, 1);
src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
dst0 = lp_build_pow(&bld->base, src0, src1);
break;
case TGSI_OPCODE_XPD:
return FALSE;
case TGSI_OPCODE_ABS:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_abs(&bld->base, src0);
break;
case TGSI_OPCODE_RCC:
/* deprecated? */
assert(0);
return FALSE;
case TGSI_OPCODE_DPH:
return FALSE;
case TGSI_OPCODE_COS:
src0 = emit_fetch(bld, inst, 0);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_cos(&bld->base, tmp0);
break;
case TGSI_OPCODE_DDX:
return FALSE;
case TGSI_OPCODE_DDY:
return FALSE;
case TGSI_OPCODE_KILP:
/* predicated kill */
return FALSE;
case TGSI_OPCODE_KIL:
/* conditional kill */
return FALSE;
case TGSI_OPCODE_PK2H:
return FALSE;
break;
case TGSI_OPCODE_PK2US:
return FALSE;
break;
case TGSI_OPCODE_PK4B:
return FALSE;
break;
case TGSI_OPCODE_PK4UB:
return FALSE;
case TGSI_OPCODE_RFL:
return FALSE;
case TGSI_OPCODE_SEQ:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_EQUAL, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_SFL:
dst0 = bld->base.zero;
break;
case TGSI_OPCODE_SGT:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_SIN:
src0 = emit_fetch(bld, inst, 0);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_sin(&bld->base, tmp0);
break;
case TGSI_OPCODE_SLE:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LEQUAL, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_SNE:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL, src0, src1);
dst0 = lp_build_select(&bld->base, tmp0, bld->base.one, bld->base.zero);
break;
case TGSI_OPCODE_STR:
dst0 = bld->base.one;
break;
case TGSI_OPCODE_TEX:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
break;
case TGSI_OPCODE_TXD:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
break;
case TGSI_OPCODE_UP2H:
/* deprecated */
assert (0);
return FALSE;
break;
case TGSI_OPCODE_UP2US:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_UP4B:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_UP4UB:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_X2D:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ARA:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ARR:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_round(&bld->base, src0);
break;
case TGSI_OPCODE_BRA:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CAL:
return FALSE;
case TGSI_OPCODE_RET:
return FALSE;
case TGSI_OPCODE_END:
*pc = -1;
break;
case TGSI_OPCODE_SSG:
/* TGSI_OPCODE_SGN */
tmp0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_sgn(&bld->base, tmp0);
break;
case TGSI_OPCODE_CMP:
src0 = emit_fetch(bld, inst, 0);
src1 = emit_fetch(bld, inst, 1);
src2 = emit_fetch(bld, inst, 2);
tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_LESS, src0, bld->base.zero);
dst0 = lp_build_select(&bld->base, tmp0, src1, src2);
break;
case TGSI_OPCODE_SCS:
return FALSE;
case TGSI_OPCODE_TXB:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
break;
case TGSI_OPCODE_NRM:
/* fall-through */
case TGSI_OPCODE_NRM4:
return FALSE;
case TGSI_OPCODE_DIV:
/* deprecated */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_DP2:
return FALSE;
case TGSI_OPCODE_TXL:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
break;
case TGSI_OPCODE_TXP:
dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
break;
case TGSI_OPCODE_BRK:
return FALSE;
case TGSI_OPCODE_IF:
return FALSE;
case TGSI_OPCODE_BGNLOOP:
return FALSE;
case TGSI_OPCODE_BGNSUB:
return FALSE;
case TGSI_OPCODE_ELSE:
return FALSE;
case TGSI_OPCODE_ENDIF:
return FALSE;
case TGSI_OPCODE_ENDLOOP:
return FALSE;
case TGSI_OPCODE_ENDSUB:
return FALSE;
case TGSI_OPCODE_PUSHA:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_POPA:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CEIL:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_ceil(&bld->base, src0);
break;
case TGSI_OPCODE_I2F:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_NOT:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TRUNC:
src0 = emit_fetch(bld, inst, 0);
dst0 = lp_build_trunc(&bld->base, src0);
break;
case TGSI_OPCODE_SHL:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_ISHR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_AND:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_OR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_MOD:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_XOR:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_SAD:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TXF:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_TXQ:
/* deprecated? */
assert(0);
return FALSE;
break;
case TGSI_OPCODE_CONT:
return FALSE;
case TGSI_OPCODE_EMIT:
return FALSE;
break;
case TGSI_OPCODE_ENDPRIM:
return FALSE;
break;
case TGSI_OPCODE_NOP:
break;
default:
return FALSE;
}
if (info->num_dst) {
emit_store(bld, inst, 0, dst0);
}
return TRUE;
}
int main(int argc, const char *argv[])
{
int is_config = (argc == 2) && (strcmp(argv[1], "config") == 0);
return is_config ? emit_config() : emit_fetch();
}
