static LLVMValueRef
lp_build_blend_swizzle(struct lp_build_blend_aos_context *bld,
LLVMValueRef rgb,
LLVMValueRef alpha,
enum lp_build_blend_swizzle rgb_swizzle,
unsigned alpha_swizzle)
{
LLVMValueRef swizzled_rgb;
switch (rgb_swizzle) {
case LP_BUILD_BLEND_SWIZZLE_RGBA:
swizzled_rgb = rgb;
break;
case LP_BUILD_BLEND_SWIZZLE_AAAA:
swizzled_rgb = lp_build_swizzle_scalar_aos(&bld->base, rgb, alpha_swizzle);
break;
default:
assert(0);
swizzled_rgb = bld->base.undef;
}
if (rgb != alpha) {
swizzled_rgb = lp_build_select_aos(&bld->base, 1 << alpha_swizzle,
alpha, swizzled_rgb);
}
return swizzled_rgb;
}
static LLVMValueRef
swizzle_scalar_aos(struct lp_build_tgsi_aos_context *bld,
LLVMValueRef a,
unsigned chan)
{
chan = bld->swizzles[chan];
return lp_build_swizzle_scalar_aos(&bld->bld_base.base, a, chan, 4);
}
LLVMValueRef
lp_build_swizzle_aos(struct lp_build_context *bld,
LLVMValueRef a,
const unsigned char swizzles[4])
{
LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
const unsigned n = type.length;
unsigned i, j;
if (swizzles[0] == PIPE_SWIZZLE_X &&
swizzles[1] == PIPE_SWIZZLE_Y &&
swizzles[2] == PIPE_SWIZZLE_Z &&
swizzles[3] == PIPE_SWIZZLE_W) {
return a;
}
if (swizzles[0] == swizzles[1] &&
swizzles[1] == swizzles[2] &&
swizzles[2] == swizzles[3]) {
switch (swizzles[0]) {
case PIPE_SWIZZLE_X:
case PIPE_SWIZZLE_Y:
case PIPE_SWIZZLE_Z:
case PIPE_SWIZZLE_W:
return lp_build_swizzle_scalar_aos(bld, a, swizzles[0], 4);
case PIPE_SWIZZLE_0:
return bld->zero;
case PIPE_SWIZZLE_1:
return bld->one;
case LP_BLD_SWIZZLE_DONTCARE:
return bld->undef;
default:
assert(0);
return bld->undef;
}
}
if (LLVMIsConstant(a) ||
type.width >= 16) {
/*
* Shuffle.
*/
LLVMValueRef undef = LLVMGetUndef(lp_build_elem_type(bld->gallivm, type));
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
LLVMValueRef aux[LP_MAX_VECTOR_LENGTH];
memset(aux, 0, sizeof aux);
for(j = 0; j < n; j += 4) {
for(i = 0; i < 4; ++i) {
unsigned shuffle;
switch (swizzles[i]) {
default:
assert(0);
/* fall through */
case PIPE_SWIZZLE_X:
case PIPE_SWIZZLE_Y:
case PIPE_SWIZZLE_Z:
case PIPE_SWIZZLE_W:
shuffle = j + swizzles[i];
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
break;
case PIPE_SWIZZLE_0:
shuffle = type.length + 0;
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
if (!aux[0]) {
aux[0] = lp_build_const_elem(bld->gallivm, type, 0.0);
}
break;
case PIPE_SWIZZLE_1:
shuffle = type.length + 1;
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
if (!aux[1]) {
aux[1] = lp_build_const_elem(bld->gallivm, type, 1.0);
}
break;
case LP_BLD_SWIZZLE_DONTCARE:
shuffles[j + i] = LLVMGetUndef(i32t);
break;
}
}
}
for (i = 0; i < n; ++i) {
if (!aux[i]) {
aux[i] = undef;
}
}
return LLVMBuildShuffleVector(builder, a,
LLVMConstVector(aux, n),
LLVMConstVector(shuffles, n), "");
} else {
/*
* Bit mask and shifts.
*
* For example, this will convert BGRA to RGBA by doing
*
* Little endian:
* rgba = (bgra & 0x00ff0000) >> 16
* | (bgra & 0xff00ff00)
* | (bgra & 0x000000ff) << 16
*
* Big endian:A
* rgba = (bgra & 0x0000ff00) << 16
* | (bgra & 0x00ff00ff)
* | (bgra & 0xff000000) >> 16
*
* This is necessary not only for faster cause, but because X86 backend
* will refuse shuffles of <4 x i8> vectors
*/
LLVMValueRef res;
struct lp_type type4;
unsigned cond = 0;
unsigned chan;
int shift;
/*
* Start with a mixture of 1 and 0.
*/
for (chan = 0; chan < 4; ++chan) {
if (swizzles[chan] == PIPE_SWIZZLE_1) {
cond |= 1 << chan;
}
}
res = lp_build_select_aos(bld, cond, bld->one, bld->zero, 4);
/*
* Build a type where each element is an integer that cover the four
* channels.
*/
type4 = type;
type4.floating = FALSE;
type4.width *= 4;
type4.length /= 4;
a = LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type4), "");
res = LLVMBuildBitCast(builder, res, lp_build_vec_type(bld->gallivm, type4), "");
/*
* Mask and shift the channels, trying to group as many channels in the
* same shift as possible. The shift amount is positive for shifts left
* and negative for shifts right.
*/
for (shift = -3; shift <= 3; ++shift) {
uint64_t mask = 0;
assert(type4.width <= sizeof(mask)*8);
/*
* Vector element numbers follow the XYZW order, so 0 is always X, etc.
* After widening 4 times we have:
*
* 3210
* Little-endian register layout: WZYX
*
* 0123
* Big-endian register layout: XYZW
*
* For little-endian, higher-numbered channels are obtained by a shift right
* (negative shift amount) and lower-numbered channels by a shift left
* (positive shift amount). The opposite is true for big-endian.
*/
for (chan = 0; chan < 4; ++chan) {
if (swizzles[chan] < 4) {
/* We need to move channel swizzles[chan] into channel chan */
#ifdef PIPE_ARCH_LITTLE_ENDIAN
if (swizzles[chan] - chan == -shift) {
mask |= ((1ULL << type.width) - 1) << (swizzles[chan] * type.width);
}
#else
if (swizzles[chan] - chan == shift) {
mask |= ((1ULL << type.width) - 1) << (type4.width - type.width) >> (swizzles[chan] * type.width);
}
#endif
}
}
if (mask) {
LLVMValueRef masked;
LLVMValueRef shifted;
if (0)
debug_printf("shift = %i, mask = %" PRIx64 "\n", shift, mask);
masked = LLVMBuildAnd(builder, a,
lp_build_const_int_vec(bld->gallivm, type4, mask), "");
if (shift > 0) {
shifted = LLVMBuildShl(builder, masked,
lp_build_const_int_vec(bld->gallivm, type4, shift*type.width), "");
} else if (shift < 0) {
shifted = LLVMBuildLShr(builder, masked,
lp_build_const_int_vec(bld->gallivm, type4, -shift*type.width), "");
} else {
shifted = masked;
}
res = LLVMBuildOr(builder, res, shifted, "");
}
}
return LLVMBuildBitCast(builder, res,
lp_build_vec_type(bld->gallivm, type), "");
}
/**
* Emit LLVM for one TGSI instruction.
* \param return TRUE for success, FALSE otherwise
*/
boolean
lp_emit_instruction_aos(
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 = NULL;
/*
* 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->bld_base.base.undef;
}
switch (inst->Instruction.Opcode) {
case TGSI_OPCODE_ARL:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_floor(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_MOV:
dst0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
break;
case TGSI_OPCODE_LIT:
return FALSE;
case TGSI_OPCODE_RCP:
/* TGSI_OPCODE_RECIP */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_rcp(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_RSQ:
/* TGSI_OPCODE_RECIPSQRT */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = lp_build_emit_llvm_unary(&bld->bld_base, TGSI_OPCODE_ABS, src0);
dst0 = lp_build_rsqrt(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_EXP:
return FALSE;
case TGSI_OPCODE_LOG:
return FALSE;
case TGSI_OPCODE_MUL:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_mul(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_ADD:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_add(&bld->bld_base.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 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_max(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_MAX:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_max(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_SLT:
/* TGSI_OPCODE_SETLT */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_SGE:
/* TGSI_OPCODE_SETGE */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_MAD:
/* TGSI_OPCODE_MADD */
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp0 = lp_build_mul(&bld->bld_base.base, src0, src1);
dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
break;
case TGSI_OPCODE_SUB:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
dst0 = lp_build_sub(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_LRP:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp0 = lp_build_sub(&bld->bld_base.base, src1, src2);
tmp0 = lp_build_mul(&bld->bld_base.base, src0, tmp0);
dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
break;
case TGSI_OPCODE_CND:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp1 = lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, 0.5);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src2, tmp1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, src0, src1);
break;
case TGSI_OPCODE_DP2A:
return FALSE;
case TGSI_OPCODE_FRC:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = lp_build_floor(&bld->bld_base.base, src0);
dst0 = lp_build_sub(&bld->bld_base.base, src0, tmp0);
break;
case TGSI_OPCODE_CLAMP:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp0 = lp_build_max(&bld->bld_base.base, src0, src1);
dst0 = lp_build_min(&bld->bld_base.base, tmp0, src2);
break;
case TGSI_OPCODE_FLR:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_floor(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_ROUND:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_round(&bld->bld_base.base, src0);
break;
case TGSI_OPCODE_EX2:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = lp_build_swizzle_scalar_aos(&bld->bld_base.base, src0, TGSI_SWIZZLE_X, TGSI_NUM_CHANNELS);
dst0 = lp_build_exp2(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_LG2:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_log2(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_POW:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
dst0 = lp_build_pow(&bld->bld_base.base, src0, src1);
break;
case TGSI_OPCODE_XPD:
return FALSE;
case TGSI_OPCODE_RCC:
/* deprecated? */
assert(0);
return FALSE;
case TGSI_OPCODE_DPH:
return FALSE;
case TGSI_OPCODE_COS:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_cos(&bld->bld_base.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 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_EQUAL, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_SFL:
dst0 = bld->bld_base.base.zero;
break;
case TGSI_OPCODE_SGT:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_SIN:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
dst0 = lp_build_sin(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_SLE:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LEQUAL, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_SNE:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_NOTEQUAL, src0, src1);
dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
break;
case TGSI_OPCODE_STR:
dst0 = bld->bld_base.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 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_round(&bld->bld_base.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 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_sgn(&bld->bld_base.base, tmp0);
break;
case TGSI_OPCODE_CMP:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, bld->bld_base.base.zero);
dst0 = lp_build_select(&bld->bld_base.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 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_ceil(&bld->bld_base.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 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
dst0 = lp_build_trunc(&bld->bld_base.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) {
lp_emit_store_aos(bld, inst, 0, dst0);
}
return TRUE;
}
LLVMValueRef
lp_build_swizzle_aos(struct lp_build_context *bld,
LLVMValueRef a,
const unsigned char swizzles[4])
{
LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
const unsigned n = type.length;
unsigned i, j;
if (swizzles[0] == PIPE_SWIZZLE_RED &&
swizzles[1] == PIPE_SWIZZLE_GREEN &&
swizzles[2] == PIPE_SWIZZLE_BLUE &&
swizzles[3] == PIPE_SWIZZLE_ALPHA) {
return a;
}
if (swizzles[0] == swizzles[1] &&
swizzles[1] == swizzles[2] &&
swizzles[2] == swizzles[3]) {
switch (swizzles[0]) {
case PIPE_SWIZZLE_RED:
case PIPE_SWIZZLE_GREEN:
case PIPE_SWIZZLE_BLUE:
case PIPE_SWIZZLE_ALPHA:
return lp_build_swizzle_scalar_aos(bld, a, swizzles[0]);
case PIPE_SWIZZLE_ZERO:
return bld->zero;
case PIPE_SWIZZLE_ONE:
return bld->one;
case LP_BLD_SWIZZLE_DONTCARE:
return bld->undef;
default:
assert(0);
return bld->undef;
}
}
if (type.width >= 16) {
/*
* Shuffle.
*/
LLVMValueRef undef = LLVMGetUndef(lp_build_elem_type(bld->gallivm, type));
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
LLVMValueRef aux[LP_MAX_VECTOR_LENGTH];
memset(aux, 0, sizeof aux);
for(j = 0; j < n; j += 4) {
for(i = 0; i < 4; ++i) {
unsigned shuffle;
switch (swizzles[i]) {
default:
assert(0);
/* fall through */
case PIPE_SWIZZLE_RED:
case PIPE_SWIZZLE_GREEN:
case PIPE_SWIZZLE_BLUE:
case PIPE_SWIZZLE_ALPHA:
shuffle = j + swizzles[i];
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
break;
case PIPE_SWIZZLE_ZERO:
shuffle = type.length + 0;
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
if (!aux[0]) {
aux[0] = lp_build_const_elem(bld->gallivm, type, 0.0);
}
break;
case PIPE_SWIZZLE_ONE:
shuffle = type.length + 1;
shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
if (!aux[1]) {
aux[1] = lp_build_const_elem(bld->gallivm, type, 1.0);
}
break;
case LP_BLD_SWIZZLE_DONTCARE:
shuffles[j + i] = LLVMGetUndef(i32t);
break;
}
}
}
for (i = 0; i < n; ++i) {
if (!aux[i]) {
aux[i] = undef;
}
}
return LLVMBuildShuffleVector(builder, a,
LLVMConstVector(aux, n),
LLVMConstVector(shuffles, n), "");
} else {
/*
* Bit mask and shifts.
*
* For example, this will convert BGRA to RGBA by doing
*
* rgba = (bgra & 0x00ff0000) >> 16
* | (bgra & 0xff00ff00)
* | (bgra & 0x000000ff) << 16
*
* This is necessary not only for faster cause, but because X86 backend
* will refuse shuffles of <4 x i8> vectors
*/
LLVMValueRef res;
struct lp_type type4;
unsigned cond = 0;
unsigned chan;
int shift;
/*
* Start with a mixture of 1 and 0.
*/
for (chan = 0; chan < 4; ++chan) {
if (swizzles[chan] == PIPE_SWIZZLE_ONE) {
cond |= 1 << chan;
}
}
res = lp_build_select_aos(bld, cond, bld->one, bld->zero);
/*
* Build a type where each element is an integer that cover the four
* channels.
*/
type4 = type;
type4.floating = FALSE;
type4.width *= 4;
type4.length /= 4;
a = LLVMBuildBitCast(builder, a, lp_build_vec_type(bld->gallivm, type4), "");
res = LLVMBuildBitCast(builder, res, lp_build_vec_type(bld->gallivm, type4), "");
/*
* Mask and shift the channels, trying to group as many channels in the
* same shift as possible
*/
for (shift = -3; shift <= 3; ++shift) {
unsigned long long mask = 0;
assert(type4.width <= sizeof(mask)*8);
for (chan = 0; chan < 4; ++chan) {
/* FIXME: big endian */
if (swizzles[chan] < 4 &&
chan - swizzles[chan] == shift) {
mask |= ((1ULL << type.width) - 1) << (swizzles[chan] * type.width);
}
}
if (mask) {
LLVMValueRef masked;
LLVMValueRef shifted;
if (0)
debug_printf("shift = %i, mask = 0x%08llx\n", shift, mask);
masked = LLVMBuildAnd(builder, a,
lp_build_const_int_vec(bld->gallivm, type4, mask), "");
if (shift > 0) {
shifted = LLVMBuildShl(builder, masked,
lp_build_const_int_vec(bld->gallivm, type4, shift*type.width), "");
} else if (shift < 0) {
shifted = LLVMBuildLShr(builder, masked,
lp_build_const_int_vec(bld->gallivm, type4, -shift*type.width), "");
} else {
shifted = masked;
}
res = LLVMBuildOr(builder, res, shifted, "");
}
}
return LLVMBuildBitCast(builder, res,
lp_build_vec_type(bld->gallivm, type), "");
}
}
