/* 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 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: */
LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
};
LLVMValueRef msb =
lp_build_intrinsic(builder, "llvm.ctlz.i32",
emit_data->dst_type, args, ARRAY_SIZE(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, "");
}
static LLVMValueRef
llvm_face_select_helper(
struct radeon_llvm_context * ctx,
const char *intrinsic, unsigned face_register,
unsigned frontcolor_register, unsigned backcolor_regiser)
{
LLVMValueRef backcolor = llvm_load_input_helper(
ctx,
intrinsic,
backcolor_regiser);
LLVMValueRef front_color = llvm_load_input_helper(
ctx,
intrinsic,
frontcolor_register);
LLVMValueRef face = llvm_load_input_helper(
ctx,
"llvm.R600.load.input",
face_register);
LLVMValueRef is_face_positive = LLVMBuildFCmp(
ctx->soa.bld_base.base.gallivm->builder,
LLVMRealUGT, face,
lp_build_const_float(ctx->soa.bld_base.base.gallivm, 0.0f),
"");
return LLVMBuildSelect(
ctx->soa.bld_base.base.gallivm->builder,
is_face_positive,
front_color,
backcolor,
"");
}
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)
{
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
LLVMRealPredicate pred;
LLVMValueRef cond;
/* Use ordered for everything but NE (which is usual for
* float comparisons)
*/
switch (emit_data->inst->Instruction.Opcode) {
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 = LLVMRealOGT; break;
default: assert(!"unknown instruction"); pred = 0; break;
}
cond = LLVMBuildFCmp(builder,
pred, emit_data->args[0], emit_data->args[1], "");
emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
cond, bld_base->base.one, bld_base->base.zero, "");
}
/* 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;
LLVMBuilderRef builder = gallivm->builder;
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. However, LLVM also assumes that ffs(x) is
* in [0, 31], but GLSL expects that ffs(0) = -1, so
* a conditional assignment to handle 0 is still required.
*/
LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
};
LLVMValueRef lsb =
lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
emit_data->dst_type, args, ARRAY_SIZE(args),
LLVMReadNoneAttribute);
/* TODO: We need an intrinsic to skip this conditional. */
/* 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), lsb, "");
}
LLVMValueRef
ac_build_imsb(struct ac_llvm_context *ctx,
LLVMValueRef arg,
LLVMTypeRef dst_type)
{
const char *intr_name = (HAVE_LLVM < 0x0400) ? "llvm.AMDGPU.flbit.i32" :
"llvm.amdgcn.sffbh.i32";
LLVMValueRef msb = ac_build_intrinsic(ctx, intr_name,
dst_type, &arg, 1,
AC_FUNC_ATTR_READNONE);
/* The HW returns the last bit index from MSB, but NIR/TGSI wants
* the index from LSB. Invert it by doing "31 - msb". */
msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
msb, "");
LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
LLVMValueRef cond = LLVMBuildOr(ctx->builder,
LLVMBuildICmp(ctx->builder, LLVMIntEQ,
arg, LLVMConstInt(ctx->i32, 0, 0), ""),
LLVMBuildICmp(ctx->builder, LLVMIntEQ,
arg, all_ones, ""), "");
return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
}
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:
case TGSI_OPCODE_I64MAX:
op = LLVMIntSGT;
break;
case TGSI_OPCODE_IMIN:
case TGSI_OPCODE_I64MIN:
op = LLVMIntSLT;
break;
case TGSI_OPCODE_UMAX:
case TGSI_OPCODE_U64MAX:
op = LLVMIntUGT;
break;
case TGSI_OPCODE_UMIN:
case TGSI_OPCODE_U64MIN:
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], "");
}
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, "");
}
/* 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)
{
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, "");
}
/**
* Build a manual selection sequence for cube face sc/tc coordinates and
* major axis vector (multiplied by 2 for consistency) for the given
* vec3 \p coords, for the face implied by \p selcoords.
*
* For the major axis, we always adjust the sign to be in the direction of
* selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
* the selcoords major axis.
*/
static void build_cube_select(LLVMBuilderRef builder,
const struct cube_selection_coords *selcoords,
const LLVMValueRef *coords,
LLVMValueRef *out_st,
LLVMValueRef *out_ma)
{
LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
LLVMValueRef is_ma_positive;
LLVMValueRef sgn_ma;
LLVMValueRef is_ma_z, is_not_ma_z;
LLVMValueRef is_ma_y;
LLVMValueRef is_ma_x;
LLVMValueRef sgn;
LLVMValueRef tmp;
is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
selcoords->ma, LLVMConstReal(f32, 0.0), "");
sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
/* Select sc */
tmp = LLVMBuildSelect(builder, is_ma_z, coords[2], coords[0], "");
sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
LLVMBuildSelect(builder, is_ma_x, sgn_ma,
LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
/* Select tc */
tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
sgn = LLVMBuildSelect(builder, is_ma_y, LLVMBuildFNeg(builder, sgn_ma, ""),
LLVMConstReal(f32, -1.0), "");
out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
/* Select ma */
tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
sgn = LLVMBuildSelect(builder, is_ma_positive,
LLVMConstReal(f32, 2.0), LLVMConstReal(f32, -2.0), "");
*out_ma = LLVMBuildFMul(builder, tmp, sgn, "");
}
static LLVMValueRef
llvm_face_select_helper(
struct radeon_llvm_context * ctx,
LLVMValueRef face, LLVMValueRef front_color, LLVMValueRef back_color)
{
const struct lp_build_context * bb = &ctx->soa.bld_base.base;
LLVMValueRef is_front = LLVMBuildFCmp(
bb->gallivm->builder, LLVMRealUGT, face,
lp_build_const_float(bb->gallivm, 0.0f), "");
return LLVMBuildSelect(bb->gallivm->builder, is_front,
front_color, back_color, "");
}
static void emit_ssg(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 cmp, val;
if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_I64SSG) {
cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int64_bld.zero, "");
val = LLVMBuildSelect(builder, cmp, bld_base->int64_bld.one, emit_data->args[0], "");
cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int64_bld.zero, "");
val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int64_bld.elem_type, -1, true), "");
} else if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
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, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
}
emit_data->output[emit_data->chan] = val;
}
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_cndlt(
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 float_zero = lp_build_const_float(
bld_base->base.gallivm, 0.0f);
LLVMValueRef cmp = LLVMBuildFCmp(
builder, LLVMRealULT, emit_data->args[0], float_zero, "");
emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
cmp, emit_data->args[1], emit_data->args[2], "");
}
SCM llvm_build_select(SCM scm_function, SCM scm_condition, SCM scm_value_if, SCM scm_value_else)
{
SCM retval;
struct llvm_value_t *self;
self = (struct llvm_value_t *)scm_gc_calloc(sizeof(struct llvm_value_t), "llvm value");
SCM_NEWSMOB(retval, llvm_value_tag, self);
struct llvm_function_t *function = get_llvm_function(scm_function);
struct llvm_value_t *condition = get_llvm_value(scm_condition);
struct llvm_value_t *value_if = get_llvm_value(scm_value_if );
struct llvm_value_t *value_else = get_llvm_value(scm_value_else);
self->value = LLVMBuildSelect(function->builder, condition->value, value_if->value, value_else->value, "x");
return retval;
}
/**
* For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
* two (adjacent) mipmap level indexes. Later, we'll sample from those
* two mipmap levels and interpolate between them.
*/
void
lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
unsigned unit,
LLVMValueRef lod_ipart,
LLVMValueRef *lod_fpart_inout,
LLVMValueRef *level0_out,
LLVMValueRef *level1_out)
{
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context *int_bld = &bld->int_bld;
struct lp_build_context *float_bld = &bld->float_bld;
LLVMValueRef last_level;
LLVMValueRef clamp_min;
LLVMValueRef clamp_max;
*level0_out = lod_ipart;
*level1_out = lp_build_add(int_bld, lod_ipart, int_bld->one);
last_level = bld->dynamic_state->last_level(bld->dynamic_state,
bld->gallivm, unit);
/*
* Clamp both lod_ipart and lod_ipart + 1 to [0, last_level], with the
* minimum number of comparisons, and zeroing lod_fpart in the extreme
* ends in the process.
*/
/* lod_ipart < 0 */
clamp_min = LLVMBuildICmp(builder, LLVMIntSLT,
lod_ipart, int_bld->zero,
"clamp_lod_to_zero");
*level0_out = LLVMBuildSelect(builder, clamp_min,
int_bld->zero, *level0_out, "");
*level1_out = LLVMBuildSelect(builder, clamp_min,
int_bld->zero, *level1_out, "");
*lod_fpart_inout = LLVMBuildSelect(builder, clamp_min,
float_bld->zero, *lod_fpart_inout, "");
/* lod_ipart >= last_level */
clamp_max = LLVMBuildICmp(builder, LLVMIntSGE,
lod_ipart, last_level,
"clamp_lod_to_last");
*level0_out = LLVMBuildSelect(builder, clamp_max,
last_level, *level0_out, "");
*level1_out = LLVMBuildSelect(builder, clamp_max,
last_level, *level1_out, "");
*lod_fpart_inout = LLVMBuildSelect(builder, clamp_max,
float_bld->zero, *lod_fpart_inout, "");
lp_build_name(*level0_out, "sampler%u_miplevel0", unit);
lp_build_name(*level1_out, "sampler%u_miplevel1", unit);
lp_build_name(*lod_fpart_inout, "sampler%u_mipweight", unit);
}
/** Helper used by lp_build_cube_lookup()
* Return (major_coord >= 0) ? pos_face : neg_face;
*/
static LLVMValueRef
lp_build_cube_face(struct lp_build_sample_context *bld,
LLVMValueRef major_coord,
unsigned pos_face, unsigned neg_face)
{
struct gallivm_state *gallivm = bld->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef cmp = LLVMBuildFCmp(builder, LLVMRealUGE,
major_coord,
bld->float_bld.zero, "");
LLVMValueRef pos = lp_build_const_int32(gallivm, pos_face);
LLVMValueRef neg = lp_build_const_int32(gallivm, neg_face);
LLVMValueRef res = LLVMBuildSelect(builder, cmp, pos, neg, "");
return res;
}
void build_mapgrant_item(
struct llvm_ctx *ctx,
LLVMValueRef *dest,
LLVMValueRef send_base,
LLVMValueRef fpage,
LLVMValueRef is_grant)
{
V grant_cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
CONST_WORD(0), is_grant, "is.grant.cond");
dest[0] = LLVMBuildOr(ctx->builder, send_base,
LLVMBuildSelect(ctx->builder, grant_cond,
CONST_WORD(0xa), CONST_WORD(0x8), "mgitem.type.bits"),
"mapgrant.info.word");
dest[1] = fpage;
}
static void emit_ucmp(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 arg0 = LLVMBuildBitCast(builder, emit_data->args[0],
bld_base->uint_bld.elem_type, "");
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], "");
}
LLVMValueRef gen_cond(struct node *ast)
{
LLVMValueRef truth;
truth = LLVMBuildICmp(builder,
LLVMIntNE,
codegen(ast->one),
CONST(0),
"");
return LLVMBuildSelect(builder,
truth,
codegen(ast->two),
codegen(ast->three),
"");
}
static void emit_bfe(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 bfe_sm5;
LLVMValueRef cond;
bfe_sm5 = lp_build_intrinsic(builder, action->intr_name,
emit_data->dst_type, emit_data->args,
emit_data->arg_count, LLVMReadNoneAttribute);
/* Correct for GLSL semantics. */
cond = LLVMBuildICmp(builder, LLVMIntUGE, emit_data->args[2],
lp_build_const_int32(gallivm, 32), "");
emit_data->output[emit_data->chan] =
LLVMBuildSelect(builder, cond, emit_data->args[0], bfe_sm5, "");
}
/* Perform front/back face culling and return true if the primitive is accepted. */
static LLVMValueRef ac_cull_face(struct ac_llvm_context *ctx,
LLVMValueRef pos[3][4],
struct ac_position_w_info *w,
bool cull_front,
bool cull_back,
bool cull_zero_area)
{
LLVMBuilderRef builder = ctx->builder;
if (cull_front && cull_back)
return ctx->i1false;
if (!cull_front && !cull_back && !cull_zero_area)
return ctx->i1true;
/* Front/back face culling. Also if the determinant == 0, the triangle
* area is 0.
*/
LLVMValueRef det_t0 = LLVMBuildFSub(builder, pos[2][0], pos[0][0], "");
LLVMValueRef det_t1 = LLVMBuildFSub(builder, pos[1][1], pos[0][1], "");
LLVMValueRef det_t2 = LLVMBuildFSub(builder, pos[0][0], pos[1][0], "");
LLVMValueRef det_t3 = LLVMBuildFSub(builder, pos[0][1], pos[2][1], "");
LLVMValueRef det_p0 = LLVMBuildFMul(builder, det_t0, det_t1, "");
LLVMValueRef det_p1 = LLVMBuildFMul(builder, det_t2, det_t3, "");
LLVMValueRef det = LLVMBuildFSub(builder, det_p0, det_p1, "");
/* Negative W negates the determinant. */
det = LLVMBuildSelect(builder, w->w_reflection,
LLVMBuildFNeg(builder, det, ""),
det, "");
LLVMValueRef accepted = NULL;
if (cull_front) {
LLVMRealPredicate cond = cull_zero_area ? LLVMRealOGT : LLVMRealOGE;
accepted = LLVMBuildFCmp(builder, cond, det, ctx->f32_0, "");
} else if (cull_back) {
LLVMRealPredicate cond = cull_zero_area ? LLVMRealOLT : LLVMRealOLE;
accepted = LLVMBuildFCmp(builder, cond, det, ctx->f32_0, "");
} else if (cull_zero_area) {
accepted = LLVMBuildFCmp(builder, LLVMRealONE, det, ctx->f32_0, "");
}
return accepted;
}
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];
LLVMValueRef bfi_sm5;
LLVMValueRef cond;
// 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], "");
bfi_args[1] = LLVMBuildShl(builder, emit_data->args[1],
emit_data->args[2], "");
bfi_args[2] = emit_data->args[0];
/* Calculate:
* (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
* Use the right-hand side, which the LLVM backend can convert to V_BFI.
*/
bfi_sm5 =
LLVMBuildXor(builder, bfi_args[2],
LLVMBuildAnd(builder, bfi_args[0],
LLVMBuildXor(builder, bfi_args[1], bfi_args[2],
""), ""), "");
/* Since shifts of >= 32 bits are undefined in LLVM IR, the backend
* uses the convenient V_BFI lowering for the above, which follows SM5
* and disagrees with GLSL semantics when bits (src3) is 32.
*/
cond = LLVMBuildICmp(builder, LLVMIntUGE, emit_data->args[3],
lp_build_const_int32(gallivm, 32), "");
emit_data->output[emit_data->chan] =
LLVMBuildSelect(builder, cond, emit_data->args[1], bfi_sm5, "");
}
LLVMValueRef
ac_build_umsb(struct ac_llvm_context *ctx,
LLVMValueRef arg,
LLVMTypeRef dst_type)
{
LLVMValueRef args[2] = {
arg,
LLVMConstInt(ctx->i1, 1, 0),
};
LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.ctlz.i32",
dst_type, args, ARRAY_SIZE(args),
AC_FUNC_ATTR_READNONE);
/* The HW returns the last bit index from MSB, but TGSI/NIR wants
* the index from LSB. Invert it by doing "31 - msb". */
msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
msb, "");
/* check for zero */
return LLVMBuildSelect(ctx->builder,
LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg,
LLVMConstInt(ctx->i32, 0, 0), ""),
LLVMConstInt(ctx->i32, -1, true), msb, "");
}
/**
* Build code to compare two values 'a' and 'b' of 'type' using the given func.
* \param func one of PIPE_FUNC_x
* The result values will be 0 for false or ~0 for true.
*/
LLVMValueRef
lp_build_compare(struct gallivm_state *gallivm,
const struct lp_type type,
unsigned func,
LLVMValueRef a,
LLVMValueRef b)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, type);
LLVMValueRef zeros = LLVMConstNull(int_vec_type);
LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
LLVMValueRef cond;
LLVMValueRef res;
assert(func >= PIPE_FUNC_NEVER);
assert(func <= PIPE_FUNC_ALWAYS);
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if(func == PIPE_FUNC_NEVER)
return zeros;
if(func == PIPE_FUNC_ALWAYS)
return ones;
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
/*
* There are no unsigned integer comparison instructions in SSE.
*/
if (!type.floating && !type.sign &&
type.width * type.length == 128 &&
util_cpu_caps.has_sse2 &&
(func == PIPE_FUNC_LESS ||
func == PIPE_FUNC_LEQUAL ||
func == PIPE_FUNC_GREATER ||
func == PIPE_FUNC_GEQUAL) &&
(gallivm_debug & GALLIVM_DEBUG_PERF)) {
debug_printf("%s: inefficient <%u x i%u> unsigned comparison\n",
__FUNCTION__, type.length, type.width);
}
#endif
#if HAVE_LLVM < 0x0207
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
if(type.width * type.length == 128) {
if(type.floating && util_cpu_caps.has_sse) {
/* float[4] comparison */
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
LLVMValueRef args[3];
unsigned cc;
boolean swap;
swap = FALSE;
switch(func) {
case PIPE_FUNC_EQUAL:
cc = 0;
break;
case PIPE_FUNC_NOTEQUAL:
cc = 4;
break;
case PIPE_FUNC_LESS:
cc = 1;
break;
case PIPE_FUNC_LEQUAL:
cc = 2;
break;
case PIPE_FUNC_GREATER:
cc = 1;
swap = TRUE;
break;
case PIPE_FUNC_GEQUAL:
cc = 2;
swap = TRUE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
if(swap) {
args[0] = b;
args[1] = a;
}
else {
args[0] = a;
args[1] = b;
}
args[2] = LLVMConstInt(LLVMInt8TypeInContext(gallivm->context), cc, 0);
res = lp_build_intrinsic(builder,
"llvm.x86.sse.cmp.ps",
vec_type,
args, 3);
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
return res;
}
else if(util_cpu_caps.has_sse2) {
/* int[4] comparison */
static const struct {
unsigned swap:1;
unsigned eq:1;
unsigned gt:1;
unsigned not:1;
} table[] = {
{0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
{1, 0, 1, 0}, /* PIPE_FUNC_LESS */
{0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
{0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
{0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
{0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
{1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
{0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
};
const char *pcmpeq;
const char *pcmpgt;
LLVMValueRef args[2];
LLVMValueRef res;
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
switch (type.width) {
case 8:
pcmpeq = "llvm.x86.sse2.pcmpeq.b";
pcmpgt = "llvm.x86.sse2.pcmpgt.b";
break;
case 16:
pcmpeq = "llvm.x86.sse2.pcmpeq.w";
pcmpgt = "llvm.x86.sse2.pcmpgt.w";
break;
case 32:
pcmpeq = "llvm.x86.sse2.pcmpeq.d";
pcmpgt = "llvm.x86.sse2.pcmpgt.d";
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
/* There are no unsigned comparison instructions. So flip the sign bit
* so that the results match.
*/
if (table[func].gt && !type.sign) {
LLVMValueRef msb = lp_build_const_int_vec(gallivm, type, (unsigned long long)1 << (type.width - 1));
a = LLVMBuildXor(builder, a, msb, "");
b = LLVMBuildXor(builder, b, msb, "");
}
if(table[func].swap) {
args[0] = b;
args[1] = a;
}
else {
args[0] = a;
args[1] = b;
}
if(table[func].eq)
res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
else if (table[func].gt)
res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
else
res = LLVMConstNull(vec_type);
if(table[func].not)
res = LLVMBuildNot(builder, res, "");
return res;
}
} /* if (type.width * type.length == 128) */
#endif
#endif /* HAVE_LLVM < 0x0207 */
/* XXX: It is not clear if we should use the ordered or unordered operators */
if(type.floating) {
LLVMRealPredicate op;
switch(func) {
case PIPE_FUNC_NEVER:
op = LLVMRealPredicateFalse;
break;
case PIPE_FUNC_ALWAYS:
op = LLVMRealPredicateTrue;
break;
case PIPE_FUNC_EQUAL:
op = LLVMRealUEQ;
break;
case PIPE_FUNC_NOTEQUAL:
op = LLVMRealUNE;
break;
case PIPE_FUNC_LESS:
op = LLVMRealULT;
break;
case PIPE_FUNC_LEQUAL:
op = LLVMRealULE;
break;
case PIPE_FUNC_GREATER:
op = LLVMRealUGT;
break;
case PIPE_FUNC_GEQUAL:
op = LLVMRealUGE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
#if HAVE_LLVM >= 0x0207
cond = LLVMBuildFCmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
if (type.length == 1) {
cond = LLVMBuildFCmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
unsigned i;
res = LLVMGetUndef(int_vec_type);
debug_printf("%s: warning: using slow element-wise float"
" vector comparison\n", __FUNCTION__);
for (i = 0; i < type.length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
cond = LLVMBuildFCmp(builder, op,
LLVMBuildExtractElement(builder, a, index, ""),
LLVMBuildExtractElement(builder, b, index, ""),
"");
cond = LLVMBuildSelect(builder, cond,
LLVMConstExtractElement(ones, index),
LLVMConstExtractElement(zeros, index),
"");
res = LLVMBuildInsertElement(builder, res, cond, index, "");
}
}
#endif
}
else {
LLVMIntPredicate op;
switch(func) {
case PIPE_FUNC_EQUAL:
op = LLVMIntEQ;
break;
case PIPE_FUNC_NOTEQUAL:
op = LLVMIntNE;
break;
case PIPE_FUNC_LESS:
op = type.sign ? LLVMIntSLT : LLVMIntULT;
break;
case PIPE_FUNC_LEQUAL:
op = type.sign ? LLVMIntSLE : LLVMIntULE;
break;
case PIPE_FUNC_GREATER:
op = type.sign ? LLVMIntSGT : LLVMIntUGT;
break;
case PIPE_FUNC_GEQUAL:
op = type.sign ? LLVMIntSGE : LLVMIntUGE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
#if HAVE_LLVM >= 0x0207
cond = LLVMBuildICmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
if (type.length == 1) {
cond = LLVMBuildICmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
unsigned i;
res = LLVMGetUndef(int_vec_type);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: using slow element-wise int"
" vector comparison\n", __FUNCTION__);
}
for(i = 0; i < type.length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
cond = LLVMBuildICmp(builder, op,
LLVMBuildExtractElement(builder, a, index, ""),
LLVMBuildExtractElement(builder, b, index, ""),
"");
cond = LLVMBuildSelect(builder, cond,
LLVMConstExtractElement(ones, index),
LLVMConstExtractElement(zeros, index),
"");
res = LLVMBuildInsertElement(builder, res, cond, index, "");
}
}
#endif
}
return res;
}
static LLVMValueRef get_stritem_len_fn(struct llvm_ctx *ctx)
{
if(ctx->stritem_len_fn != NULL) return ctx->stritem_len_fn;
/* returns (i32 len, i32 new_tpos)
* params (word *utcbptr, i32 tpos)
*
* when return value "new_tpos" > tmax + 1, the result is invalid. the function
* should also not be called when tpos > tmax + 1.
*/
LLVMTypeRef ret_types[2] = { ctx->i32t, ctx->i32t },
parm_types[2] = { LLVMPointerType(ctx->wordt, 0), ctx->i32t },
ret_type = LLVMStructTypeInContext(ctx->ctx, ret_types, 2, 0),
fn_type = LLVMFunctionType(ret_type, parm_types, 2, 0);
LLVMValueRef fn = LLVMAddFunction(ctx->module, "__muidl_get_stritem_len",
fn_type);
LLVMSetVisibility(fn, LLVMHiddenVisibility);
LLVMSetLinkage(fn, LLVMInternalLinkage);
V fn_args[2];
LLVMGetParams(fn, fn_args);
LLVMAddAttribute(fn_args[0], LLVMNoCaptureAttribute);
for(int i=0; i<2; i++) {
LLVMAddAttribute(fn_args[i], LLVMInRegAttribute);
}
ctx->stritem_len_fn = fn;
LLVMBuilderRef old_builder = ctx->builder;
ctx->builder = LLVMCreateBuilderInContext(ctx->ctx);
LLVMBasicBlockRef entry_bb = LLVMAppendBasicBlockInContext(ctx->ctx, fn,
"EntryBlock"),
loop_bb = LLVMAppendBasicBlockInContext(ctx->ctx, fn, "loop"),
valid_bb = LLVMAppendBasicBlockInContext(ctx->ctx, fn, "valid"),
exit_bb = LLVMAppendBasicBlockInContext(ctx->ctx, fn, "exit");
LLVMPositionBuilderAtEnd(ctx->builder, entry_bb);
LLVMValueRef old_utcb = ctx->utcb, old_tpos = ctx->tpos;
ctx->utcb = fn_args[0];
ctx->tpos = fn_args[1];
LLVMBuildBr(ctx->builder, loop_bb);
LLVMPositionBuilderAtEnd(ctx->builder, exit_bb);
LLVMValueRef exit_len_phi = LLVMBuildPhi(ctx->builder, ctx->i32t,
"exit.len.phi"),
exit_tpos_phi = LLVMBuildPhi(ctx->builder, ctx->i32t,
"exit.tpos.phi");
LLVMValueRef rvals[2] = { exit_len_phi, exit_tpos_phi };
LLVMBuildAggregateRet(ctx->builder, rvals, 2);
LLVMPositionBuilderAtEnd(ctx->builder, loop_bb);
LLVMValueRef len_phi = LLVMBuildPhi(ctx->builder, ctx->i32t, "len.phi"),
tpos_phi = LLVMBuildPhi(ctx->builder, ctx->i32t, "tpos.phi");
LLVMAddIncoming(len_phi, &ctx->zero, &entry_bb, 1);
LLVMAddIncoming(tpos_phi, &ctx->tpos, &entry_bb, 1);
ctx->tpos = tpos_phi;
/* test: if *tpos doesn't look like a string item, conk out. */
LLVMValueRef infoword = build_utcb_load(ctx, ctx->tpos, "si.info");
LLVMValueRef is_cond = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
ctx->zero, LLVMBuildAnd(ctx->builder, infoword,
CONST_WORD(1 << 4), "infoword.si.mask"),
"infoword.si.cond");
/* anything + 100 is sure to be > tmax + 1. */
LLVMValueRef fucked_tpos = LLVMBuildAdd(ctx->builder, tpos_phi,
CONST_INT(100), "f****d.tpos");
branch_set_phi(ctx, exit_len_phi, len_phi);
branch_set_phi(ctx, exit_tpos_phi, fucked_tpos);
LLVMBuildCondBr(ctx->builder, is_cond, valid_bb, exit_bb);
LLVMPositionBuilderAtEnd(ctx->builder, valid_bb);
LLVMValueRef string_length = LLVMBuildTruncOrBitCast(ctx->builder,
LLVMBuildLShr(ctx->builder, infoword,
CONST_INT(10), "si.info.len"),
ctx->i32t, "si.info.len.int"),
string_j = LLVMBuildTruncOrBitCast(ctx->builder,
LLVMBuildAnd(ctx->builder, CONST_WORD(0x1f),
LLVMBuildLShr(ctx->builder, infoword, CONST_WORD(4),
"si.info.j.shift"),
"si.info.j.masked"),
ctx->i32t, "si.info.j"),
string_c = LLVMBuildTruncOrBitCast(ctx->builder,
LLVMBuildAnd(ctx->builder, CONST_WORD(1 << 9),
infoword, "si.info.c.masked"),
ctx->i32t, "si.info.c.masked.int"),
c_cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
string_c, CONST_WORD(0), "si.info.c.cond"),
new_len = LLVMBuildAdd(ctx->builder, len_phi,
LLVMBuildMul(ctx->builder, string_length,
LLVMBuildAdd(ctx->builder, string_j,
CONST_INT(1), "j.plus.one"),
"len.incr"),
"len.new"),
new_tpos = LLVMBuildAdd(ctx->builder, ctx->tpos,
LLVMBuildSelect(ctx->builder, c_cond,
LLVMBuildAdd(ctx->builder, CONST_INT(2),
string_j, "cont.tpos.bump"),
CONST_INT(2), "tpos.bump"),
"tpos.new");
LLVMAddIncoming(len_phi, &new_len, &valid_bb, 1);
LLVMAddIncoming(tpos_phi, &new_tpos, &valid_bb, 1);
LLVMAddIncoming(exit_len_phi, &new_len, &valid_bb, 1);
LLVMAddIncoming(exit_tpos_phi, &new_tpos, &valid_bb, 1);
LLVMBuildCondBr(ctx->builder, c_cond, loop_bb, exit_bb);
LLVMDisposeBuilder(ctx->builder);
ctx->builder = old_builder;
ctx->utcb = old_utcb;
ctx->tpos = old_tpos;
return ctx->stritem_len_fn;
}
/**
* Return mask ? a : b;
*
* mask is a bitwise mask, composed of 0 or ~0 for each element. Any other value
* will yield unpredictable results.
*/
LLVMValueRef
lp_build_select(struct lp_build_context *bld,
LLVMValueRef mask,
LLVMValueRef a,
LLVMValueRef b)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMContextRef lc = bld->gallivm->context;
struct lp_type type = bld->type;
LLVMValueRef res;
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if(a == b)
return a;
if (type.length == 1) {
mask = LLVMBuildTrunc(builder, mask, LLVMInt1TypeInContext(lc), "");
res = LLVMBuildSelect(builder, mask, a, b, "");
}
else if (0) {
/* Generate a vector select.
*
* XXX: Using vector selects would avoid emitting intrinsics, but they aren't
* properly supported yet.
*
* LLVM 3.0 includes experimental support provided the -promote-elements
* options is passed to LLVM's command line (e.g., via
* llvm::cl::ParseCommandLineOptions), but resulting code quality is much
* worse, probably because some optimization passes don't know how to
* handle vector selects.
*
* See also:
* - http://lists.cs.uiuc.edu/pipermail/llvmdev/2011-October/043659.html
*/
/* Convert the mask to a vector of booleans.
* XXX: There are two ways to do this. Decide what's best.
*/
if (1) {
LLVMTypeRef bool_vec_type = LLVMVectorType(LLVMInt1TypeInContext(lc), type.length);
mask = LLVMBuildTrunc(builder, mask, bool_vec_type, "");
} else {
mask = LLVMBuildICmp(builder, LLVMIntNE, mask, LLVMConstNull(bld->int_vec_type), "");
}
res = LLVMBuildSelect(builder, mask, a, b, "");
}
else if (((util_cpu_caps.has_sse4_1 &&
type.width * type.length == 128) ||
(util_cpu_caps.has_avx &&
type.width * type.length == 256 && type.width >= 32)) &&
!LLVMIsConstant(a) &&
!LLVMIsConstant(b) &&
!LLVMIsConstant(mask)) {
const char *intrinsic;
LLVMTypeRef arg_type;
LLVMValueRef args[3];
/*
* There's only float blend in AVX but can just cast i32/i64
* to float.
*/
if (type.width * type.length == 256) {
if (type.width == 64) {
intrinsic = "llvm.x86.avx.blendv.pd.256";
arg_type = LLVMVectorType(LLVMDoubleTypeInContext(lc), 4);
}
else {
intrinsic = "llvm.x86.avx.blendv.ps.256";
arg_type = LLVMVectorType(LLVMFloatTypeInContext(lc), 8);
}
}
else if (type.floating &&
type.width == 64) {
intrinsic = "llvm.x86.sse41.blendvpd";
arg_type = LLVMVectorType(LLVMDoubleTypeInContext(lc), 2);
} else if (type.floating &&
type.width == 32) {
intrinsic = "llvm.x86.sse41.blendvps";
arg_type = LLVMVectorType(LLVMFloatTypeInContext(lc), 4);
} else {
intrinsic = "llvm.x86.sse41.pblendvb";
arg_type = LLVMVectorType(LLVMInt8TypeInContext(lc), 16);
}
if (arg_type != bld->int_vec_type) {
mask = LLVMBuildBitCast(builder, mask, arg_type, "");
}
if (arg_type != bld->vec_type) {
a = LLVMBuildBitCast(builder, a, arg_type, "");
b = LLVMBuildBitCast(builder, b, arg_type, "");
}
args[0] = b;
args[1] = a;
args[2] = mask;
res = lp_build_intrinsic(builder, intrinsic,
arg_type, args, Elements(args));
if (arg_type != bld->vec_type) {
res = LLVMBuildBitCast(builder, res, bld->vec_type, "");
}
}
else {
res = lp_build_select_bitwise(bld, mask, a, b);
}
return res;
}
static void declare_input_fs(
struct si_shader_context * si_shader_ctx,
unsigned input_index,
const struct tgsi_full_declaration *decl)
{
struct si_shader *shader = &si_shader_ctx->shader->shader;
struct lp_build_context * base =
&si_shader_ctx->radeon_bld.soa.bld_base.base;
struct gallivm_state * gallivm = base->gallivm;
LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
LLVMValueRef main_fn = si_shader_ctx->radeon_bld.main_fn;
LLVMValueRef interp_param;
const char * intr_name;
/* This value is:
* [15:0] NewPrimMask (Bit mask for each quad. It is set it the
* quad begins a new primitive. Bit 0 always needs
* to be unset)
* [32:16] ParamOffset
*
*/
LLVMValueRef params = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_PRIM_MASK);
LLVMValueRef attr_number;
unsigned chan;
if (decl->Semantic.Name == TGSI_SEMANTIC_POSITION) {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
unsigned soa_index =
radeon_llvm_reg_index_soa(input_index, chan);
si_shader_ctx->radeon_bld.inputs[soa_index] =
LLVMGetParam(main_fn, SI_PARAM_POS_X_FLOAT + chan);
if (chan == 3)
/* RCP for fragcoord.w */
si_shader_ctx->radeon_bld.inputs[soa_index] =
LLVMBuildFDiv(gallivm->builder,
lp_build_const_float(gallivm, 1.0f),
si_shader_ctx->radeon_bld.inputs[soa_index],
"");
}
return;
}
if (decl->Semantic.Name == TGSI_SEMANTIC_FACE) {
LLVMValueRef face, is_face_positive;
face = LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE);
is_face_positive = LLVMBuildFCmp(gallivm->builder,
LLVMRealUGT, face,
lp_build_const_float(gallivm, 0.0f),
"");
si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 0)] =
LLVMBuildSelect(gallivm->builder,
is_face_positive,
lp_build_const_float(gallivm, 1.0f),
lp_build_const_float(gallivm, 0.0f),
"");
si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 1)] =
si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 2)] =
lp_build_const_float(gallivm, 0.0f);
si_shader_ctx->radeon_bld.inputs[radeon_llvm_reg_index_soa(input_index, 3)] =
lp_build_const_float(gallivm, 1.0f);
return;
}
shader->input[input_index].param_offset = shader->ninterp++;
attr_number = lp_build_const_int32(gallivm,
shader->input[input_index].param_offset);
/* XXX: Handle all possible interpolation modes */
switch (decl->Interp.Interpolate) {
case TGSI_INTERPOLATE_COLOR:
if (si_shader_ctx->shader->key.ps.flatshade) {
interp_param = 0;
} else {
if (decl->Interp.Centroid)
interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTROID);
else
interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTER);
}
break;
case TGSI_INTERPOLATE_CONSTANT:
interp_param = 0;
break;
case TGSI_INTERPOLATE_LINEAR:
if (decl->Interp.Centroid)
interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_CENTROID);
else
interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_CENTER);
break;
case TGSI_INTERPOLATE_PERSPECTIVE:
if (decl->Interp.Centroid)
interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTROID);
else
interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTER);
break;
default:
fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
return;
}
intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
/* XXX: Could there be more than TGSI_NUM_CHANNELS (4) ? */
if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR &&
si_shader_ctx->shader->key.ps.color_two_side) {
LLVMValueRef args[4];
LLVMValueRef face, is_face_positive;
LLVMValueRef back_attr_number =
lp_build_const_int32(gallivm,
shader->input[input_index].param_offset + 1);
face = LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE);
is_face_positive = LLVMBuildFCmp(gallivm->builder,
LLVMRealUGT, face,
lp_build_const_float(gallivm, 0.0f),
"");
args[2] = params;
args[3] = interp_param;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
LLVMValueRef front, back;
args[0] = llvm_chan;
args[1] = attr_number;
front = build_intrinsic(base->gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
args[1] = back_attr_number;
back = build_intrinsic(base->gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
si_shader_ctx->radeon_bld.inputs[soa_index] =
LLVMBuildSelect(gallivm->builder,
is_face_positive,
front,
back,
"");
}
shader->ninterp++;
} else {
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
LLVMValueRef args[4];
LLVMValueRef llvm_chan = lp_build_const_int32(gallivm, chan);
unsigned soa_index = radeon_llvm_reg_index_soa(input_index, chan);
args[0] = llvm_chan;
args[1] = attr_number;
args[2] = params;
args[3] = interp_param;
si_shader_ctx->radeon_bld.inputs[soa_index] =
build_intrinsic(base->gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
}
}
}
