2 * Copyright 2014 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
28 #include <llvm-c/Core.h>
30 #include "c11/threads.h"
35 #include "ac_llvm_util.h"
36 #include "ac_exp_param.h"
37 #include "util/bitscan.h"
38 #include "util/macros.h"
39 #include "util/u_atomic.h"
42 #include "shader_enums.h"
44 /* Initialize module-independent parts of the context.
46 * The caller is responsible for initializing ctx::module and ctx::builder.
49 ac_llvm_context_init(struct ac_llvm_context
*ctx
, LLVMContextRef context
,
50 enum chip_class chip_class
, enum radeon_family family
)
54 ctx
->chip_class
= chip_class
;
57 ctx
->context
= context
;
61 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
62 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
63 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
64 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
65 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
66 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
67 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
68 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
69 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
70 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
71 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
72 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
73 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
74 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
75 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
77 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
78 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
79 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
80 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
81 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
82 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
84 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
85 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
87 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
90 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
91 "invariant.load", 14);
93 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
95 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
96 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
98 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
99 "amdgpu.uniform", 14);
101 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
105 ac_get_llvm_num_components(LLVMValueRef value
)
107 LLVMTypeRef type
= LLVMTypeOf(value
);
108 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
109 ? LLVMGetVectorSize(type
)
111 return num_components
;
115 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
119 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
124 return LLVMBuildExtractElement(ac
->builder
, value
,
125 LLVMConstInt(ac
->i32
, index
, false), "");
129 ac_get_type_size(LLVMTypeRef type
)
131 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
134 case LLVMIntegerTypeKind
:
135 return LLVMGetIntTypeWidth(type
) / 8;
136 case LLVMFloatTypeKind
:
138 case LLVMDoubleTypeKind
:
139 case LLVMPointerTypeKind
:
141 case LLVMVectorTypeKind
:
142 return LLVMGetVectorSize(type
) *
143 ac_get_type_size(LLVMGetElementType(type
));
144 case LLVMArrayTypeKind
:
145 return LLVMGetArrayLength(type
) *
146 ac_get_type_size(LLVMGetElementType(type
));
153 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
155 if (t
== ctx
->f16
|| t
== ctx
->i16
)
157 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
159 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
162 unreachable("Unhandled integer size");
166 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
168 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
169 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
170 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
171 LLVMGetVectorSize(t
));
173 return to_integer_type_scalar(ctx
, t
);
177 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
179 LLVMTypeRef type
= LLVMTypeOf(v
);
180 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
183 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
185 if (t
== ctx
->i16
|| t
== ctx
->f16
)
187 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
189 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
192 unreachable("Unhandled float size");
196 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
198 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
199 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
200 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
201 LLVMGetVectorSize(t
));
203 return to_float_type_scalar(ctx
, t
);
207 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
209 LLVMTypeRef type
= LLVMTypeOf(v
);
210 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
215 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
216 LLVMTypeRef return_type
, LLVMValueRef
*params
,
217 unsigned param_count
, unsigned attrib_mask
)
219 LLVMValueRef function
, call
;
220 bool set_callsite_attrs
= HAVE_LLVM
>= 0x0400 &&
221 !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
223 function
= LLVMGetNamedFunction(ctx
->module
, name
);
225 LLVMTypeRef param_types
[32], function_type
;
228 assert(param_count
<= 32);
230 for (i
= 0; i
< param_count
; ++i
) {
232 param_types
[i
] = LLVMTypeOf(params
[i
]);
235 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
236 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
238 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
239 LLVMSetLinkage(function
, LLVMExternalLinkage
);
241 if (!set_callsite_attrs
)
242 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
245 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
246 if (set_callsite_attrs
)
247 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
252 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
255 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
257 LLVMTypeRef elem_type
= type
;
259 assert(bufsize
>= 8);
261 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
262 int ret
= snprintf(buf
, bufsize
, "v%u",
263 LLVMGetVectorSize(type
));
265 char *type_name
= LLVMPrintTypeToString(type
);
266 fprintf(stderr
, "Error building type name for: %s\n",
270 elem_type
= LLVMGetElementType(type
);
274 switch (LLVMGetTypeKind(elem_type
)) {
276 case LLVMIntegerTypeKind
:
277 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
279 case LLVMFloatTypeKind
:
280 snprintf(buf
, bufsize
, "f32");
282 case LLVMDoubleTypeKind
:
283 snprintf(buf
, bufsize
, "f64");
289 * Helper function that builds an LLVM IR PHI node and immediately adds
293 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
294 unsigned count_incoming
, LLVMValueRef
*values
,
295 LLVMBasicBlockRef
*blocks
)
297 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
298 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
302 /* Prevent optimizations (at least of memory accesses) across the current
303 * point in the program by emitting empty inline assembly that is marked as
304 * having side effects.
306 * Optionally, a value can be passed through the inline assembly to prevent
307 * LLVM from hoisting calls to ReadNone functions.
310 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
313 static int counter
= 0;
315 LLVMBuilderRef builder
= ctx
->builder
;
318 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
321 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
322 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
323 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
325 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
326 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
327 LLVMValueRef vgpr
= *pvgpr
;
328 LLVMTypeRef vgpr_type
= LLVMTypeOf(vgpr
);
329 unsigned vgpr_size
= ac_get_type_size(vgpr_type
);
332 assert(vgpr_size
% 4 == 0);
334 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
335 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
336 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
337 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
338 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
345 ac_build_ballot(struct ac_llvm_context
*ctx
,
348 LLVMValueRef args
[3] = {
351 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
354 /* We currently have no other way to prevent LLVM from lifting the icmp
355 * calls to a dominating basic block.
357 ac_build_optimization_barrier(ctx
, &args
[0]);
359 if (LLVMTypeOf(args
[0]) != ctx
->i32
)
360 args
[0] = LLVMBuildBitCast(ctx
->builder
, args
[0], ctx
->i32
, "");
362 return ac_build_intrinsic(ctx
,
363 "llvm.amdgcn.icmp.i32",
365 AC_FUNC_ATTR_NOUNWIND
|
366 AC_FUNC_ATTR_READNONE
|
367 AC_FUNC_ATTR_CONVERGENT
);
371 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
373 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
374 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
375 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
379 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
381 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
382 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
383 LLVMConstInt(ctx
->i64
, 0, 0), "");
387 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
389 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
390 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
392 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
393 vote_set
, active_set
, "");
394 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
396 LLVMConstInt(ctx
->i64
, 0, 0), "");
397 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
401 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
402 unsigned value_count
, unsigned component
)
404 LLVMValueRef vec
= NULL
;
406 if (value_count
== 1) {
407 return values
[component
];
408 } else if (!value_count
)
409 unreachable("value_count is 0");
411 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
412 LLVMValueRef value
= values
[i
];
415 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
416 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
417 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
423 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
424 LLVMValueRef
*values
,
425 unsigned value_count
,
426 unsigned value_stride
,
430 LLVMBuilderRef builder
= ctx
->builder
;
431 LLVMValueRef vec
= NULL
;
434 if (value_count
== 1 && !always_vector
) {
436 return LLVMBuildLoad(builder
, values
[0], "");
438 } else if (!value_count
)
439 unreachable("value_count is 0");
441 for (i
= 0; i
< value_count
; i
++) {
442 LLVMValueRef value
= values
[i
* value_stride
];
444 value
= LLVMBuildLoad(builder
, value
, "");
447 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
448 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
449 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
455 ac_build_gather_values(struct ac_llvm_context
*ctx
,
456 LLVMValueRef
*values
,
457 unsigned value_count
)
459 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
463 ac_build_fdiv(struct ac_llvm_context
*ctx
,
467 LLVMValueRef ret
= LLVMBuildFDiv(ctx
->builder
, num
, den
, "");
469 /* Use v_rcp_f32 instead of precise division. */
470 if (!LLVMIsConstant(ret
))
471 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
475 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
476 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
477 * already multiplied by two. id is the cube face number.
479 struct cube_selection_coords
{
486 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
488 struct cube_selection_coords
*out
)
490 LLVMTypeRef f32
= ctx
->f32
;
492 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
493 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
494 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
495 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
496 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
497 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
498 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
499 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
503 * Build a manual selection sequence for cube face sc/tc coordinates and
504 * major axis vector (multiplied by 2 for consistency) for the given
505 * vec3 \p coords, for the face implied by \p selcoords.
507 * For the major axis, we always adjust the sign to be in the direction of
508 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
509 * the selcoords major axis.
511 static void build_cube_select(struct ac_llvm_context
*ctx
,
512 const struct cube_selection_coords
*selcoords
,
513 const LLVMValueRef
*coords
,
514 LLVMValueRef
*out_st
,
515 LLVMValueRef
*out_ma
)
517 LLVMBuilderRef builder
= ctx
->builder
;
518 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
519 LLVMValueRef is_ma_positive
;
521 LLVMValueRef is_ma_z
, is_not_ma_z
;
522 LLVMValueRef is_ma_y
;
523 LLVMValueRef is_ma_x
;
527 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
528 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
529 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
530 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
532 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
533 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
534 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
535 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
536 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
539 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
540 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
541 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
542 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
543 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
546 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
547 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
548 LLVMConstReal(f32
, -1.0), "");
549 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
552 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
553 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
554 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
555 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
556 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
560 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
561 bool is_deriv
, bool is_array
, bool is_lod
,
562 LLVMValueRef
*coords_arg
,
563 LLVMValueRef
*derivs_arg
)
566 LLVMBuilderRef builder
= ctx
->builder
;
567 struct cube_selection_coords selcoords
;
568 LLVMValueRef coords
[3];
571 if (is_array
&& !is_lod
) {
572 LLVMValueRef tmp
= coords_arg
[3];
573 tmp
= ac_build_intrinsic(ctx
, "llvm.rint.f32", ctx
->f32
, &tmp
, 1, 0);
575 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
577 * "For Array forms, the array layer used will be
579 * max(0, min(d−1, floor(layer+0.5)))
581 * where d is the depth of the texture array and layer
582 * comes from the component indicated in the tables below.
583 * Workaroudn for an issue where the layer is taken from a
584 * helper invocation which happens to fall on a different
585 * layer due to extrapolation."
587 * VI and earlier attempt to implement this in hardware by
588 * clamping the value of coords[2] = (8 * layer) + face.
589 * Unfortunately, this means that the we end up with the wrong
590 * face when clamping occurs.
592 * Clamp the layer earlier to work around the issue.
594 if (ctx
->chip_class
<= VI
) {
596 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
597 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
603 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
605 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
606 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
607 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
609 for (int i
= 0; i
< 2; ++i
)
610 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
612 coords
[2] = selcoords
.id
;
614 if (is_deriv
&& derivs_arg
) {
615 LLVMValueRef derivs
[4];
618 /* Convert cube derivatives to 2D derivatives. */
619 for (axis
= 0; axis
< 2; axis
++) {
620 LLVMValueRef deriv_st
[2];
621 LLVMValueRef deriv_ma
;
623 /* Transform the derivative alongside the texture
624 * coordinate. Mathematically, the correct formula is
625 * as follows. Assume we're projecting onto the +Z face
626 * and denote by dx/dh the derivative of the (original)
627 * X texture coordinate with respect to horizontal
628 * window coordinates. The projection onto the +Z face
633 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
634 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
636 * This motivatives the implementation below.
638 * Whether this actually gives the expected results for
639 * apps that might feed in derivatives obtained via
640 * finite differences is anyone's guess. The OpenGL spec
641 * seems awfully quiet about how textureGrad for cube
642 * maps should be handled.
644 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
645 deriv_st
, &deriv_ma
);
647 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
649 for (int i
= 0; i
< 2; ++i
)
650 derivs
[axis
* 2 + i
] =
651 LLVMBuildFSub(builder
,
652 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
653 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
656 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
659 /* Shift the texture coordinate. This must be applied after the
660 * derivative calculation.
662 for (int i
= 0; i
< 2; ++i
)
663 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
666 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
667 /* coords_arg.w component - array_index for cube arrays */
668 LLVMValueRef tmp
= LLVMBuildFMul(ctx
->builder
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), "");
669 coords
[2] = LLVMBuildFAdd(ctx
->builder
, tmp
, coords
[2], "");
672 memcpy(coords_arg
, coords
, sizeof(coords
));
677 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
678 LLVMValueRef llvm_chan
,
679 LLVMValueRef attr_number
,
684 LLVMValueRef args
[5];
687 if (HAVE_LLVM
< 0x0400) {
689 ij
[0] = LLVMBuildBitCast(ctx
->builder
, i
, ctx
->i32
, "");
690 ij
[1] = LLVMBuildBitCast(ctx
->builder
, j
, ctx
->i32
, "");
693 args
[1] = attr_number
;
695 args
[3] = ac_build_gather_values(ctx
, ij
, 2);
696 return ac_build_intrinsic(ctx
, "llvm.SI.fs.interp",
698 AC_FUNC_ATTR_READNONE
);
703 args
[2] = attr_number
;
706 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
707 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
712 args
[3] = attr_number
;
715 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
716 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
720 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
721 LLVMValueRef parameter
,
722 LLVMValueRef llvm_chan
,
723 LLVMValueRef attr_number
,
726 LLVMValueRef args
[4];
727 if (HAVE_LLVM
< 0x0400) {
729 args
[1] = attr_number
;
732 return ac_build_intrinsic(ctx
,
733 "llvm.SI.fs.constant",
735 AC_FUNC_ATTR_READNONE
);
740 args
[2] = attr_number
;
743 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
744 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
748 ac_build_gep0(struct ac_llvm_context
*ctx
,
749 LLVMValueRef base_ptr
,
752 LLVMValueRef indices
[2] = {
753 LLVMConstInt(ctx
->i32
, 0, 0),
756 return LLVMBuildGEP(ctx
->builder
, base_ptr
,
761 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
762 LLVMValueRef base_ptr
, LLVMValueRef index
,
765 LLVMBuildStore(ctx
->builder
, value
,
766 ac_build_gep0(ctx
, base_ptr
, index
));
770 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
771 * It's equivalent to doing a load from &base_ptr[index].
773 * \param base_ptr Where the array starts.
774 * \param index The element index into the array.
775 * \param uniform Whether the base_ptr and index can be assumed to be
776 * dynamically uniform (i.e. load to an SGPR)
777 * \param invariant Whether the load is invariant (no other opcodes affect it)
780 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
781 LLVMValueRef index
, bool uniform
, bool invariant
)
783 LLVMValueRef pointer
, result
;
785 pointer
= ac_build_gep0(ctx
, base_ptr
, index
);
787 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
788 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
790 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
794 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
797 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false);
800 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
801 LLVMValueRef base_ptr
, LLVMValueRef index
)
803 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true);
806 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
807 LLVMValueRef base_ptr
, LLVMValueRef index
)
809 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true);
812 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
813 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
814 * or v4i32 (num_channels=3,4).
817 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
820 unsigned num_channels
,
821 LLVMValueRef voffset
,
822 LLVMValueRef soffset
,
823 unsigned inst_offset
,
826 bool writeonly_memory
,
827 bool swizzle_enable_hint
)
829 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
830 * (voffset is swizzled, but soffset isn't swizzled).
831 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
833 if (!swizzle_enable_hint
) {
834 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
836 if (num_channels
== 3) {
837 LLVMValueRef v
[3], v01
;
839 for (int i
= 0; i
< 3; i
++) {
840 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
841 LLVMConstInt(ctx
->i32
, i
, 0), "");
843 v01
= ac_build_gather_values(ctx
, v
, 2);
845 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
846 soffset
, inst_offset
, glc
, slc
,
847 writeonly_memory
, swizzle_enable_hint
);
848 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
849 soffset
, inst_offset
+ 8,
851 writeonly_memory
, swizzle_enable_hint
);
855 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
856 static const char *types
[] = {"f32", "v2f32", "v4f32"};
858 LLVMValueRef offset
= soffset
;
861 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
862 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
864 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
866 LLVMValueRef args
[] = {
867 ac_to_float(ctx
, vdata
),
868 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
869 LLVMConstInt(ctx
->i32
, 0, 0),
871 LLVMConstInt(ctx
->i1
, glc
, 0),
872 LLVMConstInt(ctx
->i1
, slc
, 0),
875 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.%s",
878 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
879 args
, ARRAY_SIZE(args
),
881 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
:
882 AC_FUNC_ATTR_WRITEONLY
);
886 static unsigned dfmt
[] = {
887 V_008F0C_BUF_DATA_FORMAT_32
,
888 V_008F0C_BUF_DATA_FORMAT_32_32
,
889 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
890 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
892 assert(num_channels
>= 1 && num_channels
<= 4);
894 LLVMValueRef args
[] = {
897 LLVMConstInt(ctx
->i32
, num_channels
, 0),
898 voffset
? voffset
: LLVMGetUndef(ctx
->i32
),
900 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
901 LLVMConstInt(ctx
->i32
, dfmt
[num_channels
- 1], 0),
902 LLVMConstInt(ctx
->i32
, V_008F0C_BUF_NUM_FORMAT_UINT
, 0),
903 LLVMConstInt(ctx
->i32
, voffset
!= NULL
, 0),
904 LLVMConstInt(ctx
->i32
, 0, 0), /* idxen */
905 LLVMConstInt(ctx
->i32
, glc
, 0),
906 LLVMConstInt(ctx
->i32
, slc
, 0),
907 LLVMConstInt(ctx
->i32
, 0, 0), /* tfe*/
910 /* The instruction offset field has 12 bits */
911 assert(voffset
|| inst_offset
< (1 << 12));
913 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
914 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
915 const char *types
[] = {"i32", "v2i32", "v4i32"};
917 snprintf(name
, sizeof(name
), "llvm.SI.tbuffer.store.%s", types
[func
]);
919 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
920 args
, ARRAY_SIZE(args
),
921 AC_FUNC_ATTR_LEGACY
);
925 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
929 LLVMValueRef voffset
,
930 LLVMValueRef soffset
,
931 unsigned inst_offset
,
937 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
939 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
941 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
943 /* TODO: VI and later generations can use SMEM with GLC=1.*/
944 if (allow_smem
&& !glc
&& !slc
) {
945 assert(vindex
== NULL
);
947 LLVMValueRef result
[4];
949 for (int i
= 0; i
< num_channels
; i
++) {
951 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
952 LLVMConstInt(ctx
->i32
, 4, 0), "");
954 LLVMValueRef args
[2] = {rsrc
, offset
};
955 result
[i
] = ac_build_intrinsic(ctx
, "llvm.SI.load.const.v4i32",
957 AC_FUNC_ATTR_READNONE
|
958 AC_FUNC_ATTR_LEGACY
);
960 if (num_channels
== 1)
963 if (num_channels
== 3)
964 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
965 return ac_build_gather_values(ctx
, result
, num_channels
);
968 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
970 LLVMValueRef args
[] = {
971 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
972 vindex
? vindex
: LLVMConstInt(ctx
->i32
, 0, 0),
974 LLVMConstInt(ctx
->i1
, glc
, 0),
975 LLVMConstInt(ctx
->i1
, slc
, 0)
978 LLVMTypeRef types
[] = {ctx
->f32
, LLVMVectorType(ctx
->f32
, 2),
980 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
983 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.%s",
986 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
988 ac_get_load_intr_attribs(can_speculate
));
991 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
994 LLVMValueRef voffset
,
997 LLVMValueRef args
[] = {
998 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
1001 ctx
->i1false
, /* glc */
1002 ctx
->i1false
, /* slc */
1005 return ac_build_intrinsic(ctx
,
1006 "llvm.amdgcn.buffer.load.format.v4f32",
1007 ctx
->v4f32
, args
, ARRAY_SIZE(args
),
1008 ac_get_load_intr_attribs(can_speculate
));
1012 * Set range metadata on an instruction. This can only be used on load and
1013 * call instructions. If you know an instruction can only produce the values
1014 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1015 * \p lo is the minimum value inclusive.
1016 * \p hi is the maximum value exclusive.
1018 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1019 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1021 LLVMValueRef range_md
, md_args
[2];
1022 LLVMTypeRef type
= LLVMTypeOf(value
);
1023 LLVMContextRef context
= LLVMGetTypeContext(type
);
1025 md_args
[0] = LLVMConstInt(type
, lo
, false);
1026 md_args
[1] = LLVMConstInt(type
, hi
, false);
1027 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1028 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
1032 ac_get_thread_id(struct ac_llvm_context
*ctx
)
1036 LLVMValueRef tid_args
[2];
1037 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
1038 tid_args
[1] = LLVMConstInt(ctx
->i32
, 0, false);
1039 tid_args
[1] = ac_build_intrinsic(ctx
,
1040 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
1041 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
1043 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
1045 2, AC_FUNC_ATTR_READNONE
);
1046 set_range_metadata(ctx
, tid
, 0, 64);
1051 * SI implements derivatives using the local data store (LDS)
1052 * All writes to the LDS happen in all executing threads at
1053 * the same time. TID is the Thread ID for the current
1054 * thread and is a value between 0 and 63, representing
1055 * the thread's position in the wavefront.
1057 * For the pixel shader threads are grouped into quads of four pixels.
1058 * The TIDs of the pixels of a quad are:
1066 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1067 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1068 * the current pixel's column, and masking with 0xfffffffe yields the TID
1069 * of the left pixel of the current pixel's row.
1071 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1072 * adding 2 yields the TID of the pixel below the top pixel.
1075 ac_build_ddxy(struct ac_llvm_context
*ctx
,
1080 LLVMValueRef tl
, trbl
, args
[2];
1081 LLVMValueRef result
;
1083 if (ctx
->chip_class
>= VI
) {
1084 LLVMValueRef thread_id
, tl_tid
, trbl_tid
;
1085 thread_id
= ac_get_thread_id(ctx
);
1087 tl_tid
= LLVMBuildAnd(ctx
->builder
, thread_id
,
1088 LLVMConstInt(ctx
->i32
, mask
, false), "");
1090 trbl_tid
= LLVMBuildAdd(ctx
->builder
, tl_tid
,
1091 LLVMConstInt(ctx
->i32
, idx
, false), "");
1093 args
[0] = LLVMBuildMul(ctx
->builder
, tl_tid
,
1094 LLVMConstInt(ctx
->i32
, 4, false), "");
1096 tl
= ac_build_intrinsic(ctx
,
1097 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
1099 AC_FUNC_ATTR_READNONE
|
1100 AC_FUNC_ATTR_CONVERGENT
);
1102 args
[0] = LLVMBuildMul(ctx
->builder
, trbl_tid
,
1103 LLVMConstInt(ctx
->i32
, 4, false), "");
1104 trbl
= ac_build_intrinsic(ctx
,
1105 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
1107 AC_FUNC_ATTR_READNONE
|
1108 AC_FUNC_ATTR_CONVERGENT
);
1110 uint32_t masks
[2] = {};
1113 case AC_TID_MASK_TOP_LEFT
:
1121 case AC_TID_MASK_TOP
:
1125 case AC_TID_MASK_LEFT
:
1134 args
[1] = LLVMConstInt(ctx
->i32
, masks
[0], false);
1136 tl
= ac_build_intrinsic(ctx
,
1137 "llvm.amdgcn.ds.swizzle", ctx
->i32
,
1139 AC_FUNC_ATTR_READNONE
|
1140 AC_FUNC_ATTR_CONVERGENT
);
1142 args
[1] = LLVMConstInt(ctx
->i32
, masks
[1], false);
1143 trbl
= ac_build_intrinsic(ctx
,
1144 "llvm.amdgcn.ds.swizzle", ctx
->i32
,
1146 AC_FUNC_ATTR_READNONE
|
1147 AC_FUNC_ATTR_CONVERGENT
);
1150 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, ctx
->f32
, "");
1151 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, ctx
->f32
, "");
1152 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
1157 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
1159 LLVMValueRef wave_id
)
1161 LLVMValueRef args
[2];
1162 const char *intr_name
= (HAVE_LLVM
< 0x0400) ? "llvm.SI.sendmsg" : "llvm.amdgcn.s.sendmsg";
1163 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
1165 ac_build_intrinsic(ctx
, intr_name
, ctx
->voidt
, args
, 2, 0);
1169 ac_build_imsb(struct ac_llvm_context
*ctx
,
1171 LLVMTypeRef dst_type
)
1173 const char *intr_name
= (HAVE_LLVM
< 0x0400) ? "llvm.AMDGPU.flbit.i32" :
1174 "llvm.amdgcn.sffbh.i32";
1175 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intr_name
,
1177 AC_FUNC_ATTR_READNONE
);
1179 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1180 * the index from LSB. Invert it by doing "31 - msb". */
1181 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1184 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
1185 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
1186 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1187 arg
, LLVMConstInt(ctx
->i32
, 0, 0), ""),
1188 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1189 arg
, all_ones
, ""), "");
1191 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
1195 ac_build_umsb(struct ac_llvm_context
*ctx
,
1197 LLVMTypeRef dst_type
)
1199 LLVMValueRef args
[2] = {
1203 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32",
1204 dst_type
, args
, ARRAY_SIZE(args
),
1205 AC_FUNC_ATTR_READNONE
);
1207 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1208 * the index from LSB. Invert it by doing "31 - msb". */
1209 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1212 /* check for zero */
1213 return LLVMBuildSelect(ctx
->builder
,
1214 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
,
1215 LLVMConstInt(ctx
->i32
, 0, 0), ""),
1216 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
1219 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1222 LLVMValueRef args
[2] = {a
, b
};
1223 return ac_build_intrinsic(ctx
, "llvm.minnum.f32", ctx
->f32
, args
, 2,
1224 AC_FUNC_ATTR_READNONE
);
1227 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1230 LLVMValueRef args
[2] = {a
, b
};
1231 return ac_build_intrinsic(ctx
, "llvm.maxnum.f32", ctx
->f32
, args
, 2,
1232 AC_FUNC_ATTR_READNONE
);
1235 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1238 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
1239 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1242 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
1244 if (HAVE_LLVM
>= 0x0500) {
1245 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, ctx
->f32_0
),
1249 LLVMValueRef args
[3] = {
1251 LLVMConstReal(ctx
->f32
, 0),
1252 LLVMConstReal(ctx
->f32
, 1),
1255 return ac_build_intrinsic(ctx
, "llvm.AMDGPU.clamp.", ctx
->f32
, args
, 3,
1256 AC_FUNC_ATTR_READNONE
|
1257 AC_FUNC_ATTR_LEGACY
);
1260 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
1262 LLVMValueRef args
[9];
1264 if (HAVE_LLVM
>= 0x0500) {
1265 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
1266 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
1269 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
1270 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
1272 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
1274 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
1276 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1277 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1279 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
1280 ctx
->voidt
, args
, 6, 0);
1282 args
[2] = a
->out
[0];
1283 args
[3] = a
->out
[1];
1284 args
[4] = a
->out
[2];
1285 args
[5] = a
->out
[3];
1286 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1287 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1289 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
1290 ctx
->voidt
, args
, 8, 0);
1295 args
[0] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
1296 args
[1] = LLVMConstInt(ctx
->i32
, a
->valid_mask
, 0);
1297 args
[2] = LLVMConstInt(ctx
->i32
, a
->done
, 0);
1298 args
[3] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
1299 args
[4] = LLVMConstInt(ctx
->i32
, a
->compr
, 0);
1300 memcpy(args
+ 5, a
->out
, sizeof(a
->out
[0]) * 4);
1302 ac_build_intrinsic(ctx
, "llvm.SI.export", ctx
->voidt
, args
, 9,
1303 AC_FUNC_ATTR_LEGACY
);
1306 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
1307 struct ac_image_args
*a
)
1309 LLVMTypeRef dst_type
;
1310 LLVMValueRef args
[11];
1311 unsigned num_args
= 0;
1312 const char *name
= NULL
;
1313 char intr_name
[128], type
[64];
1315 if (HAVE_LLVM
>= 0x0400) {
1316 bool sample
= a
->opcode
== ac_image_sample
||
1317 a
->opcode
== ac_image_gather4
||
1318 a
->opcode
== ac_image_get_lod
;
1321 args
[num_args
++] = ac_to_float(ctx
, a
->addr
);
1323 args
[num_args
++] = a
->addr
;
1325 args
[num_args
++] = a
->resource
;
1327 args
[num_args
++] = a
->sampler
;
1328 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, 0);
1330 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, 0);
1331 args
[num_args
++] = ctx
->i1false
; /* glc */
1332 args
[num_args
++] = ctx
->i1false
; /* slc */
1333 args
[num_args
++] = ctx
->i1false
; /* lwe */
1334 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->da
, 0);
1336 switch (a
->opcode
) {
1337 case ac_image_sample
:
1338 name
= "llvm.amdgcn.image.sample";
1340 case ac_image_gather4
:
1341 name
= "llvm.amdgcn.image.gather4";
1344 name
= "llvm.amdgcn.image.load";
1346 case ac_image_load_mip
:
1347 name
= "llvm.amdgcn.image.load.mip";
1349 case ac_image_get_lod
:
1350 name
= "llvm.amdgcn.image.getlod";
1352 case ac_image_get_resinfo
:
1353 name
= "llvm.amdgcn.image.getresinfo";
1356 unreachable("invalid image opcode");
1359 ac_build_type_name_for_intr(LLVMTypeOf(args
[0]), type
,
1362 snprintf(intr_name
, sizeof(intr_name
), "%s%s%s%s.v4f32.%s.v8i32",
1364 a
->compare
? ".c" : "",
1368 a
->level_zero
? ".lz" : "",
1369 a
->offset
? ".o" : "",
1372 LLVMValueRef result
=
1373 ac_build_intrinsic(ctx
, intr_name
,
1374 ctx
->v4f32
, args
, num_args
,
1375 AC_FUNC_ATTR_READNONE
);
1377 result
= LLVMBuildBitCast(ctx
->builder
, result
,
1383 args
[num_args
++] = a
->addr
;
1384 args
[num_args
++] = a
->resource
;
1386 if (a
->opcode
== ac_image_load
||
1387 a
->opcode
== ac_image_load_mip
||
1388 a
->opcode
== ac_image_get_resinfo
) {
1389 dst_type
= ctx
->v4i32
;
1391 dst_type
= ctx
->v4f32
;
1392 args
[num_args
++] = a
->sampler
;
1395 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, 0);
1396 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->unorm
, 0);
1397 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* r128 */
1398 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->da
, 0);
1399 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* glc */
1400 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* slc */
1401 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* tfe */
1402 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* lwe */
1404 switch (a
->opcode
) {
1405 case ac_image_sample
:
1406 name
= "llvm.SI.image.sample";
1408 case ac_image_gather4
:
1409 name
= "llvm.SI.gather4";
1412 name
= "llvm.SI.image.load";
1414 case ac_image_load_mip
:
1415 name
= "llvm.SI.image.load.mip";
1417 case ac_image_get_lod
:
1418 name
= "llvm.SI.getlod";
1420 case ac_image_get_resinfo
:
1421 name
= "llvm.SI.getresinfo";
1425 ac_build_type_name_for_intr(LLVMTypeOf(a
->addr
), type
, sizeof(type
));
1426 snprintf(intr_name
, sizeof(intr_name
), "%s%s%s%s.%s",
1428 a
->compare
? ".c" : "",
1432 a
->level_zero
? ".lz" : "",
1433 a
->offset
? ".o" : "",
1436 return ac_build_intrinsic(ctx
, intr_name
,
1437 dst_type
, args
, num_args
,
1438 AC_FUNC_ATTR_READNONE
|
1439 AC_FUNC_ATTR_LEGACY
);
1442 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
1443 LLVMValueRef args
[2])
1445 if (HAVE_LLVM
>= 0x0500) {
1447 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
1449 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz",
1451 AC_FUNC_ATTR_READNONE
);
1452 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1455 return ac_build_intrinsic(ctx
, "llvm.SI.packf16", ctx
->i32
, args
, 2,
1456 AC_FUNC_ATTR_READNONE
|
1457 AC_FUNC_ATTR_LEGACY
);
1460 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
1462 assert(HAVE_LLVM
>= 0x0600);
1463 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
1464 &i1
, 1, AC_FUNC_ATTR_READNONE
);
1467 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
1469 if (HAVE_LLVM
>= 0x0600) {
1470 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
1475 LLVMValueRef value
= LLVMBuildSelect(ctx
->builder
, i1
,
1476 LLVMConstReal(ctx
->f32
, 1),
1477 LLVMConstReal(ctx
->f32
, -1), "");
1478 ac_build_intrinsic(ctx
, "llvm.AMDGPU.kill", ctx
->voidt
,
1479 &value
, 1, AC_FUNC_ATTR_LEGACY
);
1482 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
1483 LLVMValueRef offset
, LLVMValueRef width
,
1486 LLVMValueRef args
[] = {
1492 if (HAVE_LLVM
>= 0x0500) {
1493 return ac_build_intrinsic(ctx
,
1494 is_signed
? "llvm.amdgcn.sbfe.i32" :
1495 "llvm.amdgcn.ubfe.i32",
1497 AC_FUNC_ATTR_READNONE
);
1500 return ac_build_intrinsic(ctx
,
1501 is_signed
? "llvm.AMDGPU.bfe.i32" :
1502 "llvm.AMDGPU.bfe.u32",
1504 AC_FUNC_ATTR_READNONE
|
1505 AC_FUNC_ATTR_LEGACY
);
1508 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned simm16
)
1510 LLVMValueRef args
[1] = {
1511 LLVMConstInt(ctx
->i32
, simm16
, false),
1513 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
1514 ctx
->voidt
, args
, 1, 0);
1517 void ac_get_image_intr_name(const char *base_name
,
1518 LLVMTypeRef data_type
,
1519 LLVMTypeRef coords_type
,
1520 LLVMTypeRef rsrc_type
,
1521 char *out_name
, unsigned out_len
)
1523 char coords_type_name
[8];
1525 ac_build_type_name_for_intr(coords_type
, coords_type_name
,
1526 sizeof(coords_type_name
));
1528 if (HAVE_LLVM
<= 0x0309) {
1529 snprintf(out_name
, out_len
, "%s.%s", base_name
, coords_type_name
);
1531 char data_type_name
[8];
1532 char rsrc_type_name
[8];
1534 ac_build_type_name_for_intr(data_type
, data_type_name
,
1535 sizeof(data_type_name
));
1536 ac_build_type_name_for_intr(rsrc_type
, rsrc_type_name
,
1537 sizeof(rsrc_type_name
));
1538 snprintf(out_name
, out_len
, "%s.%s.%s.%s", base_name
,
1539 data_type_name
, coords_type_name
, rsrc_type_name
);
1543 #define AC_EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
1544 #define AC_EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
1552 struct ac_vs_exp_chan
1556 enum ac_ir_type type
;
1559 struct ac_vs_exp_inst
{
1562 struct ac_vs_exp_chan chan
[4];
1565 struct ac_vs_exports
{
1567 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
1570 /* Return true if the PARAM export has been eliminated. */
1571 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
1572 uint32_t num_outputs
,
1573 struct ac_vs_exp_inst
*exp
)
1575 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
1576 bool is_zero
[4] = {}, is_one
[4] = {};
1578 for (i
= 0; i
< 4; i
++) {
1579 /* It's a constant expression. Undef outputs are eliminated too. */
1580 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
1583 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
1584 if (exp
->chan
[i
].const_float
== 0)
1586 else if (exp
->chan
[i
].const_float
== 1)
1589 return false; /* other constant */
1594 /* Only certain combinations of 0 and 1 can be eliminated. */
1595 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
1596 default_val
= is_zero
[3] ? 0 : 1;
1597 else if (is_one
[0] && is_one
[1] && is_one
[2])
1598 default_val
= is_zero
[3] ? 2 : 3;
1602 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
1603 LLVMInstructionEraseFromParent(exp
->inst
);
1605 /* Change OFFSET to DEFAULT_VAL. */
1606 for (i
= 0; i
< num_outputs
; i
++) {
1607 if (vs_output_param_offset
[i
] == exp
->offset
) {
1608 vs_output_param_offset
[i
] =
1609 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
1616 static bool ac_eliminate_duplicated_output(uint8_t *vs_output_param_offset
,
1617 uint32_t num_outputs
,
1618 struct ac_vs_exports
*processed
,
1619 struct ac_vs_exp_inst
*exp
)
1621 unsigned p
, copy_back_channels
= 0;
1623 /* See if the output is already in the list of processed outputs.
1624 * The LLVMValueRef comparison relies on SSA.
1626 for (p
= 0; p
< processed
->num
; p
++) {
1627 bool different
= false;
1629 for (unsigned j
= 0; j
< 4; j
++) {
1630 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
1631 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
1633 /* Treat undef as a match. */
1634 if (c2
->type
== AC_IR_UNDEF
)
1637 /* If c1 is undef but c2 isn't, we can copy c2 to c1
1638 * and consider the instruction duplicated.
1640 if (c1
->type
== AC_IR_UNDEF
) {
1641 copy_back_channels
|= 1 << j
;
1645 /* Test whether the channels are not equal. */
1646 if (c1
->type
!= c2
->type
||
1647 (c1
->type
== AC_IR_CONST
&&
1648 c1
->const_float
!= c2
->const_float
) ||
1649 (c1
->type
== AC_IR_VALUE
&&
1650 c1
->value
!= c2
->value
)) {
1658 copy_back_channels
= 0;
1660 if (p
== processed
->num
)
1663 /* If a match was found, but the matching export has undef where the new
1664 * one has a normal value, copy the normal value to the undef channel.
1666 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
1668 while (copy_back_channels
) {
1669 unsigned chan
= u_bit_scan(©_back_channels
);
1671 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
1672 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
1673 exp
->chan
[chan
].value
);
1674 match
->chan
[chan
] = exp
->chan
[chan
];
1677 /* The PARAM export is duplicated. Kill it. */
1678 LLVMInstructionEraseFromParent(exp
->inst
);
1680 /* Change OFFSET to the matching export. */
1681 for (unsigned i
= 0; i
< num_outputs
; i
++) {
1682 if (vs_output_param_offset
[i
] == exp
->offset
) {
1683 vs_output_param_offset
[i
] = match
->offset
;
1690 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
1691 LLVMValueRef main_fn
,
1692 uint8_t *vs_output_param_offset
,
1693 uint32_t num_outputs
,
1694 uint8_t *num_param_exports
)
1696 LLVMBasicBlockRef bb
;
1697 bool removed_any
= false;
1698 struct ac_vs_exports exports
;
1702 /* Process all LLVM instructions. */
1703 bb
= LLVMGetFirstBasicBlock(main_fn
);
1705 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
1708 LLVMValueRef cur
= inst
;
1709 inst
= LLVMGetNextInstruction(inst
);
1710 struct ac_vs_exp_inst exp
;
1712 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
1715 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
1717 if (!ac_llvm_is_function(callee
))
1720 const char *name
= LLVMGetValueName(callee
);
1721 unsigned num_args
= LLVMCountParams(callee
);
1723 /* Check if this is an export instruction. */
1724 if ((num_args
!= 9 && num_args
!= 8) ||
1725 (strcmp(name
, "llvm.SI.export") &&
1726 strcmp(name
, "llvm.amdgcn.exp.f32")))
1729 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
1730 unsigned target
= LLVMConstIntGetZExtValue(arg
);
1732 if (target
< V_008DFC_SQ_EXP_PARAM
)
1735 target
-= V_008DFC_SQ_EXP_PARAM
;
1737 /* Parse the instruction. */
1738 memset(&exp
, 0, sizeof(exp
));
1739 exp
.offset
= target
;
1742 for (unsigned i
= 0; i
< 4; i
++) {
1743 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
1745 exp
.chan
[i
].value
= v
;
1747 if (LLVMIsUndef(v
)) {
1748 exp
.chan
[i
].type
= AC_IR_UNDEF
;
1749 } else if (LLVMIsAConstantFP(v
)) {
1750 LLVMBool loses_info
;
1751 exp
.chan
[i
].type
= AC_IR_CONST
;
1752 exp
.chan
[i
].const_float
=
1753 LLVMConstRealGetDouble(v
, &loses_info
);
1755 exp
.chan
[i
].type
= AC_IR_VALUE
;
1759 /* Eliminate constant and duplicated PARAM exports. */
1760 if (ac_eliminate_const_output(vs_output_param_offset
,
1761 num_outputs
, &exp
) ||
1762 ac_eliminate_duplicated_output(vs_output_param_offset
,
1763 num_outputs
, &exports
,
1767 exports
.exp
[exports
.num
++] = exp
;
1770 bb
= LLVMGetNextBasicBlock(bb
);
1773 /* Remove holes in export memory due to removed PARAM exports.
1774 * This is done by renumbering all PARAM exports.
1777 uint8_t old_offset
[VARYING_SLOT_MAX
];
1780 /* Make a copy of the offsets. We need the old version while
1781 * we are modifying some of them. */
1782 memcpy(old_offset
, vs_output_param_offset
,
1783 sizeof(old_offset
));
1785 for (i
= 0; i
< exports
.num
; i
++) {
1786 unsigned offset
= exports
.exp
[i
].offset
;
1788 /* Update vs_output_param_offset. Multiple outputs can
1789 * have the same offset.
1791 for (out
= 0; out
< num_outputs
; out
++) {
1792 if (old_offset
[out
] == offset
)
1793 vs_output_param_offset
[out
] = i
;
1796 /* Change the PARAM offset in the instruction. */
1797 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
1798 LLVMConstInt(ctx
->i32
,
1799 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
1801 *num_param_exports
= exports
.num
;
1805 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
1807 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
1808 ac_build_intrinsic(ctx
,
1809 "llvm.amdgcn.init.exec", ctx
->voidt
,
1810 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
1813 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
1815 unsigned lds_size
= ctx
->chip_class
>= CIK
? 65536 : 32768;
1816 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
1817 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_LOCAL_ADDR_SPACE
),
1821 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
1822 LLVMValueRef dw_addr
)
1824 return ac_build_load(ctx
, ctx
->lds
, dw_addr
);
1827 void ac_lds_store(struct ac_llvm_context
*ctx
,
1828 LLVMValueRef dw_addr
,
1831 value
= ac_to_integer(ctx
, value
);
1832 ac_build_indexed_store(ctx
, ctx
->lds
,
1836 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
1837 LLVMTypeRef dst_type
,
1840 LLVMValueRef params
[2] = {
1843 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
1844 * add special code to check for x=0. The reason is that
1845 * the LLVM behavior for x=0 is different from what we
1846 * need here. However, LLVM also assumes that ffs(x) is
1847 * in [0, 31], but GLSL expects that ffs(0) = -1, so
1848 * a conditional assignment to handle 0 is still required.
1850 * The hardware already implements the correct behavior.
1852 LLVMConstInt(ctx
->i1
, 1, false),
1855 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, "llvm.cttz.i32", ctx
->i32
,
1857 AC_FUNC_ATTR_READNONE
);
1859 /* TODO: We need an intrinsic to skip this conditional. */
1860 /* Check for zero: */
1861 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
1864 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");