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
)
54 ctx
->chip_class
= chip_class
;
56 ctx
->context
= context
;
60 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
61 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
62 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
63 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
64 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
65 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
66 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
67 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
68 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
69 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
70 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
71 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
73 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
74 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
75 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
76 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
78 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
81 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
82 "invariant.load", 14);
84 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
86 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
87 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
89 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
90 "amdgpu.uniform", 14);
92 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
96 ac_get_type_size(LLVMTypeRef type
)
98 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
101 case LLVMIntegerTypeKind
:
102 return LLVMGetIntTypeWidth(type
) / 8;
103 case LLVMFloatTypeKind
:
105 case LLVMDoubleTypeKind
:
106 case LLVMPointerTypeKind
:
108 case LLVMVectorTypeKind
:
109 return LLVMGetVectorSize(type
) *
110 ac_get_type_size(LLVMGetElementType(type
));
111 case LLVMArrayTypeKind
:
112 return LLVMGetArrayLength(type
) *
113 ac_get_type_size(LLVMGetElementType(type
));
120 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
122 if (t
== ctx
->f16
|| t
== ctx
->i16
)
124 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
126 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
129 unreachable("Unhandled integer size");
133 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
135 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
136 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
137 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
138 LLVMGetVectorSize(t
));
140 return to_integer_type_scalar(ctx
, t
);
144 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
146 LLVMTypeRef type
= LLVMTypeOf(v
);
147 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
150 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
152 if (t
== ctx
->i16
|| t
== ctx
->f16
)
154 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
156 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
159 unreachable("Unhandled float size");
163 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
165 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
166 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
167 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
168 LLVMGetVectorSize(t
));
170 return to_float_type_scalar(ctx
, t
);
174 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
176 LLVMTypeRef type
= LLVMTypeOf(v
);
177 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
182 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
183 LLVMTypeRef return_type
, LLVMValueRef
*params
,
184 unsigned param_count
, unsigned attrib_mask
)
186 LLVMValueRef function
, call
;
187 bool set_callsite_attrs
= HAVE_LLVM
>= 0x0400 &&
188 !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
190 function
= LLVMGetNamedFunction(ctx
->module
, name
);
192 LLVMTypeRef param_types
[32], function_type
;
195 assert(param_count
<= 32);
197 for (i
= 0; i
< param_count
; ++i
) {
199 param_types
[i
] = LLVMTypeOf(params
[i
]);
202 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
203 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
205 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
206 LLVMSetLinkage(function
, LLVMExternalLinkage
);
208 if (!set_callsite_attrs
)
209 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
212 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
213 if (set_callsite_attrs
)
214 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
219 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
222 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
224 LLVMTypeRef elem_type
= type
;
226 assert(bufsize
>= 8);
228 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
229 int ret
= snprintf(buf
, bufsize
, "v%u",
230 LLVMGetVectorSize(type
));
232 char *type_name
= LLVMPrintTypeToString(type
);
233 fprintf(stderr
, "Error building type name for: %s\n",
237 elem_type
= LLVMGetElementType(type
);
241 switch (LLVMGetTypeKind(elem_type
)) {
243 case LLVMIntegerTypeKind
:
244 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
246 case LLVMFloatTypeKind
:
247 snprintf(buf
, bufsize
, "f32");
249 case LLVMDoubleTypeKind
:
250 snprintf(buf
, bufsize
, "f64");
255 /* Prevent optimizations (at least of memory accesses) across the current
256 * point in the program by emitting empty inline assembly that is marked as
257 * having side effects.
259 * Optionally, a value can be passed through the inline assembly to prevent
260 * LLVM from hoisting calls to ReadNone functions.
263 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
266 static int counter
= 0;
268 LLVMBuilderRef builder
= ctx
->builder
;
271 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
274 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
275 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
276 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
278 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
279 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
280 LLVMValueRef vgpr
= *pvgpr
;
281 LLVMTypeRef vgpr_type
= LLVMTypeOf(vgpr
);
282 unsigned vgpr_size
= ac_get_type_size(vgpr_type
);
285 assert(vgpr_size
% 4 == 0);
287 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
288 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
289 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
290 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
291 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
298 ac_build_ballot(struct ac_llvm_context
*ctx
,
301 LLVMValueRef args
[3] = {
304 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
307 /* We currently have no other way to prevent LLVM from lifting the icmp
308 * calls to a dominating basic block.
310 ac_build_optimization_barrier(ctx
, &args
[0]);
312 if (LLVMTypeOf(args
[0]) != ctx
->i32
)
313 args
[0] = LLVMBuildBitCast(ctx
->builder
, args
[0], ctx
->i32
, "");
315 return ac_build_intrinsic(ctx
,
316 "llvm.amdgcn.icmp.i32",
318 AC_FUNC_ATTR_NOUNWIND
|
319 AC_FUNC_ATTR_READNONE
|
320 AC_FUNC_ATTR_CONVERGENT
);
324 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
326 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
327 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
328 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
332 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
334 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
335 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
336 LLVMConstInt(ctx
->i64
, 0, 0), "");
340 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
342 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
343 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
345 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
346 vote_set
, active_set
, "");
347 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
349 LLVMConstInt(ctx
->i64
, 0, 0), "");
350 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
354 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
355 LLVMValueRef
*values
,
356 unsigned value_count
,
357 unsigned value_stride
,
361 LLVMBuilderRef builder
= ctx
->builder
;
362 LLVMValueRef vec
= NULL
;
365 if (value_count
== 1 && !always_vector
) {
367 return LLVMBuildLoad(builder
, values
[0], "");
369 } else if (!value_count
)
370 unreachable("value_count is 0");
372 for (i
= 0; i
< value_count
; i
++) {
373 LLVMValueRef value
= values
[i
* value_stride
];
375 value
= LLVMBuildLoad(builder
, value
, "");
378 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
379 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
380 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
386 ac_build_gather_values(struct ac_llvm_context
*ctx
,
387 LLVMValueRef
*values
,
388 unsigned value_count
)
390 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
394 ac_build_fdiv(struct ac_llvm_context
*ctx
,
398 LLVMValueRef ret
= LLVMBuildFDiv(ctx
->builder
, num
, den
, "");
400 if (!LLVMIsConstant(ret
))
401 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
405 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
406 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
407 * already multiplied by two. id is the cube face number.
409 struct cube_selection_coords
{
416 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
418 struct cube_selection_coords
*out
)
420 LLVMTypeRef f32
= ctx
->f32
;
422 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
423 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
424 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
425 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
426 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
427 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
428 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
429 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
433 * Build a manual selection sequence for cube face sc/tc coordinates and
434 * major axis vector (multiplied by 2 for consistency) for the given
435 * vec3 \p coords, for the face implied by \p selcoords.
437 * For the major axis, we always adjust the sign to be in the direction of
438 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
439 * the selcoords major axis.
441 static void build_cube_select(struct ac_llvm_context
*ctx
,
442 const struct cube_selection_coords
*selcoords
,
443 const LLVMValueRef
*coords
,
444 LLVMValueRef
*out_st
,
445 LLVMValueRef
*out_ma
)
447 LLVMBuilderRef builder
= ctx
->builder
;
448 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
449 LLVMValueRef is_ma_positive
;
451 LLVMValueRef is_ma_z
, is_not_ma_z
;
452 LLVMValueRef is_ma_y
;
453 LLVMValueRef is_ma_x
;
457 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
458 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
459 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
460 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
462 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
463 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
464 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
465 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
466 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
469 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
470 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
471 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
472 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
473 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
476 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
477 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
478 LLVMConstReal(f32
, -1.0), "");
479 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
482 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
483 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
484 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
485 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
486 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
490 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
491 bool is_deriv
, bool is_array
, bool is_lod
,
492 LLVMValueRef
*coords_arg
,
493 LLVMValueRef
*derivs_arg
)
496 LLVMBuilderRef builder
= ctx
->builder
;
497 struct cube_selection_coords selcoords
;
498 LLVMValueRef coords
[3];
501 if (is_array
&& !is_lod
) {
502 LLVMValueRef tmp
= coords_arg
[3];
503 tmp
= ac_build_intrinsic(ctx
, "llvm.rint.f32", ctx
->f32
, &tmp
, 1, 0);
505 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
507 * "For Array forms, the array layer used will be
509 * max(0, min(d−1, floor(layer+0.5)))
511 * where d is the depth of the texture array and layer
512 * comes from the component indicated in the tables below.
513 * Workaroudn for an issue where the layer is taken from a
514 * helper invocation which happens to fall on a different
515 * layer due to extrapolation."
517 * VI and earlier attempt to implement this in hardware by
518 * clamping the value of coords[2] = (8 * layer) + face.
519 * Unfortunately, this means that the we end up with the wrong
520 * face when clamping occurs.
522 * Clamp the layer earlier to work around the issue.
524 if (ctx
->chip_class
<= VI
) {
526 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
527 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
533 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
535 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
536 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
537 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
539 for (int i
= 0; i
< 2; ++i
)
540 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
542 coords
[2] = selcoords
.id
;
544 if (is_deriv
&& derivs_arg
) {
545 LLVMValueRef derivs
[4];
548 /* Convert cube derivatives to 2D derivatives. */
549 for (axis
= 0; axis
< 2; axis
++) {
550 LLVMValueRef deriv_st
[2];
551 LLVMValueRef deriv_ma
;
553 /* Transform the derivative alongside the texture
554 * coordinate. Mathematically, the correct formula is
555 * as follows. Assume we're projecting onto the +Z face
556 * and denote by dx/dh the derivative of the (original)
557 * X texture coordinate with respect to horizontal
558 * window coordinates. The projection onto the +Z face
563 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
564 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
566 * This motivatives the implementation below.
568 * Whether this actually gives the expected results for
569 * apps that might feed in derivatives obtained via
570 * finite differences is anyone's guess. The OpenGL spec
571 * seems awfully quiet about how textureGrad for cube
572 * maps should be handled.
574 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
575 deriv_st
, &deriv_ma
);
577 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
579 for (int i
= 0; i
< 2; ++i
)
580 derivs
[axis
* 2 + i
] =
581 LLVMBuildFSub(builder
,
582 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
583 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
586 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
589 /* Shift the texture coordinate. This must be applied after the
590 * derivative calculation.
592 for (int i
= 0; i
< 2; ++i
)
593 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
596 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
597 /* coords_arg.w component - array_index for cube arrays */
598 LLVMValueRef tmp
= LLVMBuildFMul(ctx
->builder
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), "");
599 coords
[2] = LLVMBuildFAdd(ctx
->builder
, tmp
, coords
[2], "");
602 memcpy(coords_arg
, coords
, sizeof(coords
));
607 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
608 LLVMValueRef llvm_chan
,
609 LLVMValueRef attr_number
,
614 LLVMValueRef args
[5];
617 if (HAVE_LLVM
< 0x0400) {
619 ij
[0] = LLVMBuildBitCast(ctx
->builder
, i
, ctx
->i32
, "");
620 ij
[1] = LLVMBuildBitCast(ctx
->builder
, j
, ctx
->i32
, "");
623 args
[1] = attr_number
;
625 args
[3] = ac_build_gather_values(ctx
, ij
, 2);
626 return ac_build_intrinsic(ctx
, "llvm.SI.fs.interp",
628 AC_FUNC_ATTR_READNONE
);
633 args
[2] = attr_number
;
636 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
637 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
642 args
[3] = attr_number
;
645 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
646 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
650 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
651 LLVMValueRef parameter
,
652 LLVMValueRef llvm_chan
,
653 LLVMValueRef attr_number
,
656 LLVMValueRef args
[4];
657 if (HAVE_LLVM
< 0x0400) {
659 args
[1] = attr_number
;
662 return ac_build_intrinsic(ctx
,
663 "llvm.SI.fs.constant",
665 AC_FUNC_ATTR_READNONE
);
670 args
[2] = attr_number
;
673 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
674 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
678 ac_build_gep0(struct ac_llvm_context
*ctx
,
679 LLVMValueRef base_ptr
,
682 LLVMValueRef indices
[2] = {
683 LLVMConstInt(ctx
->i32
, 0, 0),
686 return LLVMBuildGEP(ctx
->builder
, base_ptr
,
691 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
692 LLVMValueRef base_ptr
, LLVMValueRef index
,
695 LLVMBuildStore(ctx
->builder
, value
,
696 ac_build_gep0(ctx
, base_ptr
, index
));
700 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
701 * It's equivalent to doing a load from &base_ptr[index].
703 * \param base_ptr Where the array starts.
704 * \param index The element index into the array.
705 * \param uniform Whether the base_ptr and index can be assumed to be
706 * dynamically uniform
709 ac_build_indexed_load(struct ac_llvm_context
*ctx
,
710 LLVMValueRef base_ptr
, LLVMValueRef index
,
713 LLVMValueRef pointer
;
715 pointer
= ac_build_gep0(ctx
, base_ptr
, index
);
717 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
718 return LLVMBuildLoad(ctx
->builder
, pointer
, "");
722 * Do a load from &base_ptr[index], but also add a flag that it's loading
723 * a constant from a dynamically uniform index.
726 ac_build_indexed_load_const(struct ac_llvm_context
*ctx
,
727 LLVMValueRef base_ptr
, LLVMValueRef index
)
729 LLVMValueRef result
= ac_build_indexed_load(ctx
, base_ptr
, index
, true);
730 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
734 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
735 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
736 * or v4i32 (num_channels=3,4).
739 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
742 unsigned num_channels
,
743 LLVMValueRef voffset
,
744 LLVMValueRef soffset
,
745 unsigned inst_offset
,
748 bool writeonly_memory
,
751 /* TODO: Fix stores with ADD_TID and remove the "has_add_tid" flag. */
753 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
755 if (num_channels
== 3) {
756 LLVMValueRef v
[3], v01
;
758 for (int i
= 0; i
< 3; i
++) {
759 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
760 LLVMConstInt(ctx
->i32
, i
, 0), "");
762 v01
= ac_build_gather_values(ctx
, v
, 2);
764 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
765 soffset
, inst_offset
, glc
, slc
,
766 writeonly_memory
, has_add_tid
);
767 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
768 soffset
, inst_offset
+ 8,
770 writeonly_memory
, has_add_tid
);
774 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
775 static const char *types
[] = {"f32", "v2f32", "v4f32"};
777 LLVMValueRef offset
= soffset
;
780 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
781 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
783 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
785 LLVMValueRef args
[] = {
786 ac_to_float(ctx
, vdata
),
787 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
788 LLVMConstInt(ctx
->i32
, 0, 0),
790 LLVMConstInt(ctx
->i1
, glc
, 0),
791 LLVMConstInt(ctx
->i1
, slc
, 0),
794 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.%s",
797 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
798 args
, ARRAY_SIZE(args
),
800 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
:
801 AC_FUNC_ATTR_WRITEONLY
);
805 static unsigned dfmt
[] = {
806 V_008F0C_BUF_DATA_FORMAT_32
,
807 V_008F0C_BUF_DATA_FORMAT_32_32
,
808 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
809 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
811 assert(num_channels
>= 1 && num_channels
<= 4);
813 LLVMValueRef args
[] = {
816 LLVMConstInt(ctx
->i32
, num_channels
, 0),
817 voffset
? voffset
: LLVMGetUndef(ctx
->i32
),
819 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
820 LLVMConstInt(ctx
->i32
, dfmt
[num_channels
- 1], 0),
821 LLVMConstInt(ctx
->i32
, V_008F0C_BUF_NUM_FORMAT_UINT
, 0),
822 LLVMConstInt(ctx
->i32
, voffset
!= NULL
, 0),
823 LLVMConstInt(ctx
->i32
, 0, 0), /* idxen */
824 LLVMConstInt(ctx
->i32
, glc
, 0),
825 LLVMConstInt(ctx
->i32
, slc
, 0),
826 LLVMConstInt(ctx
->i32
, 0, 0), /* tfe*/
829 /* The instruction offset field has 12 bits */
830 assert(voffset
|| inst_offset
< (1 << 12));
832 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
833 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
834 const char *types
[] = {"i32", "v2i32", "v4i32"};
836 snprintf(name
, sizeof(name
), "llvm.SI.tbuffer.store.%s", types
[func
]);
838 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
839 args
, ARRAY_SIZE(args
),
840 AC_FUNC_ATTR_LEGACY
);
844 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
848 LLVMValueRef voffset
,
849 LLVMValueRef soffset
,
850 unsigned inst_offset
,
856 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
858 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
860 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
862 /* TODO: VI and later generations can use SMEM with GLC=1.*/
863 if (allow_smem
&& !glc
&& !slc
) {
864 assert(vindex
== NULL
);
866 LLVMValueRef result
[4];
868 for (int i
= 0; i
< num_channels
; i
++) {
870 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
871 LLVMConstInt(ctx
->i32
, 4, 0), "");
873 LLVMValueRef args
[2] = {rsrc
, offset
};
874 result
[i
] = ac_build_intrinsic(ctx
, "llvm.SI.load.const.v4i32",
876 AC_FUNC_ATTR_READNONE
|
877 AC_FUNC_ATTR_LEGACY
);
879 if (num_channels
== 1)
882 if (num_channels
== 3)
883 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
884 return ac_build_gather_values(ctx
, result
, num_channels
);
887 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
889 LLVMValueRef args
[] = {
890 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
891 vindex
? vindex
: LLVMConstInt(ctx
->i32
, 0, 0),
893 LLVMConstInt(ctx
->i1
, glc
, 0),
894 LLVMConstInt(ctx
->i1
, slc
, 0)
897 LLVMTypeRef types
[] = {ctx
->f32
, LLVMVectorType(ctx
->f32
, 2),
899 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
902 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.%s",
905 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
907 /* READNONE means writes can't affect it, while
908 * READONLY means that writes can affect it. */
909 can_speculate
&& HAVE_LLVM
>= 0x0400 ?
910 AC_FUNC_ATTR_READNONE
:
911 AC_FUNC_ATTR_READONLY
);
914 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
917 LLVMValueRef voffset
,
920 LLVMValueRef args
[] = {
921 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
924 LLVMConstInt(ctx
->i1
, 0, 0), /* glc */
925 LLVMConstInt(ctx
->i1
, 0, 0), /* slc */
928 return ac_build_intrinsic(ctx
,
929 "llvm.amdgcn.buffer.load.format.v4f32",
930 ctx
->v4f32
, args
, ARRAY_SIZE(args
),
931 /* READNONE means writes can't affect it, while
932 * READONLY means that writes can affect it. */
933 can_speculate
&& HAVE_LLVM
>= 0x0400 ?
934 AC_FUNC_ATTR_READNONE
:
935 AC_FUNC_ATTR_READONLY
);
939 * Set range metadata on an instruction. This can only be used on load and
940 * call instructions. If you know an instruction can only produce the values
941 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
942 * \p lo is the minimum value inclusive.
943 * \p hi is the maximum value exclusive.
945 static void set_range_metadata(struct ac_llvm_context
*ctx
,
946 LLVMValueRef value
, unsigned lo
, unsigned hi
)
948 LLVMValueRef range_md
, md_args
[2];
949 LLVMTypeRef type
= LLVMTypeOf(value
);
950 LLVMContextRef context
= LLVMGetTypeContext(type
);
952 md_args
[0] = LLVMConstInt(type
, lo
, false);
953 md_args
[1] = LLVMConstInt(type
, hi
, false);
954 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
955 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
959 ac_get_thread_id(struct ac_llvm_context
*ctx
)
963 LLVMValueRef tid_args
[2];
964 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
965 tid_args
[1] = LLVMConstInt(ctx
->i32
, 0, false);
966 tid_args
[1] = ac_build_intrinsic(ctx
,
967 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
968 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
970 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
972 2, AC_FUNC_ATTR_READNONE
);
973 set_range_metadata(ctx
, tid
, 0, 64);
978 * SI implements derivatives using the local data store (LDS)
979 * All writes to the LDS happen in all executing threads at
980 * the same time. TID is the Thread ID for the current
981 * thread and is a value between 0 and 63, representing
982 * the thread's position in the wavefront.
984 * For the pixel shader threads are grouped into quads of four pixels.
985 * The TIDs of the pixels of a quad are:
993 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
994 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
995 * the current pixel's column, and masking with 0xfffffffe yields the TID
996 * of the left pixel of the current pixel's row.
998 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
999 * adding 2 yields the TID of the pixel below the top pixel.
1002 ac_build_ddxy(struct ac_llvm_context
*ctx
,
1007 LLVMValueRef tl
, trbl
, args
[2];
1008 LLVMValueRef result
;
1010 if (ctx
->chip_class
>= VI
) {
1011 LLVMValueRef thread_id
, tl_tid
, trbl_tid
;
1012 thread_id
= ac_get_thread_id(ctx
);
1014 tl_tid
= LLVMBuildAnd(ctx
->builder
, thread_id
,
1015 LLVMConstInt(ctx
->i32
, mask
, false), "");
1017 trbl_tid
= LLVMBuildAdd(ctx
->builder
, tl_tid
,
1018 LLVMConstInt(ctx
->i32
, idx
, false), "");
1020 args
[0] = LLVMBuildMul(ctx
->builder
, tl_tid
,
1021 LLVMConstInt(ctx
->i32
, 4, false), "");
1023 tl
= ac_build_intrinsic(ctx
,
1024 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
1026 AC_FUNC_ATTR_READNONE
|
1027 AC_FUNC_ATTR_CONVERGENT
);
1029 args
[0] = LLVMBuildMul(ctx
->builder
, trbl_tid
,
1030 LLVMConstInt(ctx
->i32
, 4, false), "");
1031 trbl
= ac_build_intrinsic(ctx
,
1032 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
1034 AC_FUNC_ATTR_READNONE
|
1035 AC_FUNC_ATTR_CONVERGENT
);
1040 case AC_TID_MASK_TOP_LEFT
:
1048 case AC_TID_MASK_TOP
:
1052 case AC_TID_MASK_LEFT
:
1059 args
[1] = LLVMConstInt(ctx
->i32
, masks
[0], false);
1061 tl
= ac_build_intrinsic(ctx
,
1062 "llvm.amdgcn.ds.swizzle", ctx
->i32
,
1064 AC_FUNC_ATTR_READNONE
|
1065 AC_FUNC_ATTR_CONVERGENT
);
1067 args
[1] = LLVMConstInt(ctx
->i32
, masks
[1], false);
1068 trbl
= ac_build_intrinsic(ctx
,
1069 "llvm.amdgcn.ds.swizzle", ctx
->i32
,
1071 AC_FUNC_ATTR_READNONE
|
1072 AC_FUNC_ATTR_CONVERGENT
);
1075 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, ctx
->f32
, "");
1076 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, ctx
->f32
, "");
1077 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
1082 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
1084 LLVMValueRef wave_id
)
1086 LLVMValueRef args
[2];
1087 const char *intr_name
= (HAVE_LLVM
< 0x0400) ? "llvm.SI.sendmsg" : "llvm.amdgcn.s.sendmsg";
1088 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
1090 ac_build_intrinsic(ctx
, intr_name
, ctx
->voidt
, args
, 2, 0);
1094 ac_build_imsb(struct ac_llvm_context
*ctx
,
1096 LLVMTypeRef dst_type
)
1098 const char *intr_name
= (HAVE_LLVM
< 0x0400) ? "llvm.AMDGPU.flbit.i32" :
1099 "llvm.amdgcn.sffbh.i32";
1100 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intr_name
,
1102 AC_FUNC_ATTR_READNONE
);
1104 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1105 * the index from LSB. Invert it by doing "31 - msb". */
1106 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1109 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
1110 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
1111 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1112 arg
, LLVMConstInt(ctx
->i32
, 0, 0), ""),
1113 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1114 arg
, all_ones
, ""), "");
1116 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
1120 ac_build_umsb(struct ac_llvm_context
*ctx
,
1122 LLVMTypeRef dst_type
)
1124 LLVMValueRef args
[2] = {
1126 LLVMConstInt(ctx
->i1
, 1, 0),
1128 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32",
1129 dst_type
, args
, ARRAY_SIZE(args
),
1130 AC_FUNC_ATTR_READNONE
);
1132 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1133 * the index from LSB. Invert it by doing "31 - msb". */
1134 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1137 /* check for zero */
1138 return LLVMBuildSelect(ctx
->builder
,
1139 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
,
1140 LLVMConstInt(ctx
->i32
, 0, 0), ""),
1141 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
1144 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1147 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
1148 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1151 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
1153 if (HAVE_LLVM
>= 0x0500) {
1154 LLVMValueRef max
[2] = {
1156 LLVMConstReal(ctx
->f32
, 0),
1158 LLVMValueRef min
[2] = {
1159 LLVMConstReal(ctx
->f32
, 1),
1162 min
[1] = ac_build_intrinsic(ctx
, "llvm.maxnum.f32",
1164 AC_FUNC_ATTR_READNONE
);
1165 return ac_build_intrinsic(ctx
, "llvm.minnum.f32",
1167 AC_FUNC_ATTR_READNONE
);
1170 LLVMValueRef args
[3] = {
1172 LLVMConstReal(ctx
->f32
, 0),
1173 LLVMConstReal(ctx
->f32
, 1),
1176 return ac_build_intrinsic(ctx
, "llvm.AMDGPU.clamp.", ctx
->f32
, args
, 3,
1177 AC_FUNC_ATTR_READNONE
|
1178 AC_FUNC_ATTR_LEGACY
);
1181 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
1183 LLVMValueRef args
[9];
1185 if (HAVE_LLVM
>= 0x0500) {
1186 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
1187 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
1190 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
1191 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
1193 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
1195 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
1197 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1198 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1200 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
1201 ctx
->voidt
, args
, 6, 0);
1203 args
[2] = a
->out
[0];
1204 args
[3] = a
->out
[1];
1205 args
[4] = a
->out
[2];
1206 args
[5] = a
->out
[3];
1207 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1208 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1210 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
1211 ctx
->voidt
, args
, 8, 0);
1216 args
[0] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
1217 args
[1] = LLVMConstInt(ctx
->i32
, a
->valid_mask
, 0);
1218 args
[2] = LLVMConstInt(ctx
->i32
, a
->done
, 0);
1219 args
[3] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
1220 args
[4] = LLVMConstInt(ctx
->i32
, a
->compr
, 0);
1221 memcpy(args
+ 5, a
->out
, sizeof(a
->out
[0]) * 4);
1223 ac_build_intrinsic(ctx
, "llvm.SI.export", ctx
->voidt
, args
, 9,
1224 AC_FUNC_ATTR_LEGACY
);
1227 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
1228 struct ac_image_args
*a
)
1230 LLVMTypeRef dst_type
;
1231 LLVMValueRef args
[11];
1232 unsigned num_args
= 0;
1234 char intr_name
[128], type
[64];
1236 if (HAVE_LLVM
>= 0x0400) {
1237 bool sample
= a
->opcode
== ac_image_sample
||
1238 a
->opcode
== ac_image_gather4
||
1239 a
->opcode
== ac_image_get_lod
;
1242 args
[num_args
++] = ac_to_float(ctx
, a
->addr
);
1244 args
[num_args
++] = a
->addr
;
1246 args
[num_args
++] = a
->resource
;
1248 args
[num_args
++] = a
->sampler
;
1249 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, 0);
1251 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, 0);
1252 args
[num_args
++] = LLVMConstInt(ctx
->i1
, 0, 0); /* glc */
1253 args
[num_args
++] = LLVMConstInt(ctx
->i1
, 0, 0); /* slc */
1254 args
[num_args
++] = LLVMConstInt(ctx
->i1
, 0, 0); /* lwe */
1255 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->da
, 0);
1257 switch (a
->opcode
) {
1258 case ac_image_sample
:
1259 name
= "llvm.amdgcn.image.sample";
1261 case ac_image_gather4
:
1262 name
= "llvm.amdgcn.image.gather4";
1265 name
= "llvm.amdgcn.image.load";
1267 case ac_image_load_mip
:
1268 name
= "llvm.amdgcn.image.load.mip";
1270 case ac_image_get_lod
:
1271 name
= "llvm.amdgcn.image.getlod";
1273 case ac_image_get_resinfo
:
1274 name
= "llvm.amdgcn.image.getresinfo";
1277 unreachable("invalid image opcode");
1280 ac_build_type_name_for_intr(LLVMTypeOf(args
[0]), type
,
1283 snprintf(intr_name
, sizeof(intr_name
), "%s%s%s%s.v4f32.%s.v8i32",
1285 a
->compare
? ".c" : "",
1289 a
->level_zero
? ".lz" : "",
1290 a
->offset
? ".o" : "",
1293 LLVMValueRef result
=
1294 ac_build_intrinsic(ctx
, intr_name
,
1295 ctx
->v4f32
, args
, num_args
,
1296 AC_FUNC_ATTR_READNONE
);
1298 result
= LLVMBuildBitCast(ctx
->builder
, result
,
1304 args
[num_args
++] = a
->addr
;
1305 args
[num_args
++] = a
->resource
;
1307 if (a
->opcode
== ac_image_load
||
1308 a
->opcode
== ac_image_load_mip
||
1309 a
->opcode
== ac_image_get_resinfo
) {
1310 dst_type
= ctx
->v4i32
;
1312 dst_type
= ctx
->v4f32
;
1313 args
[num_args
++] = a
->sampler
;
1316 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, 0);
1317 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->unorm
, 0);
1318 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* r128 */
1319 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->da
, 0);
1320 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* glc */
1321 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* slc */
1322 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* tfe */
1323 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* lwe */
1325 switch (a
->opcode
) {
1326 case ac_image_sample
:
1327 name
= "llvm.SI.image.sample";
1329 case ac_image_gather4
:
1330 name
= "llvm.SI.gather4";
1333 name
= "llvm.SI.image.load";
1335 case ac_image_load_mip
:
1336 name
= "llvm.SI.image.load.mip";
1338 case ac_image_get_lod
:
1339 name
= "llvm.SI.getlod";
1341 case ac_image_get_resinfo
:
1342 name
= "llvm.SI.getresinfo";
1346 ac_build_type_name_for_intr(LLVMTypeOf(a
->addr
), type
, sizeof(type
));
1347 snprintf(intr_name
, sizeof(intr_name
), "%s%s%s%s.%s",
1349 a
->compare
? ".c" : "",
1353 a
->level_zero
? ".lz" : "",
1354 a
->offset
? ".o" : "",
1357 return ac_build_intrinsic(ctx
, intr_name
,
1358 dst_type
, args
, num_args
,
1359 AC_FUNC_ATTR_READNONE
|
1360 AC_FUNC_ATTR_LEGACY
);
1363 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
1364 LLVMValueRef args
[2])
1366 if (HAVE_LLVM
>= 0x0500) {
1368 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
1370 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz",
1372 AC_FUNC_ATTR_READNONE
);
1373 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1376 return ac_build_intrinsic(ctx
, "llvm.SI.packf16", ctx
->i32
, args
, 2,
1377 AC_FUNC_ATTR_READNONE
|
1378 AC_FUNC_ATTR_LEGACY
);
1382 * KILL, AKA discard in GLSL.
1384 * \param value kill if value < 0.0 or value == NULL.
1386 void ac_build_kill(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
1389 ac_build_intrinsic(ctx
, "llvm.AMDGPU.kill", ctx
->voidt
,
1390 &value
, 1, AC_FUNC_ATTR_LEGACY
);
1392 ac_build_intrinsic(ctx
, "llvm.AMDGPU.kilp", ctx
->voidt
,
1393 NULL
, 0, AC_FUNC_ATTR_LEGACY
);
1397 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
1398 LLVMValueRef offset
, LLVMValueRef width
,
1401 LLVMValueRef args
[] = {
1407 if (HAVE_LLVM
>= 0x0500) {
1408 return ac_build_intrinsic(ctx
,
1409 is_signed
? "llvm.amdgcn.sbfe.i32" :
1410 "llvm.amdgcn.ubfe.i32",
1412 AC_FUNC_ATTR_READNONE
);
1415 return ac_build_intrinsic(ctx
,
1416 is_signed
? "llvm.AMDGPU.bfe.i32" :
1417 "llvm.AMDGPU.bfe.u32",
1419 AC_FUNC_ATTR_READNONE
|
1420 AC_FUNC_ATTR_LEGACY
);
1423 void ac_get_image_intr_name(const char *base_name
,
1424 LLVMTypeRef data_type
,
1425 LLVMTypeRef coords_type
,
1426 LLVMTypeRef rsrc_type
,
1427 char *out_name
, unsigned out_len
)
1429 char coords_type_name
[8];
1431 ac_build_type_name_for_intr(coords_type
, coords_type_name
,
1432 sizeof(coords_type_name
));
1434 if (HAVE_LLVM
<= 0x0309) {
1435 snprintf(out_name
, out_len
, "%s.%s", base_name
, coords_type_name
);
1437 char data_type_name
[8];
1438 char rsrc_type_name
[8];
1440 ac_build_type_name_for_intr(data_type
, data_type_name
,
1441 sizeof(data_type_name
));
1442 ac_build_type_name_for_intr(rsrc_type
, rsrc_type_name
,
1443 sizeof(rsrc_type_name
));
1444 snprintf(out_name
, out_len
, "%s.%s.%s.%s", base_name
,
1445 data_type_name
, coords_type_name
, rsrc_type_name
);
1449 #define AC_EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
1450 #define AC_EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
1458 struct ac_vs_exp_chan
1462 enum ac_ir_type type
;
1465 struct ac_vs_exp_inst
{
1468 struct ac_vs_exp_chan chan
[4];
1471 struct ac_vs_exports
{
1473 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
1476 /* Return true if the PARAM export has been eliminated. */
1477 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
1478 uint32_t num_outputs
,
1479 struct ac_vs_exp_inst
*exp
)
1481 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
1482 bool is_zero
[4] = {}, is_one
[4] = {};
1484 for (i
= 0; i
< 4; i
++) {
1485 /* It's a constant expression. Undef outputs are eliminated too. */
1486 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
1489 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
1490 if (exp
->chan
[i
].const_float
== 0)
1492 else if (exp
->chan
[i
].const_float
== 1)
1495 return false; /* other constant */
1500 /* Only certain combinations of 0 and 1 can be eliminated. */
1501 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
1502 default_val
= is_zero
[3] ? 0 : 1;
1503 else if (is_one
[0] && is_one
[1] && is_one
[2])
1504 default_val
= is_zero
[3] ? 2 : 3;
1508 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
1509 LLVMInstructionEraseFromParent(exp
->inst
);
1511 /* Change OFFSET to DEFAULT_VAL. */
1512 for (i
= 0; i
< num_outputs
; i
++) {
1513 if (vs_output_param_offset
[i
] == exp
->offset
) {
1514 vs_output_param_offset
[i
] =
1515 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
1522 static bool ac_eliminate_duplicated_output(uint8_t *vs_output_param_offset
,
1523 uint32_t num_outputs
,
1524 struct ac_vs_exports
*processed
,
1525 struct ac_vs_exp_inst
*exp
)
1527 unsigned p
, copy_back_channels
= 0;
1529 /* See if the output is already in the list of processed outputs.
1530 * The LLVMValueRef comparison relies on SSA.
1532 for (p
= 0; p
< processed
->num
; p
++) {
1533 bool different
= false;
1535 for (unsigned j
= 0; j
< 4; j
++) {
1536 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
1537 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
1539 /* Treat undef as a match. */
1540 if (c2
->type
== AC_IR_UNDEF
)
1543 /* If c1 is undef but c2 isn't, we can copy c2 to c1
1544 * and consider the instruction duplicated.
1546 if (c1
->type
== AC_IR_UNDEF
) {
1547 copy_back_channels
|= 1 << j
;
1551 /* Test whether the channels are not equal. */
1552 if (c1
->type
!= c2
->type
||
1553 (c1
->type
== AC_IR_CONST
&&
1554 c1
->const_float
!= c2
->const_float
) ||
1555 (c1
->type
== AC_IR_VALUE
&&
1556 c1
->value
!= c2
->value
)) {
1564 copy_back_channels
= 0;
1566 if (p
== processed
->num
)
1569 /* If a match was found, but the matching export has undef where the new
1570 * one has a normal value, copy the normal value to the undef channel.
1572 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
1574 while (copy_back_channels
) {
1575 unsigned chan
= u_bit_scan(©_back_channels
);
1577 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
1578 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
1579 exp
->chan
[chan
].value
);
1580 match
->chan
[chan
] = exp
->chan
[chan
];
1583 /* The PARAM export is duplicated. Kill it. */
1584 LLVMInstructionEraseFromParent(exp
->inst
);
1586 /* Change OFFSET to the matching export. */
1587 for (unsigned i
= 0; i
< num_outputs
; i
++) {
1588 if (vs_output_param_offset
[i
] == exp
->offset
) {
1589 vs_output_param_offset
[i
] = match
->offset
;
1596 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
1597 LLVMValueRef main_fn
,
1598 uint8_t *vs_output_param_offset
,
1599 uint32_t num_outputs
,
1600 uint8_t *num_param_exports
)
1602 LLVMBasicBlockRef bb
;
1603 bool removed_any
= false;
1604 struct ac_vs_exports exports
;
1608 /* Process all LLVM instructions. */
1609 bb
= LLVMGetFirstBasicBlock(main_fn
);
1611 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
1614 LLVMValueRef cur
= inst
;
1615 inst
= LLVMGetNextInstruction(inst
);
1616 struct ac_vs_exp_inst exp
;
1618 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
1621 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
1623 if (!ac_llvm_is_function(callee
))
1626 const char *name
= LLVMGetValueName(callee
);
1627 unsigned num_args
= LLVMCountParams(callee
);
1629 /* Check if this is an export instruction. */
1630 if ((num_args
!= 9 && num_args
!= 8) ||
1631 (strcmp(name
, "llvm.SI.export") &&
1632 strcmp(name
, "llvm.amdgcn.exp.f32")))
1635 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
1636 unsigned target
= LLVMConstIntGetZExtValue(arg
);
1638 if (target
< V_008DFC_SQ_EXP_PARAM
)
1641 target
-= V_008DFC_SQ_EXP_PARAM
;
1643 /* Parse the instruction. */
1644 memset(&exp
, 0, sizeof(exp
));
1645 exp
.offset
= target
;
1648 for (unsigned i
= 0; i
< 4; i
++) {
1649 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
1651 exp
.chan
[i
].value
= v
;
1653 if (LLVMIsUndef(v
)) {
1654 exp
.chan
[i
].type
= AC_IR_UNDEF
;
1655 } else if (LLVMIsAConstantFP(v
)) {
1656 LLVMBool loses_info
;
1657 exp
.chan
[i
].type
= AC_IR_CONST
;
1658 exp
.chan
[i
].const_float
=
1659 LLVMConstRealGetDouble(v
, &loses_info
);
1661 exp
.chan
[i
].type
= AC_IR_VALUE
;
1665 /* Eliminate constant and duplicated PARAM exports. */
1666 if (ac_eliminate_const_output(vs_output_param_offset
,
1667 num_outputs
, &exp
) ||
1668 ac_eliminate_duplicated_output(vs_output_param_offset
,
1669 num_outputs
, &exports
,
1673 exports
.exp
[exports
.num
++] = exp
;
1676 bb
= LLVMGetNextBasicBlock(bb
);
1679 /* Remove holes in export memory due to removed PARAM exports.
1680 * This is done by renumbering all PARAM exports.
1683 uint8_t old_offset
[VARYING_SLOT_MAX
];
1686 /* Make a copy of the offsets. We need the old version while
1687 * we are modifying some of them. */
1688 memcpy(old_offset
, vs_output_param_offset
,
1689 sizeof(old_offset
));
1691 for (i
= 0; i
< exports
.num
; i
++) {
1692 unsigned offset
= exports
.exp
[i
].offset
;
1694 /* Update vs_output_param_offset. Multiple outputs can
1695 * have the same offset.
1697 for (out
= 0; out
< num_outputs
; out
++) {
1698 if (old_offset
[out
] == offset
)
1699 vs_output_param_offset
[out
] = i
;
1702 /* Change the PARAM offset in the instruction. */
1703 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
1704 LLVMConstInt(ctx
->i32
,
1705 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
1707 *num_param_exports
= exports
.num
;