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
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
80 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
81 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
82 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
83 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
84 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
86 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
87 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
89 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
92 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
93 "invariant.load", 14);
95 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
97 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
98 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
100 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
101 "amdgpu.uniform", 14);
103 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
107 ac_get_llvm_num_components(LLVMValueRef value
)
109 LLVMTypeRef type
= LLVMTypeOf(value
);
110 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
111 ? LLVMGetVectorSize(type
)
113 return num_components
;
117 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
121 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
126 return LLVMBuildExtractElement(ac
->builder
, value
,
127 LLVMConstInt(ac
->i32
, index
, false), "");
131 ac_get_type_size(LLVMTypeRef type
)
133 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
136 case LLVMIntegerTypeKind
:
137 return LLVMGetIntTypeWidth(type
) / 8;
138 case LLVMFloatTypeKind
:
140 case LLVMDoubleTypeKind
:
141 case LLVMPointerTypeKind
:
143 case LLVMVectorTypeKind
:
144 return LLVMGetVectorSize(type
) *
145 ac_get_type_size(LLVMGetElementType(type
));
146 case LLVMArrayTypeKind
:
147 return LLVMGetArrayLength(type
) *
148 ac_get_type_size(LLVMGetElementType(type
));
155 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
157 if (t
== ctx
->f16
|| t
== ctx
->i16
)
159 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
161 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
164 unreachable("Unhandled integer size");
168 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
170 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
171 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
172 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
173 LLVMGetVectorSize(t
));
175 return to_integer_type_scalar(ctx
, t
);
179 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
181 LLVMTypeRef type
= LLVMTypeOf(v
);
182 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
185 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
187 if (t
== ctx
->i16
|| t
== ctx
->f16
)
189 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
191 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
194 unreachable("Unhandled float size");
198 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
200 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
201 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
202 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
203 LLVMGetVectorSize(t
));
205 return to_float_type_scalar(ctx
, t
);
209 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
211 LLVMTypeRef type
= LLVMTypeOf(v
);
212 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
217 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
218 LLVMTypeRef return_type
, LLVMValueRef
*params
,
219 unsigned param_count
, unsigned attrib_mask
)
221 LLVMValueRef function
, call
;
222 bool set_callsite_attrs
= HAVE_LLVM
>= 0x0400 &&
223 !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
225 function
= LLVMGetNamedFunction(ctx
->module
, name
);
227 LLVMTypeRef param_types
[32], function_type
;
230 assert(param_count
<= 32);
232 for (i
= 0; i
< param_count
; ++i
) {
234 param_types
[i
] = LLVMTypeOf(params
[i
]);
237 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
238 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
240 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
241 LLVMSetLinkage(function
, LLVMExternalLinkage
);
243 if (!set_callsite_attrs
)
244 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
247 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
248 if (set_callsite_attrs
)
249 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
254 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
257 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
259 LLVMTypeRef elem_type
= type
;
261 assert(bufsize
>= 8);
263 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
264 int ret
= snprintf(buf
, bufsize
, "v%u",
265 LLVMGetVectorSize(type
));
267 char *type_name
= LLVMPrintTypeToString(type
);
268 fprintf(stderr
, "Error building type name for: %s\n",
272 elem_type
= LLVMGetElementType(type
);
276 switch (LLVMGetTypeKind(elem_type
)) {
278 case LLVMIntegerTypeKind
:
279 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
281 case LLVMFloatTypeKind
:
282 snprintf(buf
, bufsize
, "f32");
284 case LLVMDoubleTypeKind
:
285 snprintf(buf
, bufsize
, "f64");
291 * Helper function that builds an LLVM IR PHI node and immediately adds
295 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
296 unsigned count_incoming
, LLVMValueRef
*values
,
297 LLVMBasicBlockRef
*blocks
)
299 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
300 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
304 /* Prevent optimizations (at least of memory accesses) across the current
305 * point in the program by emitting empty inline assembly that is marked as
306 * having side effects.
308 * Optionally, a value can be passed through the inline assembly to prevent
309 * LLVM from hoisting calls to ReadNone functions.
312 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
315 static int counter
= 0;
317 LLVMBuilderRef builder
= ctx
->builder
;
320 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
323 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
324 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
325 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
327 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
328 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
329 LLVMValueRef vgpr
= *pvgpr
;
330 LLVMTypeRef vgpr_type
= LLVMTypeOf(vgpr
);
331 unsigned vgpr_size
= ac_get_type_size(vgpr_type
);
334 assert(vgpr_size
% 4 == 0);
336 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
337 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
338 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
339 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
340 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
347 ac_build_ballot(struct ac_llvm_context
*ctx
,
350 LLVMValueRef args
[3] = {
353 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
356 /* We currently have no other way to prevent LLVM from lifting the icmp
357 * calls to a dominating basic block.
359 ac_build_optimization_barrier(ctx
, &args
[0]);
361 if (LLVMTypeOf(args
[0]) != ctx
->i32
)
362 args
[0] = LLVMBuildBitCast(ctx
->builder
, args
[0], ctx
->i32
, "");
364 return ac_build_intrinsic(ctx
,
365 "llvm.amdgcn.icmp.i32",
367 AC_FUNC_ATTR_NOUNWIND
|
368 AC_FUNC_ATTR_READNONE
|
369 AC_FUNC_ATTR_CONVERGENT
);
373 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
375 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
376 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
377 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
381 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
383 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
384 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
385 LLVMConstInt(ctx
->i64
, 0, 0), "");
389 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
391 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
392 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
394 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
395 vote_set
, active_set
, "");
396 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
398 LLVMConstInt(ctx
->i64
, 0, 0), "");
399 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
403 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
404 unsigned value_count
, unsigned component
)
406 LLVMValueRef vec
= NULL
;
408 if (value_count
== 1) {
409 return values
[component
];
410 } else if (!value_count
)
411 unreachable("value_count is 0");
413 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
414 LLVMValueRef value
= values
[i
];
417 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
418 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
419 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
425 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
426 LLVMValueRef
*values
,
427 unsigned value_count
,
428 unsigned value_stride
,
432 LLVMBuilderRef builder
= ctx
->builder
;
433 LLVMValueRef vec
= NULL
;
436 if (value_count
== 1 && !always_vector
) {
438 return LLVMBuildLoad(builder
, values
[0], "");
440 } else if (!value_count
)
441 unreachable("value_count is 0");
443 for (i
= 0; i
< value_count
; i
++) {
444 LLVMValueRef value
= values
[i
* value_stride
];
446 value
= LLVMBuildLoad(builder
, value
, "");
449 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
450 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
451 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
457 ac_build_gather_values(struct ac_llvm_context
*ctx
,
458 LLVMValueRef
*values
,
459 unsigned value_count
)
461 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
464 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
465 * with undef. Extract at most num_channels components from the input.
467 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
469 unsigned num_channels
)
471 LLVMTypeRef elemtype
;
472 LLVMValueRef chan
[4];
474 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
475 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
476 num_channels
= MIN2(num_channels
, vec_size
);
478 if (num_channels
>= 4)
481 for (unsigned i
= 0; i
< num_channels
; i
++)
482 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
484 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
487 assert(num_channels
== 1);
490 elemtype
= LLVMTypeOf(value
);
493 while (num_channels
< 4)
494 chan
[num_channels
++] = LLVMGetUndef(elemtype
);
496 return ac_build_gather_values(ctx
, chan
, 4);
500 ac_build_fdiv(struct ac_llvm_context
*ctx
,
504 LLVMValueRef ret
= LLVMBuildFDiv(ctx
->builder
, num
, den
, "");
506 /* Use v_rcp_f32 instead of precise division. */
507 if (!LLVMIsConstant(ret
))
508 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
512 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
513 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
514 * already multiplied by two. id is the cube face number.
516 struct cube_selection_coords
{
523 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
525 struct cube_selection_coords
*out
)
527 LLVMTypeRef f32
= ctx
->f32
;
529 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
530 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
531 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
532 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
533 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
534 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
535 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
536 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
540 * Build a manual selection sequence for cube face sc/tc coordinates and
541 * major axis vector (multiplied by 2 for consistency) for the given
542 * vec3 \p coords, for the face implied by \p selcoords.
544 * For the major axis, we always adjust the sign to be in the direction of
545 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
546 * the selcoords major axis.
548 static void build_cube_select(struct ac_llvm_context
*ctx
,
549 const struct cube_selection_coords
*selcoords
,
550 const LLVMValueRef
*coords
,
551 LLVMValueRef
*out_st
,
552 LLVMValueRef
*out_ma
)
554 LLVMBuilderRef builder
= ctx
->builder
;
555 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
556 LLVMValueRef is_ma_positive
;
558 LLVMValueRef is_ma_z
, is_not_ma_z
;
559 LLVMValueRef is_ma_y
;
560 LLVMValueRef is_ma_x
;
564 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
565 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
566 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
567 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
569 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
570 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
571 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
572 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
573 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
576 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
577 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
578 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
579 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
580 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
583 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
584 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
585 LLVMConstReal(f32
, -1.0), "");
586 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
589 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
590 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
591 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
592 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
593 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
597 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
598 bool is_deriv
, bool is_array
, bool is_lod
,
599 LLVMValueRef
*coords_arg
,
600 LLVMValueRef
*derivs_arg
)
603 LLVMBuilderRef builder
= ctx
->builder
;
604 struct cube_selection_coords selcoords
;
605 LLVMValueRef coords
[3];
608 if (is_array
&& !is_lod
) {
609 LLVMValueRef tmp
= coords_arg
[3];
610 tmp
= ac_build_intrinsic(ctx
, "llvm.rint.f32", ctx
->f32
, &tmp
, 1, 0);
612 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
614 * "For Array forms, the array layer used will be
616 * max(0, min(d−1, floor(layer+0.5)))
618 * where d is the depth of the texture array and layer
619 * comes from the component indicated in the tables below.
620 * Workaroudn for an issue where the layer is taken from a
621 * helper invocation which happens to fall on a different
622 * layer due to extrapolation."
624 * VI and earlier attempt to implement this in hardware by
625 * clamping the value of coords[2] = (8 * layer) + face.
626 * Unfortunately, this means that the we end up with the wrong
627 * face when clamping occurs.
629 * Clamp the layer earlier to work around the issue.
631 if (ctx
->chip_class
<= VI
) {
633 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
634 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
640 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
642 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
643 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
644 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
646 for (int i
= 0; i
< 2; ++i
)
647 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
649 coords
[2] = selcoords
.id
;
651 if (is_deriv
&& derivs_arg
) {
652 LLVMValueRef derivs
[4];
655 /* Convert cube derivatives to 2D derivatives. */
656 for (axis
= 0; axis
< 2; axis
++) {
657 LLVMValueRef deriv_st
[2];
658 LLVMValueRef deriv_ma
;
660 /* Transform the derivative alongside the texture
661 * coordinate. Mathematically, the correct formula is
662 * as follows. Assume we're projecting onto the +Z face
663 * and denote by dx/dh the derivative of the (original)
664 * X texture coordinate with respect to horizontal
665 * window coordinates. The projection onto the +Z face
670 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
671 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
673 * This motivatives the implementation below.
675 * Whether this actually gives the expected results for
676 * apps that might feed in derivatives obtained via
677 * finite differences is anyone's guess. The OpenGL spec
678 * seems awfully quiet about how textureGrad for cube
679 * maps should be handled.
681 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
682 deriv_st
, &deriv_ma
);
684 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
686 for (int i
= 0; i
< 2; ++i
)
687 derivs
[axis
* 2 + i
] =
688 LLVMBuildFSub(builder
,
689 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
690 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
693 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
696 /* Shift the texture coordinate. This must be applied after the
697 * derivative calculation.
699 for (int i
= 0; i
< 2; ++i
)
700 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
703 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
704 /* coords_arg.w component - array_index for cube arrays */
705 LLVMValueRef tmp
= LLVMBuildFMul(ctx
->builder
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), "");
706 coords
[2] = LLVMBuildFAdd(ctx
->builder
, tmp
, coords
[2], "");
709 memcpy(coords_arg
, coords
, sizeof(coords
));
714 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
715 LLVMValueRef llvm_chan
,
716 LLVMValueRef attr_number
,
721 LLVMValueRef args
[5];
724 if (HAVE_LLVM
< 0x0400) {
726 ij
[0] = LLVMBuildBitCast(ctx
->builder
, i
, ctx
->i32
, "");
727 ij
[1] = LLVMBuildBitCast(ctx
->builder
, j
, ctx
->i32
, "");
730 args
[1] = attr_number
;
732 args
[3] = ac_build_gather_values(ctx
, ij
, 2);
733 return ac_build_intrinsic(ctx
, "llvm.SI.fs.interp",
735 AC_FUNC_ATTR_READNONE
);
740 args
[2] = attr_number
;
743 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
744 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
749 args
[3] = attr_number
;
752 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
753 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
757 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
758 LLVMValueRef parameter
,
759 LLVMValueRef llvm_chan
,
760 LLVMValueRef attr_number
,
763 LLVMValueRef args
[4];
764 if (HAVE_LLVM
< 0x0400) {
766 args
[1] = attr_number
;
769 return ac_build_intrinsic(ctx
,
770 "llvm.SI.fs.constant",
772 AC_FUNC_ATTR_READNONE
);
777 args
[2] = attr_number
;
780 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
781 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
785 ac_build_gep0(struct ac_llvm_context
*ctx
,
786 LLVMValueRef base_ptr
,
789 LLVMValueRef indices
[2] = {
790 LLVMConstInt(ctx
->i32
, 0, 0),
793 return LLVMBuildGEP(ctx
->builder
, base_ptr
,
798 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
799 LLVMValueRef base_ptr
, LLVMValueRef index
,
802 LLVMBuildStore(ctx
->builder
, value
,
803 ac_build_gep0(ctx
, base_ptr
, index
));
807 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
808 * It's equivalent to doing a load from &base_ptr[index].
810 * \param base_ptr Where the array starts.
811 * \param index The element index into the array.
812 * \param uniform Whether the base_ptr and index can be assumed to be
813 * dynamically uniform (i.e. load to an SGPR)
814 * \param invariant Whether the load is invariant (no other opcodes affect it)
817 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
818 LLVMValueRef index
, bool uniform
, bool invariant
)
820 LLVMValueRef pointer
, result
;
822 pointer
= ac_build_gep0(ctx
, base_ptr
, index
);
824 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
825 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
827 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
831 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
834 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false);
837 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
838 LLVMValueRef base_ptr
, LLVMValueRef index
)
840 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true);
843 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
844 LLVMValueRef base_ptr
, LLVMValueRef index
)
846 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true);
849 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
850 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
851 * or v4i32 (num_channels=3,4).
854 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
857 unsigned num_channels
,
858 LLVMValueRef voffset
,
859 LLVMValueRef soffset
,
860 unsigned inst_offset
,
863 bool writeonly_memory
,
864 bool swizzle_enable_hint
)
866 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
867 * (voffset is swizzled, but soffset isn't swizzled).
868 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
870 if (!swizzle_enable_hint
) {
871 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
873 if (num_channels
== 3) {
874 LLVMValueRef v
[3], v01
;
876 for (int i
= 0; i
< 3; i
++) {
877 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
878 LLVMConstInt(ctx
->i32
, i
, 0), "");
880 v01
= ac_build_gather_values(ctx
, v
, 2);
882 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
883 soffset
, inst_offset
, glc
, slc
,
884 writeonly_memory
, swizzle_enable_hint
);
885 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
886 soffset
, inst_offset
+ 8,
888 writeonly_memory
, swizzle_enable_hint
);
892 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
893 static const char *types
[] = {"f32", "v2f32", "v4f32"};
895 LLVMValueRef offset
= soffset
;
898 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
899 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
901 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
903 LLVMValueRef args
[] = {
904 ac_to_float(ctx
, vdata
),
905 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
906 LLVMConstInt(ctx
->i32
, 0, 0),
908 LLVMConstInt(ctx
->i1
, glc
, 0),
909 LLVMConstInt(ctx
->i1
, slc
, 0),
912 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.%s",
915 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
916 args
, ARRAY_SIZE(args
),
918 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
:
919 AC_FUNC_ATTR_WRITEONLY
);
923 static unsigned dfmt
[] = {
924 V_008F0C_BUF_DATA_FORMAT_32
,
925 V_008F0C_BUF_DATA_FORMAT_32_32
,
926 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
927 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
929 assert(num_channels
>= 1 && num_channels
<= 4);
931 LLVMValueRef args
[] = {
934 LLVMConstInt(ctx
->i32
, num_channels
, 0),
935 voffset
? voffset
: LLVMGetUndef(ctx
->i32
),
937 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
938 LLVMConstInt(ctx
->i32
, dfmt
[num_channels
- 1], 0),
939 LLVMConstInt(ctx
->i32
, V_008F0C_BUF_NUM_FORMAT_UINT
, 0),
940 LLVMConstInt(ctx
->i32
, voffset
!= NULL
, 0),
941 LLVMConstInt(ctx
->i32
, 0, 0), /* idxen */
942 LLVMConstInt(ctx
->i32
, glc
, 0),
943 LLVMConstInt(ctx
->i32
, slc
, 0),
944 LLVMConstInt(ctx
->i32
, 0, 0), /* tfe*/
947 /* The instruction offset field has 12 bits */
948 assert(voffset
|| inst_offset
< (1 << 12));
950 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
951 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
952 const char *types
[] = {"i32", "v2i32", "v4i32"};
954 snprintf(name
, sizeof(name
), "llvm.SI.tbuffer.store.%s", types
[func
]);
956 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
957 args
, ARRAY_SIZE(args
),
958 AC_FUNC_ATTR_LEGACY
);
962 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
965 LLVMValueRef voffset
,
966 unsigned num_channels
,
972 LLVMValueRef args
[] = {
973 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
974 vindex
? vindex
: LLVMConstInt(ctx
->i32
, 0, 0),
976 LLVMConstInt(ctx
->i1
, glc
, 0),
977 LLVMConstInt(ctx
->i1
, slc
, 0)
979 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
981 LLVMTypeRef types
[] = {ctx
->f32
, ctx
->v2f32
, ctx
->v4f32
};
982 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
986 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.format.%s",
989 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.%s",
993 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
995 ac_get_load_intr_attribs(can_speculate
));
999 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1002 LLVMValueRef vindex
,
1003 LLVMValueRef voffset
,
1004 LLVMValueRef soffset
,
1005 unsigned inst_offset
,
1011 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1013 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1015 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1017 /* TODO: VI and later generations can use SMEM with GLC=1.*/
1018 if (allow_smem
&& !glc
&& !slc
) {
1019 assert(vindex
== NULL
);
1021 LLVMValueRef result
[8];
1023 for (int i
= 0; i
< num_channels
; i
++) {
1025 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1026 LLVMConstInt(ctx
->i32
, 4, 0), "");
1028 LLVMValueRef args
[2] = {rsrc
, offset
};
1029 result
[i
] = ac_build_intrinsic(ctx
, "llvm.SI.load.const.v4i32",
1031 AC_FUNC_ATTR_READNONE
|
1032 AC_FUNC_ATTR_LEGACY
);
1034 if (num_channels
== 1)
1037 if (num_channels
== 3)
1038 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1039 return ac_build_gather_values(ctx
, result
, num_channels
);
1042 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, offset
,
1043 num_channels
, glc
, slc
,
1044 can_speculate
, false);
1047 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1049 LLVMValueRef vindex
,
1050 LLVMValueRef voffset
,
1051 unsigned num_channels
,
1054 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1055 num_channels
, false, false,
1056 can_speculate
, true);
1060 * Set range metadata on an instruction. This can only be used on load and
1061 * call instructions. If you know an instruction can only produce the values
1062 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1063 * \p lo is the minimum value inclusive.
1064 * \p hi is the maximum value exclusive.
1066 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1067 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1069 LLVMValueRef range_md
, md_args
[2];
1070 LLVMTypeRef type
= LLVMTypeOf(value
);
1071 LLVMContextRef context
= LLVMGetTypeContext(type
);
1073 md_args
[0] = LLVMConstInt(type
, lo
, false);
1074 md_args
[1] = LLVMConstInt(type
, hi
, false);
1075 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1076 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
1080 ac_get_thread_id(struct ac_llvm_context
*ctx
)
1084 LLVMValueRef tid_args
[2];
1085 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
1086 tid_args
[1] = LLVMConstInt(ctx
->i32
, 0, false);
1087 tid_args
[1] = ac_build_intrinsic(ctx
,
1088 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
1089 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
1091 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
1093 2, AC_FUNC_ATTR_READNONE
);
1094 set_range_metadata(ctx
, tid
, 0, 64);
1099 * SI implements derivatives using the local data store (LDS)
1100 * All writes to the LDS happen in all executing threads at
1101 * the same time. TID is the Thread ID for the current
1102 * thread and is a value between 0 and 63, representing
1103 * the thread's position in the wavefront.
1105 * For the pixel shader threads are grouped into quads of four pixels.
1106 * The TIDs of the pixels of a quad are:
1114 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1115 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1116 * the current pixel's column, and masking with 0xfffffffe yields the TID
1117 * of the left pixel of the current pixel's row.
1119 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1120 * adding 2 yields the TID of the pixel below the top pixel.
1123 ac_build_ddxy(struct ac_llvm_context
*ctx
,
1128 LLVMValueRef tl
, trbl
, args
[2];
1129 LLVMValueRef result
;
1131 if (ctx
->chip_class
>= VI
) {
1132 LLVMValueRef thread_id
, tl_tid
, trbl_tid
;
1133 thread_id
= ac_get_thread_id(ctx
);
1135 tl_tid
= LLVMBuildAnd(ctx
->builder
, thread_id
,
1136 LLVMConstInt(ctx
->i32
, mask
, false), "");
1138 trbl_tid
= LLVMBuildAdd(ctx
->builder
, tl_tid
,
1139 LLVMConstInt(ctx
->i32
, idx
, false), "");
1141 args
[0] = LLVMBuildMul(ctx
->builder
, tl_tid
,
1142 LLVMConstInt(ctx
->i32
, 4, false), "");
1144 tl
= ac_build_intrinsic(ctx
,
1145 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
1147 AC_FUNC_ATTR_READNONE
|
1148 AC_FUNC_ATTR_CONVERGENT
);
1150 args
[0] = LLVMBuildMul(ctx
->builder
, trbl_tid
,
1151 LLVMConstInt(ctx
->i32
, 4, false), "");
1152 trbl
= ac_build_intrinsic(ctx
,
1153 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
1155 AC_FUNC_ATTR_READNONE
|
1156 AC_FUNC_ATTR_CONVERGENT
);
1158 uint32_t masks
[2] = {};
1161 case AC_TID_MASK_TOP_LEFT
:
1169 case AC_TID_MASK_TOP
:
1173 case AC_TID_MASK_LEFT
:
1182 args
[1] = LLVMConstInt(ctx
->i32
, masks
[0], false);
1184 tl
= ac_build_intrinsic(ctx
,
1185 "llvm.amdgcn.ds.swizzle", ctx
->i32
,
1187 AC_FUNC_ATTR_READNONE
|
1188 AC_FUNC_ATTR_CONVERGENT
);
1190 args
[1] = LLVMConstInt(ctx
->i32
, masks
[1], false);
1191 trbl
= ac_build_intrinsic(ctx
,
1192 "llvm.amdgcn.ds.swizzle", ctx
->i32
,
1194 AC_FUNC_ATTR_READNONE
|
1195 AC_FUNC_ATTR_CONVERGENT
);
1198 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, ctx
->f32
, "");
1199 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, ctx
->f32
, "");
1200 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
1205 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
1207 LLVMValueRef wave_id
)
1209 LLVMValueRef args
[2];
1210 const char *intr_name
= (HAVE_LLVM
< 0x0400) ? "llvm.SI.sendmsg" : "llvm.amdgcn.s.sendmsg";
1211 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
1213 ac_build_intrinsic(ctx
, intr_name
, ctx
->voidt
, args
, 2, 0);
1217 ac_build_imsb(struct ac_llvm_context
*ctx
,
1219 LLVMTypeRef dst_type
)
1221 const char *intr_name
= (HAVE_LLVM
< 0x0400) ? "llvm.AMDGPU.flbit.i32" :
1222 "llvm.amdgcn.sffbh.i32";
1223 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intr_name
,
1225 AC_FUNC_ATTR_READNONE
);
1227 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1228 * the index from LSB. Invert it by doing "31 - msb". */
1229 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1232 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
1233 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
1234 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1235 arg
, LLVMConstInt(ctx
->i32
, 0, 0), ""),
1236 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1237 arg
, all_ones
, ""), "");
1239 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
1243 ac_build_umsb(struct ac_llvm_context
*ctx
,
1245 LLVMTypeRef dst_type
)
1247 LLVMValueRef args
[2] = {
1251 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.ctlz.i32",
1252 dst_type
, args
, ARRAY_SIZE(args
),
1253 AC_FUNC_ATTR_READNONE
);
1255 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1256 * the index from LSB. Invert it by doing "31 - msb". */
1257 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1260 /* check for zero */
1261 return LLVMBuildSelect(ctx
->builder
,
1262 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
,
1263 LLVMConstInt(ctx
->i32
, 0, 0), ""),
1264 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
1267 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1270 LLVMValueRef args
[2] = {a
, b
};
1271 return ac_build_intrinsic(ctx
, "llvm.minnum.f32", ctx
->f32
, args
, 2,
1272 AC_FUNC_ATTR_READNONE
);
1275 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1278 LLVMValueRef args
[2] = {a
, b
};
1279 return ac_build_intrinsic(ctx
, "llvm.maxnum.f32", ctx
->f32
, args
, 2,
1280 AC_FUNC_ATTR_READNONE
);
1283 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1286 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
1287 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1290 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
1292 if (HAVE_LLVM
>= 0x0500) {
1293 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, ctx
->f32_0
),
1297 LLVMValueRef args
[3] = {
1299 LLVMConstReal(ctx
->f32
, 0),
1300 LLVMConstReal(ctx
->f32
, 1),
1303 return ac_build_intrinsic(ctx
, "llvm.AMDGPU.clamp.", ctx
->f32
, args
, 3,
1304 AC_FUNC_ATTR_READNONE
|
1305 AC_FUNC_ATTR_LEGACY
);
1308 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
1310 LLVMValueRef args
[9];
1312 if (HAVE_LLVM
>= 0x0500) {
1313 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
1314 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
1317 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
1318 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
1320 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
1322 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
1324 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1325 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1327 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
1328 ctx
->voidt
, args
, 6, 0);
1330 args
[2] = a
->out
[0];
1331 args
[3] = a
->out
[1];
1332 args
[4] = a
->out
[2];
1333 args
[5] = a
->out
[3];
1334 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1335 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1337 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
1338 ctx
->voidt
, args
, 8, 0);
1343 args
[0] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
1344 args
[1] = LLVMConstInt(ctx
->i32
, a
->valid_mask
, 0);
1345 args
[2] = LLVMConstInt(ctx
->i32
, a
->done
, 0);
1346 args
[3] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
1347 args
[4] = LLVMConstInt(ctx
->i32
, a
->compr
, 0);
1348 memcpy(args
+ 5, a
->out
, sizeof(a
->out
[0]) * 4);
1350 ac_build_intrinsic(ctx
, "llvm.SI.export", ctx
->voidt
, args
, 9,
1351 AC_FUNC_ATTR_LEGACY
);
1354 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
1355 struct ac_image_args
*a
)
1357 LLVMTypeRef dst_type
;
1358 LLVMValueRef args
[11];
1359 unsigned num_args
= 0;
1360 const char *name
= NULL
;
1361 char intr_name
[128], type
[64];
1363 if (HAVE_LLVM
>= 0x0400) {
1364 bool sample
= a
->opcode
== ac_image_sample
||
1365 a
->opcode
== ac_image_gather4
||
1366 a
->opcode
== ac_image_get_lod
;
1369 args
[num_args
++] = ac_to_float(ctx
, a
->addr
);
1371 args
[num_args
++] = a
->addr
;
1373 args
[num_args
++] = a
->resource
;
1375 args
[num_args
++] = a
->sampler
;
1376 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, 0);
1378 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, 0);
1379 args
[num_args
++] = ctx
->i1false
; /* glc */
1380 args
[num_args
++] = ctx
->i1false
; /* slc */
1381 args
[num_args
++] = ctx
->i1false
; /* lwe */
1382 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->da
, 0);
1384 switch (a
->opcode
) {
1385 case ac_image_sample
:
1386 name
= "llvm.amdgcn.image.sample";
1388 case ac_image_gather4
:
1389 name
= "llvm.amdgcn.image.gather4";
1392 name
= "llvm.amdgcn.image.load";
1394 case ac_image_load_mip
:
1395 name
= "llvm.amdgcn.image.load.mip";
1397 case ac_image_get_lod
:
1398 name
= "llvm.amdgcn.image.getlod";
1400 case ac_image_get_resinfo
:
1401 name
= "llvm.amdgcn.image.getresinfo";
1404 unreachable("invalid image opcode");
1407 ac_build_type_name_for_intr(LLVMTypeOf(args
[0]), type
,
1410 snprintf(intr_name
, sizeof(intr_name
), "%s%s%s%s.v4f32.%s.v8i32",
1412 a
->compare
? ".c" : "",
1416 a
->level_zero
? ".lz" : "",
1417 a
->offset
? ".o" : "",
1420 LLVMValueRef result
=
1421 ac_build_intrinsic(ctx
, intr_name
,
1422 ctx
->v4f32
, args
, num_args
,
1423 AC_FUNC_ATTR_READNONE
);
1425 result
= LLVMBuildBitCast(ctx
->builder
, result
,
1431 args
[num_args
++] = a
->addr
;
1432 args
[num_args
++] = a
->resource
;
1434 if (a
->opcode
== ac_image_load
||
1435 a
->opcode
== ac_image_load_mip
||
1436 a
->opcode
== ac_image_get_resinfo
) {
1437 dst_type
= ctx
->v4i32
;
1439 dst_type
= ctx
->v4f32
;
1440 args
[num_args
++] = a
->sampler
;
1443 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, 0);
1444 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->unorm
, 0);
1445 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* r128 */
1446 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->da
, 0);
1447 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* glc */
1448 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* slc */
1449 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* tfe */
1450 args
[num_args
++] = LLVMConstInt(ctx
->i32
, 0, 0); /* lwe */
1452 switch (a
->opcode
) {
1453 case ac_image_sample
:
1454 name
= "llvm.SI.image.sample";
1456 case ac_image_gather4
:
1457 name
= "llvm.SI.gather4";
1460 name
= "llvm.SI.image.load";
1462 case ac_image_load_mip
:
1463 name
= "llvm.SI.image.load.mip";
1465 case ac_image_get_lod
:
1466 name
= "llvm.SI.getlod";
1468 case ac_image_get_resinfo
:
1469 name
= "llvm.SI.getresinfo";
1473 ac_build_type_name_for_intr(LLVMTypeOf(a
->addr
), type
, sizeof(type
));
1474 snprintf(intr_name
, sizeof(intr_name
), "%s%s%s%s.%s",
1476 a
->compare
? ".c" : "",
1480 a
->level_zero
? ".lz" : "",
1481 a
->offset
? ".o" : "",
1484 return ac_build_intrinsic(ctx
, intr_name
,
1485 dst_type
, args
, num_args
,
1486 AC_FUNC_ATTR_READNONE
|
1487 AC_FUNC_ATTR_LEGACY
);
1490 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
1491 LLVMValueRef args
[2])
1493 if (HAVE_LLVM
>= 0x0500) {
1495 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
1497 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz",
1499 AC_FUNC_ATTR_READNONE
);
1500 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1503 return ac_build_intrinsic(ctx
, "llvm.SI.packf16", ctx
->i32
, args
, 2,
1504 AC_FUNC_ATTR_READNONE
|
1505 AC_FUNC_ATTR_LEGACY
);
1508 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
1510 assert(HAVE_LLVM
>= 0x0600);
1511 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
1512 &i1
, 1, AC_FUNC_ATTR_READNONE
);
1515 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
1517 if (HAVE_LLVM
>= 0x0600) {
1518 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
1523 LLVMValueRef value
= LLVMBuildSelect(ctx
->builder
, i1
,
1524 LLVMConstReal(ctx
->f32
, 1),
1525 LLVMConstReal(ctx
->f32
, -1), "");
1526 ac_build_intrinsic(ctx
, "llvm.AMDGPU.kill", ctx
->voidt
,
1527 &value
, 1, AC_FUNC_ATTR_LEGACY
);
1530 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
1531 LLVMValueRef offset
, LLVMValueRef width
,
1534 LLVMValueRef args
[] = {
1540 if (HAVE_LLVM
>= 0x0500) {
1541 return ac_build_intrinsic(ctx
,
1542 is_signed
? "llvm.amdgcn.sbfe.i32" :
1543 "llvm.amdgcn.ubfe.i32",
1545 AC_FUNC_ATTR_READNONE
);
1548 return ac_build_intrinsic(ctx
,
1549 is_signed
? "llvm.AMDGPU.bfe.i32" :
1550 "llvm.AMDGPU.bfe.u32",
1552 AC_FUNC_ATTR_READNONE
|
1553 AC_FUNC_ATTR_LEGACY
);
1556 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned simm16
)
1558 LLVMValueRef args
[1] = {
1559 LLVMConstInt(ctx
->i32
, simm16
, false),
1561 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
1562 ctx
->voidt
, args
, 1, 0);
1565 void ac_get_image_intr_name(const char *base_name
,
1566 LLVMTypeRef data_type
,
1567 LLVMTypeRef coords_type
,
1568 LLVMTypeRef rsrc_type
,
1569 char *out_name
, unsigned out_len
)
1571 char coords_type_name
[8];
1573 ac_build_type_name_for_intr(coords_type
, coords_type_name
,
1574 sizeof(coords_type_name
));
1576 if (HAVE_LLVM
<= 0x0309) {
1577 snprintf(out_name
, out_len
, "%s.%s", base_name
, coords_type_name
);
1579 char data_type_name
[8];
1580 char rsrc_type_name
[8];
1582 ac_build_type_name_for_intr(data_type
, data_type_name
,
1583 sizeof(data_type_name
));
1584 ac_build_type_name_for_intr(rsrc_type
, rsrc_type_name
,
1585 sizeof(rsrc_type_name
));
1586 snprintf(out_name
, out_len
, "%s.%s.%s.%s", base_name
,
1587 data_type_name
, coords_type_name
, rsrc_type_name
);
1591 #define AC_EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
1592 #define AC_EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
1600 struct ac_vs_exp_chan
1604 enum ac_ir_type type
;
1607 struct ac_vs_exp_inst
{
1610 struct ac_vs_exp_chan chan
[4];
1613 struct ac_vs_exports
{
1615 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
1618 /* Return true if the PARAM export has been eliminated. */
1619 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
1620 uint32_t num_outputs
,
1621 struct ac_vs_exp_inst
*exp
)
1623 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
1624 bool is_zero
[4] = {}, is_one
[4] = {};
1626 for (i
= 0; i
< 4; i
++) {
1627 /* It's a constant expression. Undef outputs are eliminated too. */
1628 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
1631 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
1632 if (exp
->chan
[i
].const_float
== 0)
1634 else if (exp
->chan
[i
].const_float
== 1)
1637 return false; /* other constant */
1642 /* Only certain combinations of 0 and 1 can be eliminated. */
1643 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
1644 default_val
= is_zero
[3] ? 0 : 1;
1645 else if (is_one
[0] && is_one
[1] && is_one
[2])
1646 default_val
= is_zero
[3] ? 2 : 3;
1650 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
1651 LLVMInstructionEraseFromParent(exp
->inst
);
1653 /* Change OFFSET to DEFAULT_VAL. */
1654 for (i
= 0; i
< num_outputs
; i
++) {
1655 if (vs_output_param_offset
[i
] == exp
->offset
) {
1656 vs_output_param_offset
[i
] =
1657 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
1664 static bool ac_eliminate_duplicated_output(uint8_t *vs_output_param_offset
,
1665 uint32_t num_outputs
,
1666 struct ac_vs_exports
*processed
,
1667 struct ac_vs_exp_inst
*exp
)
1669 unsigned p
, copy_back_channels
= 0;
1671 /* See if the output is already in the list of processed outputs.
1672 * The LLVMValueRef comparison relies on SSA.
1674 for (p
= 0; p
< processed
->num
; p
++) {
1675 bool different
= false;
1677 for (unsigned j
= 0; j
< 4; j
++) {
1678 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
1679 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
1681 /* Treat undef as a match. */
1682 if (c2
->type
== AC_IR_UNDEF
)
1685 /* If c1 is undef but c2 isn't, we can copy c2 to c1
1686 * and consider the instruction duplicated.
1688 if (c1
->type
== AC_IR_UNDEF
) {
1689 copy_back_channels
|= 1 << j
;
1693 /* Test whether the channels are not equal. */
1694 if (c1
->type
!= c2
->type
||
1695 (c1
->type
== AC_IR_CONST
&&
1696 c1
->const_float
!= c2
->const_float
) ||
1697 (c1
->type
== AC_IR_VALUE
&&
1698 c1
->value
!= c2
->value
)) {
1706 copy_back_channels
= 0;
1708 if (p
== processed
->num
)
1711 /* If a match was found, but the matching export has undef where the new
1712 * one has a normal value, copy the normal value to the undef channel.
1714 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
1716 while (copy_back_channels
) {
1717 unsigned chan
= u_bit_scan(©_back_channels
);
1719 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
1720 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
1721 exp
->chan
[chan
].value
);
1722 match
->chan
[chan
] = exp
->chan
[chan
];
1725 /* The PARAM export is duplicated. Kill it. */
1726 LLVMInstructionEraseFromParent(exp
->inst
);
1728 /* Change OFFSET to the matching export. */
1729 for (unsigned i
= 0; i
< num_outputs
; i
++) {
1730 if (vs_output_param_offset
[i
] == exp
->offset
) {
1731 vs_output_param_offset
[i
] = match
->offset
;
1738 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
1739 LLVMValueRef main_fn
,
1740 uint8_t *vs_output_param_offset
,
1741 uint32_t num_outputs
,
1742 uint8_t *num_param_exports
)
1744 LLVMBasicBlockRef bb
;
1745 bool removed_any
= false;
1746 struct ac_vs_exports exports
;
1750 /* Process all LLVM instructions. */
1751 bb
= LLVMGetFirstBasicBlock(main_fn
);
1753 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
1756 LLVMValueRef cur
= inst
;
1757 inst
= LLVMGetNextInstruction(inst
);
1758 struct ac_vs_exp_inst exp
;
1760 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
1763 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
1765 if (!ac_llvm_is_function(callee
))
1768 const char *name
= LLVMGetValueName(callee
);
1769 unsigned num_args
= LLVMCountParams(callee
);
1771 /* Check if this is an export instruction. */
1772 if ((num_args
!= 9 && num_args
!= 8) ||
1773 (strcmp(name
, "llvm.SI.export") &&
1774 strcmp(name
, "llvm.amdgcn.exp.f32")))
1777 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
1778 unsigned target
= LLVMConstIntGetZExtValue(arg
);
1780 if (target
< V_008DFC_SQ_EXP_PARAM
)
1783 target
-= V_008DFC_SQ_EXP_PARAM
;
1785 /* Parse the instruction. */
1786 memset(&exp
, 0, sizeof(exp
));
1787 exp
.offset
= target
;
1790 for (unsigned i
= 0; i
< 4; i
++) {
1791 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
1793 exp
.chan
[i
].value
= v
;
1795 if (LLVMIsUndef(v
)) {
1796 exp
.chan
[i
].type
= AC_IR_UNDEF
;
1797 } else if (LLVMIsAConstantFP(v
)) {
1798 LLVMBool loses_info
;
1799 exp
.chan
[i
].type
= AC_IR_CONST
;
1800 exp
.chan
[i
].const_float
=
1801 LLVMConstRealGetDouble(v
, &loses_info
);
1803 exp
.chan
[i
].type
= AC_IR_VALUE
;
1807 /* Eliminate constant and duplicated PARAM exports. */
1808 if (ac_eliminate_const_output(vs_output_param_offset
,
1809 num_outputs
, &exp
) ||
1810 ac_eliminate_duplicated_output(vs_output_param_offset
,
1811 num_outputs
, &exports
,
1815 exports
.exp
[exports
.num
++] = exp
;
1818 bb
= LLVMGetNextBasicBlock(bb
);
1821 /* Remove holes in export memory due to removed PARAM exports.
1822 * This is done by renumbering all PARAM exports.
1825 uint8_t old_offset
[VARYING_SLOT_MAX
];
1828 /* Make a copy of the offsets. We need the old version while
1829 * we are modifying some of them. */
1830 memcpy(old_offset
, vs_output_param_offset
,
1831 sizeof(old_offset
));
1833 for (i
= 0; i
< exports
.num
; i
++) {
1834 unsigned offset
= exports
.exp
[i
].offset
;
1836 /* Update vs_output_param_offset. Multiple outputs can
1837 * have the same offset.
1839 for (out
= 0; out
< num_outputs
; out
++) {
1840 if (old_offset
[out
] == offset
)
1841 vs_output_param_offset
[out
] = i
;
1844 /* Change the PARAM offset in the instruction. */
1845 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
1846 LLVMConstInt(ctx
->i32
,
1847 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
1849 *num_param_exports
= exports
.num
;
1853 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
1855 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
1856 ac_build_intrinsic(ctx
,
1857 "llvm.amdgcn.init.exec", ctx
->voidt
,
1858 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
1861 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
1863 unsigned lds_size
= ctx
->chip_class
>= CIK
? 65536 : 32768;
1864 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
1865 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_LOCAL_ADDR_SPACE
),
1869 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
1870 LLVMValueRef dw_addr
)
1872 return ac_build_load(ctx
, ctx
->lds
, dw_addr
);
1875 void ac_lds_store(struct ac_llvm_context
*ctx
,
1876 LLVMValueRef dw_addr
,
1879 value
= ac_to_integer(ctx
, value
);
1880 ac_build_indexed_store(ctx
, ctx
->lds
,
1884 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
1885 LLVMTypeRef dst_type
,
1888 LLVMValueRef params
[2] = {
1891 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
1892 * add special code to check for x=0. The reason is that
1893 * the LLVM behavior for x=0 is different from what we
1894 * need here. However, LLVM also assumes that ffs(x) is
1895 * in [0, 31], but GLSL expects that ffs(0) = -1, so
1896 * a conditional assignment to handle 0 is still required.
1898 * The hardware already implements the correct behavior.
1900 LLVMConstInt(ctx
->i1
, 1, false),
1903 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, "llvm.cttz.i32", ctx
->i32
,
1905 AC_FUNC_ATTR_READNONE
);
1907 /* TODO: We need an intrinsic to skip this conditional. */
1908 /* Check for zero: */
1909 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
1912 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
1915 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
1917 return LLVMPointerType(LLVMArrayType(elem_type
, 0),
1918 AC_CONST_ADDR_SPACE
);