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"
40 #include "util/u_math.h"
43 #include "shader_enums.h"
45 #define AC_LLVM_INITIAL_CF_DEPTH 4
47 /* Data for if/else/endif and bgnloop/endloop control flow structures.
50 /* Loop exit or next part of if/else/endif. */
51 LLVMBasicBlockRef next_block
;
52 LLVMBasicBlockRef loop_entry_block
;
55 /* Initialize module-independent parts of the context.
57 * The caller is responsible for initializing ctx::module and ctx::builder.
60 ac_llvm_context_init(struct ac_llvm_context
*ctx
,
61 enum chip_class chip_class
, enum radeon_family family
)
65 ctx
->context
= LLVMContextCreate();
67 ctx
->chip_class
= chip_class
;
72 ctx
->voidt
= LLVMVoidTypeInContext(ctx
->context
);
73 ctx
->i1
= LLVMInt1TypeInContext(ctx
->context
);
74 ctx
->i8
= LLVMInt8TypeInContext(ctx
->context
);
75 ctx
->i16
= LLVMIntTypeInContext(ctx
->context
, 16);
76 ctx
->i32
= LLVMIntTypeInContext(ctx
->context
, 32);
77 ctx
->i64
= LLVMIntTypeInContext(ctx
->context
, 64);
78 ctx
->intptr
= ctx
->i32
;
79 ctx
->f16
= LLVMHalfTypeInContext(ctx
->context
);
80 ctx
->f32
= LLVMFloatTypeInContext(ctx
->context
);
81 ctx
->f64
= LLVMDoubleTypeInContext(ctx
->context
);
82 ctx
->v2i16
= LLVMVectorType(ctx
->i16
, 2);
83 ctx
->v2i32
= LLVMVectorType(ctx
->i32
, 2);
84 ctx
->v3i32
= LLVMVectorType(ctx
->i32
, 3);
85 ctx
->v4i32
= LLVMVectorType(ctx
->i32
, 4);
86 ctx
->v2f32
= LLVMVectorType(ctx
->f32
, 2);
87 ctx
->v4f32
= LLVMVectorType(ctx
->f32
, 4);
88 ctx
->v8i32
= LLVMVectorType(ctx
->i32
, 8);
90 ctx
->i16_0
= LLVMConstInt(ctx
->i16
, 0, false);
91 ctx
->i16_1
= LLVMConstInt(ctx
->i16
, 1, false);
92 ctx
->i32_0
= LLVMConstInt(ctx
->i32
, 0, false);
93 ctx
->i32_1
= LLVMConstInt(ctx
->i32
, 1, false);
94 ctx
->i64_0
= LLVMConstInt(ctx
->i64
, 0, false);
95 ctx
->i64_1
= LLVMConstInt(ctx
->i64
, 1, false);
96 ctx
->f32_0
= LLVMConstReal(ctx
->f32
, 0.0);
97 ctx
->f32_1
= LLVMConstReal(ctx
->f32
, 1.0);
98 ctx
->f64_0
= LLVMConstReal(ctx
->f64
, 0.0);
99 ctx
->f64_1
= LLVMConstReal(ctx
->f64
, 1.0);
101 ctx
->i1false
= LLVMConstInt(ctx
->i1
, 0, false);
102 ctx
->i1true
= LLVMConstInt(ctx
->i1
, 1, false);
104 ctx
->range_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
107 ctx
->invariant_load_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
108 "invariant.load", 14);
110 ctx
->fpmath_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
, "fpmath", 6);
112 args
[0] = LLVMConstReal(ctx
->f32
, 2.5);
113 ctx
->fpmath_md_2p5_ulp
= LLVMMDNodeInContext(ctx
->context
, args
, 1);
115 ctx
->uniform_md_kind
= LLVMGetMDKindIDInContext(ctx
->context
,
116 "amdgpu.uniform", 14);
118 ctx
->empty_md
= LLVMMDNodeInContext(ctx
->context
, NULL
, 0);
122 ac_llvm_context_dispose(struct ac_llvm_context
*ctx
)
126 ctx
->flow_depth_max
= 0;
130 ac_get_llvm_num_components(LLVMValueRef value
)
132 LLVMTypeRef type
= LLVMTypeOf(value
);
133 unsigned num_components
= LLVMGetTypeKind(type
) == LLVMVectorTypeKind
134 ? LLVMGetVectorSize(type
)
136 return num_components
;
140 ac_llvm_extract_elem(struct ac_llvm_context
*ac
,
144 if (LLVMGetTypeKind(LLVMTypeOf(value
)) != LLVMVectorTypeKind
) {
149 return LLVMBuildExtractElement(ac
->builder
, value
,
150 LLVMConstInt(ac
->i32
, index
, false), "");
154 ac_get_elem_bits(struct ac_llvm_context
*ctx
, LLVMTypeRef type
)
156 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
)
157 type
= LLVMGetElementType(type
);
159 if (LLVMGetTypeKind(type
) == LLVMIntegerTypeKind
)
160 return LLVMGetIntTypeWidth(type
);
162 if (type
== ctx
->f16
)
164 if (type
== ctx
->f32
)
166 if (type
== ctx
->f64
)
169 unreachable("Unhandled type kind in get_elem_bits");
173 ac_get_type_size(LLVMTypeRef type
)
175 LLVMTypeKind kind
= LLVMGetTypeKind(type
);
178 case LLVMIntegerTypeKind
:
179 return LLVMGetIntTypeWidth(type
) / 8;
180 case LLVMHalfTypeKind
:
182 case LLVMFloatTypeKind
:
184 case LLVMDoubleTypeKind
:
186 case LLVMPointerTypeKind
:
187 if (LLVMGetPointerAddressSpace(type
) == AC_ADDR_SPACE_CONST_32BIT
)
190 case LLVMVectorTypeKind
:
191 return LLVMGetVectorSize(type
) *
192 ac_get_type_size(LLVMGetElementType(type
));
193 case LLVMArrayTypeKind
:
194 return LLVMGetArrayLength(type
) *
195 ac_get_type_size(LLVMGetElementType(type
));
202 static LLVMTypeRef
to_integer_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
204 if (t
== ctx
->f16
|| t
== ctx
->i16
)
206 else if (t
== ctx
->f32
|| t
== ctx
->i32
)
208 else if (t
== ctx
->f64
|| t
== ctx
->i64
)
211 unreachable("Unhandled integer size");
215 ac_to_integer_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
217 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
218 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
219 return LLVMVectorType(to_integer_type_scalar(ctx
, elem_type
),
220 LLVMGetVectorSize(t
));
222 if (LLVMGetTypeKind(t
) == LLVMPointerTypeKind
) {
223 switch (LLVMGetPointerAddressSpace(t
)) {
224 case AC_ADDR_SPACE_GLOBAL
:
226 case AC_ADDR_SPACE_LDS
:
229 unreachable("unhandled address space");
232 return to_integer_type_scalar(ctx
, t
);
236 ac_to_integer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
238 LLVMTypeRef type
= LLVMTypeOf(v
);
239 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
) {
240 return LLVMBuildPtrToInt(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
242 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_integer_type(ctx
, type
), "");
246 ac_to_integer_or_pointer(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
248 LLVMTypeRef type
= LLVMTypeOf(v
);
249 if (LLVMGetTypeKind(type
) == LLVMPointerTypeKind
)
251 return ac_to_integer(ctx
, v
);
254 static LLVMTypeRef
to_float_type_scalar(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
256 if (t
== ctx
->i16
|| t
== ctx
->f16
)
258 else if (t
== ctx
->i32
|| t
== ctx
->f32
)
260 else if (t
== ctx
->i64
|| t
== ctx
->f64
)
263 unreachable("Unhandled float size");
267 ac_to_float_type(struct ac_llvm_context
*ctx
, LLVMTypeRef t
)
269 if (LLVMGetTypeKind(t
) == LLVMVectorTypeKind
) {
270 LLVMTypeRef elem_type
= LLVMGetElementType(t
);
271 return LLVMVectorType(to_float_type_scalar(ctx
, elem_type
),
272 LLVMGetVectorSize(t
));
274 return to_float_type_scalar(ctx
, t
);
278 ac_to_float(struct ac_llvm_context
*ctx
, LLVMValueRef v
)
280 LLVMTypeRef type
= LLVMTypeOf(v
);
281 return LLVMBuildBitCast(ctx
->builder
, v
, ac_to_float_type(ctx
, type
), "");
286 ac_build_intrinsic(struct ac_llvm_context
*ctx
, const char *name
,
287 LLVMTypeRef return_type
, LLVMValueRef
*params
,
288 unsigned param_count
, unsigned attrib_mask
)
290 LLVMValueRef function
, call
;
291 bool set_callsite_attrs
= !(attrib_mask
& AC_FUNC_ATTR_LEGACY
);
293 function
= LLVMGetNamedFunction(ctx
->module
, name
);
295 LLVMTypeRef param_types
[32], function_type
;
298 assert(param_count
<= 32);
300 for (i
= 0; i
< param_count
; ++i
) {
302 param_types
[i
] = LLVMTypeOf(params
[i
]);
305 LLVMFunctionType(return_type
, param_types
, param_count
, 0);
306 function
= LLVMAddFunction(ctx
->module
, name
, function_type
);
308 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
309 LLVMSetLinkage(function
, LLVMExternalLinkage
);
311 if (!set_callsite_attrs
)
312 ac_add_func_attributes(ctx
->context
, function
, attrib_mask
);
315 call
= LLVMBuildCall(ctx
->builder
, function
, params
, param_count
, "");
316 if (set_callsite_attrs
)
317 ac_add_func_attributes(ctx
->context
, call
, attrib_mask
);
322 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
325 void ac_build_type_name_for_intr(LLVMTypeRef type
, char *buf
, unsigned bufsize
)
327 LLVMTypeRef elem_type
= type
;
329 assert(bufsize
>= 8);
331 if (LLVMGetTypeKind(type
) == LLVMVectorTypeKind
) {
332 int ret
= snprintf(buf
, bufsize
, "v%u",
333 LLVMGetVectorSize(type
));
335 char *type_name
= LLVMPrintTypeToString(type
);
336 fprintf(stderr
, "Error building type name for: %s\n",
340 elem_type
= LLVMGetElementType(type
);
344 switch (LLVMGetTypeKind(elem_type
)) {
346 case LLVMIntegerTypeKind
:
347 snprintf(buf
, bufsize
, "i%d", LLVMGetIntTypeWidth(elem_type
));
349 case LLVMHalfTypeKind
:
350 snprintf(buf
, bufsize
, "f16");
352 case LLVMFloatTypeKind
:
353 snprintf(buf
, bufsize
, "f32");
355 case LLVMDoubleTypeKind
:
356 snprintf(buf
, bufsize
, "f64");
362 * Helper function that builds an LLVM IR PHI node and immediately adds
366 ac_build_phi(struct ac_llvm_context
*ctx
, LLVMTypeRef type
,
367 unsigned count_incoming
, LLVMValueRef
*values
,
368 LLVMBasicBlockRef
*blocks
)
370 LLVMValueRef phi
= LLVMBuildPhi(ctx
->builder
, type
, "");
371 LLVMAddIncoming(phi
, values
, blocks
, count_incoming
);
375 void ac_build_s_barrier(struct ac_llvm_context
*ctx
)
377 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.barrier", ctx
->voidt
, NULL
,
378 0, AC_FUNC_ATTR_CONVERGENT
);
381 /* Prevent optimizations (at least of memory accesses) across the current
382 * point in the program by emitting empty inline assembly that is marked as
383 * having side effects.
385 * Optionally, a value can be passed through the inline assembly to prevent
386 * LLVM from hoisting calls to ReadNone functions.
389 ac_build_optimization_barrier(struct ac_llvm_context
*ctx
,
392 static int counter
= 0;
394 LLVMBuilderRef builder
= ctx
->builder
;
397 snprintf(code
, sizeof(code
), "; %d", p_atomic_inc_return(&counter
));
400 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->voidt
, NULL
, 0, false);
401 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "", true, false);
402 LLVMBuildCall(builder
, inlineasm
, NULL
, 0, "");
404 LLVMTypeRef ftype
= LLVMFunctionType(ctx
->i32
, &ctx
->i32
, 1, false);
405 LLVMValueRef inlineasm
= LLVMConstInlineAsm(ftype
, code
, "=v,0", true, false);
406 LLVMValueRef vgpr
= *pvgpr
;
407 LLVMTypeRef vgpr_type
= LLVMTypeOf(vgpr
);
408 unsigned vgpr_size
= ac_get_type_size(vgpr_type
);
411 assert(vgpr_size
% 4 == 0);
413 vgpr
= LLVMBuildBitCast(builder
, vgpr
, LLVMVectorType(ctx
->i32
, vgpr_size
/ 4), "");
414 vgpr0
= LLVMBuildExtractElement(builder
, vgpr
, ctx
->i32_0
, "");
415 vgpr0
= LLVMBuildCall(builder
, inlineasm
, &vgpr0
, 1, "");
416 vgpr
= LLVMBuildInsertElement(builder
, vgpr
, vgpr0
, ctx
->i32_0
, "");
417 vgpr
= LLVMBuildBitCast(builder
, vgpr
, vgpr_type
, "");
424 ac_build_shader_clock(struct ac_llvm_context
*ctx
)
426 LLVMValueRef tmp
= ac_build_intrinsic(ctx
, "llvm.readcyclecounter",
427 ctx
->i64
, NULL
, 0, 0);
428 return LLVMBuildBitCast(ctx
->builder
, tmp
, ctx
->v2i32
, "");
432 ac_build_ballot(struct ac_llvm_context
*ctx
,
435 LLVMValueRef args
[3] = {
438 LLVMConstInt(ctx
->i32
, LLVMIntNE
, 0)
441 /* We currently have no other way to prevent LLVM from lifting the icmp
442 * calls to a dominating basic block.
444 ac_build_optimization_barrier(ctx
, &args
[0]);
446 args
[0] = ac_to_integer(ctx
, args
[0]);
448 return ac_build_intrinsic(ctx
,
449 "llvm.amdgcn.icmp.i32",
451 AC_FUNC_ATTR_NOUNWIND
|
452 AC_FUNC_ATTR_READNONE
|
453 AC_FUNC_ATTR_CONVERGENT
);
457 ac_build_vote_all(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
459 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
460 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
461 return LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, vote_set
, active_set
, "");
465 ac_build_vote_any(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
467 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
468 return LLVMBuildICmp(ctx
->builder
, LLVMIntNE
, vote_set
,
469 LLVMConstInt(ctx
->i64
, 0, 0), "");
473 ac_build_vote_eq(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
475 LLVMValueRef active_set
= ac_build_ballot(ctx
, ctx
->i32_1
);
476 LLVMValueRef vote_set
= ac_build_ballot(ctx
, value
);
478 LLVMValueRef all
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
479 vote_set
, active_set
, "");
480 LLVMValueRef none
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
482 LLVMConstInt(ctx
->i64
, 0, 0), "");
483 return LLVMBuildOr(ctx
->builder
, all
, none
, "");
487 ac_build_varying_gather_values(struct ac_llvm_context
*ctx
, LLVMValueRef
*values
,
488 unsigned value_count
, unsigned component
)
490 LLVMValueRef vec
= NULL
;
492 if (value_count
== 1) {
493 return values
[component
];
494 } else if (!value_count
)
495 unreachable("value_count is 0");
497 for (unsigned i
= component
; i
< value_count
+ component
; i
++) {
498 LLVMValueRef value
= values
[i
];
501 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
502 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
- component
, false);
503 vec
= LLVMBuildInsertElement(ctx
->builder
, vec
, value
, index
, "");
509 ac_build_gather_values_extended(struct ac_llvm_context
*ctx
,
510 LLVMValueRef
*values
,
511 unsigned value_count
,
512 unsigned value_stride
,
516 LLVMBuilderRef builder
= ctx
->builder
;
517 LLVMValueRef vec
= NULL
;
520 if (value_count
== 1 && !always_vector
) {
522 return LLVMBuildLoad(builder
, values
[0], "");
524 } else if (!value_count
)
525 unreachable("value_count is 0");
527 for (i
= 0; i
< value_count
; i
++) {
528 LLVMValueRef value
= values
[i
* value_stride
];
530 value
= LLVMBuildLoad(builder
, value
, "");
533 vec
= LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value
), value_count
));
534 LLVMValueRef index
= LLVMConstInt(ctx
->i32
, i
, false);
535 vec
= LLVMBuildInsertElement(builder
, vec
, value
, index
, "");
541 ac_build_gather_values(struct ac_llvm_context
*ctx
,
542 LLVMValueRef
*values
,
543 unsigned value_count
)
545 return ac_build_gather_values_extended(ctx
, values
, value_count
, 1, false, false);
548 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
549 * channels with undef. Extract at most src_channels components from the input.
551 LLVMValueRef
ac_build_expand(struct ac_llvm_context
*ctx
,
553 unsigned src_channels
,
554 unsigned dst_channels
)
556 LLVMTypeRef elemtype
;
557 LLVMValueRef chan
[dst_channels
];
559 if (LLVMGetTypeKind(LLVMTypeOf(value
)) == LLVMVectorTypeKind
) {
560 unsigned vec_size
= LLVMGetVectorSize(LLVMTypeOf(value
));
562 if (src_channels
== dst_channels
&& vec_size
== dst_channels
)
565 src_channels
= MIN2(src_channels
, vec_size
);
567 for (unsigned i
= 0; i
< src_channels
; i
++)
568 chan
[i
] = ac_llvm_extract_elem(ctx
, value
, i
);
570 elemtype
= LLVMGetElementType(LLVMTypeOf(value
));
573 assert(src_channels
== 1);
576 elemtype
= LLVMTypeOf(value
);
579 for (unsigned i
= src_channels
; i
< dst_channels
; i
++)
580 chan
[i
] = LLVMGetUndef(elemtype
);
582 return ac_build_gather_values(ctx
, chan
, dst_channels
);
585 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
586 * with undef. Extract at most num_channels components from the input.
588 LLVMValueRef
ac_build_expand_to_vec4(struct ac_llvm_context
*ctx
,
590 unsigned num_channels
)
592 return ac_build_expand(ctx
, value
, num_channels
, 4);
595 LLVMValueRef
ac_build_round(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
597 unsigned type_size
= ac_get_type_size(LLVMTypeOf(value
));
601 name
= "llvm.rint.f16";
602 else if (type_size
== 4)
603 name
= "llvm.rint.f32";
605 name
= "llvm.rint.f64";
607 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(value
), &value
, 1,
608 AC_FUNC_ATTR_READNONE
);
612 ac_build_fdiv(struct ac_llvm_context
*ctx
,
616 /* If we do (num / den), LLVM >= 7.0 does:
617 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
619 * If we do (num * (1 / den)), LLVM does:
620 * return num * v_rcp_f32(den);
622 LLVMValueRef one
= LLVMTypeOf(num
) == ctx
->f64
? ctx
->f64_1
: ctx
->f32_1
;
623 LLVMValueRef rcp
= LLVMBuildFDiv(ctx
->builder
, one
, den
, "");
624 LLVMValueRef ret
= LLVMBuildFMul(ctx
->builder
, num
, rcp
, "");
626 /* Use v_rcp_f32 instead of precise division. */
627 if (!LLVMIsConstant(ret
))
628 LLVMSetMetadata(ret
, ctx
->fpmath_md_kind
, ctx
->fpmath_md_2p5_ulp
);
632 /* See fast_idiv_by_const.h. */
633 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
634 LLVMValueRef
ac_build_fast_udiv(struct ac_llvm_context
*ctx
,
636 LLVMValueRef multiplier
,
637 LLVMValueRef pre_shift
,
638 LLVMValueRef post_shift
,
639 LLVMValueRef increment
)
641 LLVMBuilderRef builder
= ctx
->builder
;
643 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
644 num
= LLVMBuildMul(builder
,
645 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
646 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
647 num
= LLVMBuildAdd(builder
, num
,
648 LLVMBuildZExt(builder
, increment
, ctx
->i64
, ""), "");
649 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
650 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
651 return LLVMBuildLShr(builder
, num
, post_shift
, "");
654 /* See fast_idiv_by_const.h. */
655 /* If num != UINT_MAX, this more efficient version can be used. */
656 /* Set: increment = util_fast_udiv_info::increment; */
657 LLVMValueRef
ac_build_fast_udiv_nuw(struct ac_llvm_context
*ctx
,
659 LLVMValueRef multiplier
,
660 LLVMValueRef pre_shift
,
661 LLVMValueRef post_shift
,
662 LLVMValueRef increment
)
664 LLVMBuilderRef builder
= ctx
->builder
;
666 num
= LLVMBuildLShr(builder
, num
, pre_shift
, "");
667 num
= LLVMBuildNUWAdd(builder
, num
, increment
, "");
668 num
= LLVMBuildMul(builder
,
669 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
670 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
671 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
672 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
673 return LLVMBuildLShr(builder
, num
, post_shift
, "");
676 /* See fast_idiv_by_const.h. */
677 /* Both operands must fit in 31 bits and the divisor must not be 1. */
678 LLVMValueRef
ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context
*ctx
,
680 LLVMValueRef multiplier
,
681 LLVMValueRef post_shift
)
683 LLVMBuilderRef builder
= ctx
->builder
;
685 num
= LLVMBuildMul(builder
,
686 LLVMBuildZExt(builder
, num
, ctx
->i64
, ""),
687 LLVMBuildZExt(builder
, multiplier
, ctx
->i64
, ""), "");
688 num
= LLVMBuildLShr(builder
, num
, LLVMConstInt(ctx
->i64
, 32, 0), "");
689 num
= LLVMBuildTrunc(builder
, num
, ctx
->i32
, "");
690 return LLVMBuildLShr(builder
, num
, post_shift
, "");
693 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
694 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
695 * already multiplied by two. id is the cube face number.
697 struct cube_selection_coords
{
704 build_cube_intrinsic(struct ac_llvm_context
*ctx
,
706 struct cube_selection_coords
*out
)
708 LLVMTypeRef f32
= ctx
->f32
;
710 out
->stc
[1] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubetc",
711 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
712 out
->stc
[0] = ac_build_intrinsic(ctx
, "llvm.amdgcn.cubesc",
713 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
714 out
->ma
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubema",
715 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
716 out
->id
= ac_build_intrinsic(ctx
, "llvm.amdgcn.cubeid",
717 f32
, in
, 3, AC_FUNC_ATTR_READNONE
);
721 * Build a manual selection sequence for cube face sc/tc coordinates and
722 * major axis vector (multiplied by 2 for consistency) for the given
723 * vec3 \p coords, for the face implied by \p selcoords.
725 * For the major axis, we always adjust the sign to be in the direction of
726 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
727 * the selcoords major axis.
729 static void build_cube_select(struct ac_llvm_context
*ctx
,
730 const struct cube_selection_coords
*selcoords
,
731 const LLVMValueRef
*coords
,
732 LLVMValueRef
*out_st
,
733 LLVMValueRef
*out_ma
)
735 LLVMBuilderRef builder
= ctx
->builder
;
736 LLVMTypeRef f32
= LLVMTypeOf(coords
[0]);
737 LLVMValueRef is_ma_positive
;
739 LLVMValueRef is_ma_z
, is_not_ma_z
;
740 LLVMValueRef is_ma_y
;
741 LLVMValueRef is_ma_x
;
745 is_ma_positive
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
746 selcoords
->ma
, LLVMConstReal(f32
, 0.0), "");
747 sgn_ma
= LLVMBuildSelect(builder
, is_ma_positive
,
748 LLVMConstReal(f32
, 1.0), LLVMConstReal(f32
, -1.0), "");
750 is_ma_z
= LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 4.0), "");
751 is_not_ma_z
= LLVMBuildNot(builder
, is_ma_z
, "");
752 is_ma_y
= LLVMBuildAnd(builder
, is_not_ma_z
,
753 LLVMBuildFCmp(builder
, LLVMRealUGE
, selcoords
->id
, LLVMConstReal(f32
, 2.0), ""), "");
754 is_ma_x
= LLVMBuildAnd(builder
, is_not_ma_z
, LLVMBuildNot(builder
, is_ma_y
, ""), "");
757 tmp
= LLVMBuildSelect(builder
, is_ma_x
, coords
[2], coords
[0], "");
758 sgn
= LLVMBuildSelect(builder
, is_ma_y
, LLVMConstReal(f32
, 1.0),
759 LLVMBuildSelect(builder
, is_ma_z
, sgn_ma
,
760 LLVMBuildFNeg(builder
, sgn_ma
, ""), ""), "");
761 out_st
[0] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
764 tmp
= LLVMBuildSelect(builder
, is_ma_y
, coords
[2], coords
[1], "");
765 sgn
= LLVMBuildSelect(builder
, is_ma_y
, sgn_ma
,
766 LLVMConstReal(f32
, -1.0), "");
767 out_st
[1] = LLVMBuildFMul(builder
, tmp
, sgn
, "");
770 tmp
= LLVMBuildSelect(builder
, is_ma_z
, coords
[2],
771 LLVMBuildSelect(builder
, is_ma_y
, coords
[1], coords
[0], ""), "");
772 tmp
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
773 ctx
->f32
, &tmp
, 1, AC_FUNC_ATTR_READNONE
);
774 *out_ma
= LLVMBuildFMul(builder
, tmp
, LLVMConstReal(f32
, 2.0), "");
778 ac_prepare_cube_coords(struct ac_llvm_context
*ctx
,
779 bool is_deriv
, bool is_array
, bool is_lod
,
780 LLVMValueRef
*coords_arg
,
781 LLVMValueRef
*derivs_arg
)
784 LLVMBuilderRef builder
= ctx
->builder
;
785 struct cube_selection_coords selcoords
;
786 LLVMValueRef coords
[3];
789 if (is_array
&& !is_lod
) {
790 LLVMValueRef tmp
= ac_build_round(ctx
, coords_arg
[3]);
792 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
794 * "For Array forms, the array layer used will be
796 * max(0, min(d−1, floor(layer+0.5)))
798 * where d is the depth of the texture array and layer
799 * comes from the component indicated in the tables below.
800 * Workaroudn for an issue where the layer is taken from a
801 * helper invocation which happens to fall on a different
802 * layer due to extrapolation."
804 * VI and earlier attempt to implement this in hardware by
805 * clamping the value of coords[2] = (8 * layer) + face.
806 * Unfortunately, this means that the we end up with the wrong
807 * face when clamping occurs.
809 * Clamp the layer earlier to work around the issue.
811 if (ctx
->chip_class
<= VI
) {
813 ge0
= LLVMBuildFCmp(builder
, LLVMRealOGE
, tmp
, ctx
->f32_0
, "");
814 tmp
= LLVMBuildSelect(builder
, ge0
, tmp
, ctx
->f32_0
, "");
820 build_cube_intrinsic(ctx
, coords_arg
, &selcoords
);
822 invma
= ac_build_intrinsic(ctx
, "llvm.fabs.f32",
823 ctx
->f32
, &selcoords
.ma
, 1, AC_FUNC_ATTR_READNONE
);
824 invma
= ac_build_fdiv(ctx
, LLVMConstReal(ctx
->f32
, 1.0), invma
);
826 for (int i
= 0; i
< 2; ++i
)
827 coords
[i
] = LLVMBuildFMul(builder
, selcoords
.stc
[i
], invma
, "");
829 coords
[2] = selcoords
.id
;
831 if (is_deriv
&& derivs_arg
) {
832 LLVMValueRef derivs
[4];
835 /* Convert cube derivatives to 2D derivatives. */
836 for (axis
= 0; axis
< 2; axis
++) {
837 LLVMValueRef deriv_st
[2];
838 LLVMValueRef deriv_ma
;
840 /* Transform the derivative alongside the texture
841 * coordinate. Mathematically, the correct formula is
842 * as follows. Assume we're projecting onto the +Z face
843 * and denote by dx/dh the derivative of the (original)
844 * X texture coordinate with respect to horizontal
845 * window coordinates. The projection onto the +Z face
850 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
851 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
853 * This motivatives the implementation below.
855 * Whether this actually gives the expected results for
856 * apps that might feed in derivatives obtained via
857 * finite differences is anyone's guess. The OpenGL spec
858 * seems awfully quiet about how textureGrad for cube
859 * maps should be handled.
861 build_cube_select(ctx
, &selcoords
, &derivs_arg
[axis
* 3],
862 deriv_st
, &deriv_ma
);
864 deriv_ma
= LLVMBuildFMul(builder
, deriv_ma
, invma
, "");
866 for (int i
= 0; i
< 2; ++i
)
867 derivs
[axis
* 2 + i
] =
868 LLVMBuildFSub(builder
,
869 LLVMBuildFMul(builder
, deriv_st
[i
], invma
, ""),
870 LLVMBuildFMul(builder
, deriv_ma
, coords
[i
], ""), "");
873 memcpy(derivs_arg
, derivs
, sizeof(derivs
));
876 /* Shift the texture coordinate. This must be applied after the
877 * derivative calculation.
879 for (int i
= 0; i
< 2; ++i
)
880 coords
[i
] = LLVMBuildFAdd(builder
, coords
[i
], LLVMConstReal(ctx
->f32
, 1.5), "");
883 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
884 /* coords_arg.w component - array_index for cube arrays */
885 coords
[2] = ac_build_fmad(ctx
, coords_arg
[3], LLVMConstReal(ctx
->f32
, 8.0), coords
[2]);
888 memcpy(coords_arg
, coords
, sizeof(coords
));
893 ac_build_fs_interp(struct ac_llvm_context
*ctx
,
894 LLVMValueRef llvm_chan
,
895 LLVMValueRef attr_number
,
900 LLVMValueRef args
[5];
905 args
[2] = attr_number
;
908 p1
= ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p1",
909 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
914 args
[3] = attr_number
;
917 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.p2",
918 ctx
->f32
, args
, 5, AC_FUNC_ATTR_READNONE
);
922 ac_build_fs_interp_mov(struct ac_llvm_context
*ctx
,
923 LLVMValueRef parameter
,
924 LLVMValueRef llvm_chan
,
925 LLVMValueRef attr_number
,
928 LLVMValueRef args
[4];
932 args
[2] = attr_number
;
935 return ac_build_intrinsic(ctx
, "llvm.amdgcn.interp.mov",
936 ctx
->f32
, args
, 4, AC_FUNC_ATTR_READNONE
);
940 ac_build_gep_ptr(struct ac_llvm_context
*ctx
,
941 LLVMValueRef base_ptr
,
944 return LLVMBuildGEP(ctx
->builder
, base_ptr
, &index
, 1, "");
948 ac_build_gep0(struct ac_llvm_context
*ctx
,
949 LLVMValueRef base_ptr
,
952 LLVMValueRef indices
[2] = {
956 return LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
959 LLVMValueRef
ac_build_pointer_add(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
962 return LLVMBuildPointerCast(ctx
->builder
,
963 ac_build_gep0(ctx
, ptr
, index
),
964 LLVMTypeOf(ptr
), "");
968 ac_build_indexed_store(struct ac_llvm_context
*ctx
,
969 LLVMValueRef base_ptr
, LLVMValueRef index
,
972 LLVMBuildStore(ctx
->builder
, value
,
973 ac_build_gep0(ctx
, base_ptr
, index
));
977 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
978 * It's equivalent to doing a load from &base_ptr[index].
980 * \param base_ptr Where the array starts.
981 * \param index The element index into the array.
982 * \param uniform Whether the base_ptr and index can be assumed to be
983 * dynamically uniform (i.e. load to an SGPR)
984 * \param invariant Whether the load is invariant (no other opcodes affect it)
985 * \param no_unsigned_wraparound
986 * For all possible re-associations and re-distributions of an expression
987 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
988 * without inbounds in base_ptr), this parameter is true if "addr + offset"
989 * does not result in an unsigned integer wraparound. This is used for
990 * optimal code generation of 32-bit pointer arithmetic.
992 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
993 * integer wraparound can't be an imm offset in s_load_dword, because
994 * the instruction performs "addr + offset" in 64 bits.
996 * Expected usage for bindless textures by chaining GEPs:
997 * // possible unsigned wraparound, don't use InBounds:
998 * ptr1 = LLVMBuildGEP(base_ptr, index);
999 * image = load(ptr1); // becomes "s_load ptr1, 0"
1001 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1002 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1005 ac_build_load_custom(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1006 LLVMValueRef index
, bool uniform
, bool invariant
,
1007 bool no_unsigned_wraparound
)
1009 LLVMValueRef pointer
, result
;
1010 LLVMValueRef indices
[2] = {ctx
->i32_0
, index
};
1012 if (no_unsigned_wraparound
&&
1013 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr
)) == AC_ADDR_SPACE_CONST_32BIT
)
1014 pointer
= LLVMBuildInBoundsGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1016 pointer
= LLVMBuildGEP(ctx
->builder
, base_ptr
, indices
, 2, "");
1019 LLVMSetMetadata(pointer
, ctx
->uniform_md_kind
, ctx
->empty_md
);
1020 result
= LLVMBuildLoad(ctx
->builder
, pointer
, "");
1022 LLVMSetMetadata(result
, ctx
->invariant_load_md_kind
, ctx
->empty_md
);
1026 LLVMValueRef
ac_build_load(struct ac_llvm_context
*ctx
, LLVMValueRef base_ptr
,
1029 return ac_build_load_custom(ctx
, base_ptr
, index
, false, false, false);
1032 LLVMValueRef
ac_build_load_invariant(struct ac_llvm_context
*ctx
,
1033 LLVMValueRef base_ptr
, LLVMValueRef index
)
1035 return ac_build_load_custom(ctx
, base_ptr
, index
, false, true, false);
1038 /* This assumes that there is no unsigned integer wraparound during the address
1039 * computation, excluding all GEPs within base_ptr. */
1040 LLVMValueRef
ac_build_load_to_sgpr(struct ac_llvm_context
*ctx
,
1041 LLVMValueRef base_ptr
, LLVMValueRef index
)
1043 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, true);
1046 /* See ac_build_load_custom() documentation. */
1047 LLVMValueRef
ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context
*ctx
,
1048 LLVMValueRef base_ptr
, LLVMValueRef index
)
1050 return ac_build_load_custom(ctx
, base_ptr
, index
, true, true, false);
1053 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1054 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1055 * or v4i32 (num_channels=3,4).
1058 ac_build_buffer_store_dword(struct ac_llvm_context
*ctx
,
1061 unsigned num_channels
,
1062 LLVMValueRef voffset
,
1063 LLVMValueRef soffset
,
1064 unsigned inst_offset
,
1067 bool writeonly_memory
,
1068 bool swizzle_enable_hint
)
1070 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
1072 if (num_channels
== 3) {
1073 LLVMValueRef v
[3], v01
;
1075 for (int i
= 0; i
< 3; i
++) {
1076 v
[i
] = LLVMBuildExtractElement(ctx
->builder
, vdata
,
1077 LLVMConstInt(ctx
->i32
, i
, 0), "");
1079 v01
= ac_build_gather_values(ctx
, v
, 2);
1081 ac_build_buffer_store_dword(ctx
, rsrc
, v01
, 2, voffset
,
1082 soffset
, inst_offset
, glc
, slc
,
1083 writeonly_memory
, swizzle_enable_hint
);
1084 ac_build_buffer_store_dword(ctx
, rsrc
, v
[2], 1, voffset
,
1085 soffset
, inst_offset
+ 8,
1087 writeonly_memory
, swizzle_enable_hint
);
1091 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1092 * (voffset is swizzled, but soffset isn't swizzled).
1093 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1095 if (!swizzle_enable_hint
) {
1096 LLVMValueRef offset
= soffset
;
1098 static const char *types
[] = {"f32", "v2f32", "v4f32"};
1101 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1102 LLVMConstInt(ctx
->i32
, inst_offset
, 0), "");
1104 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1106 LLVMValueRef args
[] = {
1107 ac_to_float(ctx
, vdata
),
1108 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
1111 LLVMConstInt(ctx
->i1
, glc
, 0),
1112 LLVMConstInt(ctx
->i1
, slc
, 0),
1116 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.store.%s",
1117 types
[CLAMP(num_channels
, 1, 3) - 1]);
1119 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
1120 args
, ARRAY_SIZE(args
),
1122 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
:
1123 AC_FUNC_ATTR_WRITEONLY
);
1127 static const unsigned dfmt
[] = {
1128 V_008F0C_BUF_DATA_FORMAT_32
,
1129 V_008F0C_BUF_DATA_FORMAT_32_32
,
1130 V_008F0C_BUF_DATA_FORMAT_32_32_32
,
1131 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1133 static const char *types
[] = {"i32", "v2i32", "v4i32"};
1134 LLVMValueRef args
[] = {
1136 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
1138 voffset
? voffset
: ctx
->i32_0
,
1140 LLVMConstInt(ctx
->i32
, inst_offset
, 0),
1141 LLVMConstInt(ctx
->i32
, dfmt
[num_channels
- 1], 0),
1142 LLVMConstInt(ctx
->i32
, V_008F0C_BUF_NUM_FORMAT_UINT
, 0),
1143 LLVMConstInt(ctx
->i1
, glc
, 0),
1144 LLVMConstInt(ctx
->i1
, slc
, 0),
1147 snprintf(name
, sizeof(name
), "llvm.amdgcn.tbuffer.store.%s",
1148 types
[CLAMP(num_channels
, 1, 3) - 1]);
1150 ac_build_intrinsic(ctx
, name
, ctx
->voidt
,
1151 args
, ARRAY_SIZE(args
),
1153 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY
:
1154 AC_FUNC_ATTR_WRITEONLY
);
1158 ac_build_buffer_load_common(struct ac_llvm_context
*ctx
,
1160 LLVMValueRef vindex
,
1161 LLVMValueRef voffset
,
1162 unsigned num_channels
,
1168 LLVMValueRef args
[] = {
1169 LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, ""),
1170 vindex
? vindex
: ctx
->i32_0
,
1172 LLVMConstInt(ctx
->i1
, glc
, 0),
1173 LLVMConstInt(ctx
->i1
, slc
, 0)
1175 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1177 LLVMTypeRef types
[] = {ctx
->f32
, ctx
->v2f32
, ctx
->v4f32
};
1178 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1182 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.format.%s",
1185 snprintf(name
, sizeof(name
), "llvm.amdgcn.buffer.load.%s",
1189 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
1191 ac_get_load_intr_attribs(can_speculate
));
1195 ac_build_llvm8_buffer_load_common(struct ac_llvm_context
*ctx
,
1197 LLVMValueRef vindex
,
1198 LLVMValueRef voffset
,
1199 LLVMValueRef soffset
,
1200 unsigned num_channels
,
1207 LLVMValueRef args
[5];
1209 args
[idx
++] = LLVMBuildBitCast(ctx
->builder
, rsrc
, ctx
->v4i32
, "");
1211 args
[idx
++] = vindex
? vindex
: ctx
->i32_0
;
1212 args
[idx
++] = voffset
? voffset
: ctx
->i32_0
;
1213 args
[idx
++] = soffset
? soffset
: ctx
->i32_0
;
1214 args
[idx
++] = LLVMConstInt(ctx
->i32
, (glc
? 1 : 0) + (slc
? 2 : 0), 0);
1215 unsigned func
= CLAMP(num_channels
, 1, 3) - 1;
1217 LLVMTypeRef types
[] = {ctx
->f32
, ctx
->v2f32
, ctx
->v4f32
};
1218 const char *type_names
[] = {"f32", "v2f32", "v4f32"};
1219 const char *indexing_kind
= structurized
? "struct" : "raw";
1223 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.format.%s",
1224 indexing_kind
, type_names
[func
]);
1226 snprintf(name
, sizeof(name
), "llvm.amdgcn.%s.buffer.load.%s",
1227 indexing_kind
, type_names
[func
]);
1230 return ac_build_intrinsic(ctx
, name
, types
[func
], args
,
1232 ac_get_load_intr_attribs(can_speculate
));
1236 ac_build_buffer_load(struct ac_llvm_context
*ctx
,
1239 LLVMValueRef vindex
,
1240 LLVMValueRef voffset
,
1241 LLVMValueRef soffset
,
1242 unsigned inst_offset
,
1248 LLVMValueRef offset
= LLVMConstInt(ctx
->i32
, inst_offset
, 0);
1250 offset
= LLVMBuildAdd(ctx
->builder
, offset
, voffset
, "");
1252 offset
= LLVMBuildAdd(ctx
->builder
, offset
, soffset
, "");
1254 if (allow_smem
&& !slc
&&
1255 (!glc
|| (HAVE_LLVM
>= 0x0800 && ctx
->chip_class
>= VI
))) {
1256 assert(vindex
== NULL
);
1258 LLVMValueRef result
[8];
1260 for (int i
= 0; i
< num_channels
; i
++) {
1262 offset
= LLVMBuildAdd(ctx
->builder
, offset
,
1263 LLVMConstInt(ctx
->i32
, 4, 0), "");
1265 const char *intrname
=
1266 HAVE_LLVM
>= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1267 : "llvm.SI.load.const.v4i32";
1268 unsigned num_args
= HAVE_LLVM
>= 0x0800 ? 3 : 2;
1269 LLVMValueRef args
[3] = {
1272 glc
? ctx
->i32_1
: ctx
->i32_0
,
1274 result
[i
] = ac_build_intrinsic(ctx
, intrname
,
1275 ctx
->f32
, args
, num_args
,
1276 AC_FUNC_ATTR_READNONE
|
1277 (HAVE_LLVM
< 0x0800 ? AC_FUNC_ATTR_LEGACY
: 0));
1279 if (num_channels
== 1)
1282 if (num_channels
== 3)
1283 result
[num_channels
++] = LLVMGetUndef(ctx
->f32
);
1284 return ac_build_gather_values(ctx
, result
, num_channels
);
1287 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, offset
,
1288 num_channels
, glc
, slc
,
1289 can_speculate
, false);
1292 LLVMValueRef
ac_build_buffer_load_format(struct ac_llvm_context
*ctx
,
1294 LLVMValueRef vindex
,
1295 LLVMValueRef voffset
,
1296 unsigned num_channels
,
1300 if (HAVE_LLVM
>= 0x800) {
1301 return ac_build_llvm8_buffer_load_common(ctx
, rsrc
, vindex
, voffset
, ctx
->i32_0
,
1302 num_channels
, glc
, false,
1303 can_speculate
, true, true);
1305 return ac_build_buffer_load_common(ctx
, rsrc
, vindex
, voffset
,
1306 num_channels
, glc
, false,
1307 can_speculate
, true);
1310 LLVMValueRef
ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context
*ctx
,
1312 LLVMValueRef vindex
,
1313 LLVMValueRef voffset
,
1314 unsigned num_channels
,
1318 if (HAVE_LLVM
>= 0x800) {
1319 return ac_build_llvm8_buffer_load_common(ctx
, rsrc
, vindex
, voffset
, ctx
->i32_0
,
1320 num_channels
, glc
, false,
1321 can_speculate
, true, true);
1324 LLVMValueRef elem_count
= LLVMBuildExtractElement(ctx
->builder
, rsrc
, LLVMConstInt(ctx
->i32
, 2, 0), "");
1325 LLVMValueRef stride
= LLVMBuildExtractElement(ctx
->builder
, rsrc
, ctx
->i32_1
, "");
1326 stride
= LLVMBuildLShr(ctx
->builder
, stride
, LLVMConstInt(ctx
->i32
, 16, 0), "");
1328 LLVMValueRef new_elem_count
= LLVMBuildSelect(ctx
->builder
,
1329 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, elem_count
, stride
, ""),
1330 elem_count
, stride
, "");
1332 LLVMValueRef new_rsrc
= LLVMBuildInsertElement(ctx
->builder
, rsrc
, new_elem_count
,
1333 LLVMConstInt(ctx
->i32
, 2, 0), "");
1335 return ac_build_buffer_load_common(ctx
, new_rsrc
, vindex
, voffset
,
1336 num_channels
, glc
, false,
1337 can_speculate
, true);
1341 ac_build_tbuffer_load_short(struct ac_llvm_context
*ctx
,
1343 LLVMValueRef vindex
,
1344 LLVMValueRef voffset
,
1345 LLVMValueRef soffset
,
1346 LLVMValueRef immoffset
,
1349 const char *name
= "llvm.amdgcn.tbuffer.load.i32";
1350 LLVMTypeRef type
= ctx
->i32
;
1351 LLVMValueRef params
[] = {
1357 LLVMConstInt(ctx
->i32
, V_008F0C_BUF_DATA_FORMAT_16
, false),
1358 LLVMConstInt(ctx
->i32
, V_008F0C_BUF_NUM_FORMAT_UINT
, false),
1362 LLVMValueRef res
= ac_build_intrinsic(ctx
, name
, type
, params
, 9, 0);
1363 return LLVMBuildTrunc(ctx
->builder
, res
, ctx
->i16
, "");
1367 * Set range metadata on an instruction. This can only be used on load and
1368 * call instructions. If you know an instruction can only produce the values
1369 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1370 * \p lo is the minimum value inclusive.
1371 * \p hi is the maximum value exclusive.
1373 static void set_range_metadata(struct ac_llvm_context
*ctx
,
1374 LLVMValueRef value
, unsigned lo
, unsigned hi
)
1376 LLVMValueRef range_md
, md_args
[2];
1377 LLVMTypeRef type
= LLVMTypeOf(value
);
1378 LLVMContextRef context
= LLVMGetTypeContext(type
);
1380 md_args
[0] = LLVMConstInt(type
, lo
, false);
1381 md_args
[1] = LLVMConstInt(type
, hi
, false);
1382 range_md
= LLVMMDNodeInContext(context
, md_args
, 2);
1383 LLVMSetMetadata(value
, ctx
->range_md_kind
, range_md
);
1387 ac_get_thread_id(struct ac_llvm_context
*ctx
)
1391 LLVMValueRef tid_args
[2];
1392 tid_args
[0] = LLVMConstInt(ctx
->i32
, 0xffffffff, false);
1393 tid_args
[1] = ctx
->i32_0
;
1394 tid_args
[1] = ac_build_intrinsic(ctx
,
1395 "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
1396 tid_args
, 2, AC_FUNC_ATTR_READNONE
);
1398 tid
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi",
1400 2, AC_FUNC_ATTR_READNONE
);
1401 set_range_metadata(ctx
, tid
, 0, 64);
1406 * SI implements derivatives using the local data store (LDS)
1407 * All writes to the LDS happen in all executing threads at
1408 * the same time. TID is the Thread ID for the current
1409 * thread and is a value between 0 and 63, representing
1410 * the thread's position in the wavefront.
1412 * For the pixel shader threads are grouped into quads of four pixels.
1413 * The TIDs of the pixels of a quad are:
1421 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1422 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1423 * the current pixel's column, and masking with 0xfffffffe yields the TID
1424 * of the left pixel of the current pixel's row.
1426 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1427 * adding 2 yields the TID of the pixel below the top pixel.
1430 ac_build_ddxy(struct ac_llvm_context
*ctx
,
1435 unsigned tl_lanes
[4], trbl_lanes
[4];
1436 LLVMValueRef tl
, trbl
;
1437 LLVMValueRef result
;
1439 for (unsigned i
= 0; i
< 4; ++i
) {
1440 tl_lanes
[i
] = i
& mask
;
1441 trbl_lanes
[i
] = (i
& mask
) + idx
;
1444 tl
= ac_build_quad_swizzle(ctx
, val
,
1445 tl_lanes
[0], tl_lanes
[1],
1446 tl_lanes
[2], tl_lanes
[3]);
1447 trbl
= ac_build_quad_swizzle(ctx
, val
,
1448 trbl_lanes
[0], trbl_lanes
[1],
1449 trbl_lanes
[2], trbl_lanes
[3]);
1451 tl
= LLVMBuildBitCast(ctx
->builder
, tl
, ctx
->f32
, "");
1452 trbl
= LLVMBuildBitCast(ctx
->builder
, trbl
, ctx
->f32
, "");
1453 result
= LLVMBuildFSub(ctx
->builder
, trbl
, tl
, "");
1455 result
= ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.f32", ctx
->f32
,
1462 ac_build_sendmsg(struct ac_llvm_context
*ctx
,
1464 LLVMValueRef wave_id
)
1466 LLVMValueRef args
[2];
1467 args
[0] = LLVMConstInt(ctx
->i32
, msg
, false);
1469 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.sendmsg", ctx
->voidt
, args
, 2, 0);
1473 ac_build_imsb(struct ac_llvm_context
*ctx
,
1475 LLVMTypeRef dst_type
)
1477 LLVMValueRef msb
= ac_build_intrinsic(ctx
, "llvm.amdgcn.sffbh.i32",
1479 AC_FUNC_ATTR_READNONE
);
1481 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1482 * the index from LSB. Invert it by doing "31 - msb". */
1483 msb
= LLVMBuildSub(ctx
->builder
, LLVMConstInt(ctx
->i32
, 31, false),
1486 LLVMValueRef all_ones
= LLVMConstInt(ctx
->i32
, -1, true);
1487 LLVMValueRef cond
= LLVMBuildOr(ctx
->builder
,
1488 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1489 arg
, ctx
->i32_0
, ""),
1490 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
,
1491 arg
, all_ones
, ""), "");
1493 return LLVMBuildSelect(ctx
->builder
, cond
, all_ones
, msb
, "");
1497 ac_build_umsb(struct ac_llvm_context
*ctx
,
1499 LLVMTypeRef dst_type
)
1501 const char *intrin_name
;
1503 LLVMValueRef highest_bit
;
1507 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(arg
));
1510 intrin_name
= "llvm.ctlz.i64";
1512 highest_bit
= LLVMConstInt(ctx
->i64
, 63, false);
1516 intrin_name
= "llvm.ctlz.i32";
1518 highest_bit
= LLVMConstInt(ctx
->i32
, 31, false);
1522 intrin_name
= "llvm.ctlz.i16";
1524 highest_bit
= LLVMConstInt(ctx
->i16
, 15, false);
1528 unreachable(!"invalid bitsize");
1532 LLVMValueRef params
[2] = {
1537 LLVMValueRef msb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
1539 AC_FUNC_ATTR_READNONE
);
1541 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1542 * the index from LSB. Invert it by doing "31 - msb". */
1543 msb
= LLVMBuildSub(ctx
->builder
, highest_bit
, msb
, "");
1544 msb
= LLVMBuildTruncOrBitCast(ctx
->builder
, msb
, ctx
->i32
, "");
1546 /* check for zero */
1547 return LLVMBuildSelect(ctx
->builder
,
1548 LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, arg
, zero
, ""),
1549 LLVMConstInt(ctx
->i32
, -1, true), msb
, "");
1552 LLVMValueRef
ac_build_fmin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1555 LLVMValueRef args
[2] = {a
, b
};
1556 return ac_build_intrinsic(ctx
, "llvm.minnum.f32", ctx
->f32
, args
, 2,
1557 AC_FUNC_ATTR_READNONE
);
1560 LLVMValueRef
ac_build_fmax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1563 LLVMValueRef args
[2] = {a
, b
};
1564 return ac_build_intrinsic(ctx
, "llvm.maxnum.f32", ctx
->f32
, args
, 2,
1565 AC_FUNC_ATTR_READNONE
);
1568 LLVMValueRef
ac_build_imin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1571 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSLE
, a
, b
, "");
1572 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1575 LLVMValueRef
ac_build_imax(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1578 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, a
, b
, "");
1579 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1582 LLVMValueRef
ac_build_umin(struct ac_llvm_context
*ctx
, LLVMValueRef a
,
1585 LLVMValueRef cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntULE
, a
, b
, "");
1586 return LLVMBuildSelect(ctx
->builder
, cmp
, a
, b
, "");
1589 LLVMValueRef
ac_build_clamp(struct ac_llvm_context
*ctx
, LLVMValueRef value
)
1591 return ac_build_fmin(ctx
, ac_build_fmax(ctx
, value
, ctx
->f32_0
),
1595 void ac_build_export(struct ac_llvm_context
*ctx
, struct ac_export_args
*a
)
1597 LLVMValueRef args
[9];
1599 args
[0] = LLVMConstInt(ctx
->i32
, a
->target
, 0);
1600 args
[1] = LLVMConstInt(ctx
->i32
, a
->enabled_channels
, 0);
1603 LLVMTypeRef i16
= LLVMInt16TypeInContext(ctx
->context
);
1604 LLVMTypeRef v2i16
= LLVMVectorType(i16
, 2);
1606 args
[2] = LLVMBuildBitCast(ctx
->builder
, a
->out
[0],
1608 args
[3] = LLVMBuildBitCast(ctx
->builder
, a
->out
[1],
1610 args
[4] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1611 args
[5] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1613 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.compr.v2i16",
1614 ctx
->voidt
, args
, 6, 0);
1616 args
[2] = a
->out
[0];
1617 args
[3] = a
->out
[1];
1618 args
[4] = a
->out
[2];
1619 args
[5] = a
->out
[3];
1620 args
[6] = LLVMConstInt(ctx
->i1
, a
->done
, 0);
1621 args
[7] = LLVMConstInt(ctx
->i1
, a
->valid_mask
, 0);
1623 ac_build_intrinsic(ctx
, "llvm.amdgcn.exp.f32",
1624 ctx
->voidt
, args
, 8, 0);
1628 void ac_build_export_null(struct ac_llvm_context
*ctx
)
1630 struct ac_export_args args
;
1632 args
.enabled_channels
= 0x0; /* enabled channels */
1633 args
.valid_mask
= 1; /* whether the EXEC mask is valid */
1634 args
.done
= 1; /* DONE bit */
1635 args
.target
= V_008DFC_SQ_EXP_NULL
;
1636 args
.compr
= 0; /* COMPR flag (0 = 32-bit export) */
1637 args
.out
[0] = LLVMGetUndef(ctx
->f32
); /* R */
1638 args
.out
[1] = LLVMGetUndef(ctx
->f32
); /* G */
1639 args
.out
[2] = LLVMGetUndef(ctx
->f32
); /* B */
1640 args
.out
[3] = LLVMGetUndef(ctx
->f32
); /* A */
1642 ac_build_export(ctx
, &args
);
1645 static unsigned ac_num_coords(enum ac_image_dim dim
)
1651 case ac_image_1darray
:
1655 case ac_image_2darray
:
1656 case ac_image_2dmsaa
:
1658 case ac_image_2darraymsaa
:
1661 unreachable("ac_num_coords: bad dim");
1665 static unsigned ac_num_derivs(enum ac_image_dim dim
)
1669 case ac_image_1darray
:
1672 case ac_image_2darray
:
1677 case ac_image_2dmsaa
:
1678 case ac_image_2darraymsaa
:
1680 unreachable("derivatives not supported");
1684 static const char *get_atomic_name(enum ac_atomic_op op
)
1687 case ac_atomic_swap
: return "swap";
1688 case ac_atomic_add
: return "add";
1689 case ac_atomic_sub
: return "sub";
1690 case ac_atomic_smin
: return "smin";
1691 case ac_atomic_umin
: return "umin";
1692 case ac_atomic_smax
: return "smax";
1693 case ac_atomic_umax
: return "umax";
1694 case ac_atomic_and
: return "and";
1695 case ac_atomic_or
: return "or";
1696 case ac_atomic_xor
: return "xor";
1698 unreachable("bad atomic op");
1701 LLVMValueRef
ac_build_image_opcode(struct ac_llvm_context
*ctx
,
1702 struct ac_image_args
*a
)
1704 const char *overload
[3] = { "", "", "" };
1705 unsigned num_overloads
= 0;
1706 LLVMValueRef args
[18];
1707 unsigned num_args
= 0;
1708 enum ac_image_dim dim
= a
->dim
;
1710 assert(!a
->lod
|| a
->lod
== ctx
->i32_0
|| a
->lod
== ctx
->f32_0
||
1712 assert((a
->opcode
!= ac_image_get_resinfo
&& a
->opcode
!= ac_image_load_mip
&&
1713 a
->opcode
!= ac_image_store_mip
) ||
1715 assert(a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
1716 (!a
->compare
&& !a
->offset
));
1717 assert((a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
||
1718 a
->opcode
== ac_image_get_lod
) ||
1720 assert((a
->bias
? 1 : 0) +
1722 (a
->level_zero
? 1 : 0) +
1723 (a
->derivs
[0] ? 1 : 0) <= 1);
1725 if (a
->opcode
== ac_image_get_lod
) {
1727 case ac_image_1darray
:
1730 case ac_image_2darray
:
1739 bool sample
= a
->opcode
== ac_image_sample
||
1740 a
->opcode
== ac_image_gather4
||
1741 a
->opcode
== ac_image_get_lod
;
1742 bool atomic
= a
->opcode
== ac_image_atomic
||
1743 a
->opcode
== ac_image_atomic_cmpswap
;
1744 LLVMTypeRef coord_type
= sample
? ctx
->f32
: ctx
->i32
;
1746 if (atomic
|| a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
) {
1747 args
[num_args
++] = a
->data
[0];
1748 if (a
->opcode
== ac_image_atomic_cmpswap
)
1749 args
[num_args
++] = a
->data
[1];
1753 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->dmask
, false);
1756 args
[num_args
++] = ac_to_integer(ctx
, a
->offset
);
1758 args
[num_args
++] = ac_to_float(ctx
, a
->bias
);
1759 overload
[num_overloads
++] = ".f32";
1762 args
[num_args
++] = ac_to_float(ctx
, a
->compare
);
1764 unsigned count
= ac_num_derivs(dim
);
1765 for (unsigned i
= 0; i
< count
; ++i
)
1766 args
[num_args
++] = ac_to_float(ctx
, a
->derivs
[i
]);
1767 overload
[num_overloads
++] = ".f32";
1769 unsigned num_coords
=
1770 a
->opcode
!= ac_image_get_resinfo
? ac_num_coords(dim
) : 0;
1771 for (unsigned i
= 0; i
< num_coords
; ++i
)
1772 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->coords
[i
], coord_type
, "");
1774 args
[num_args
++] = LLVMBuildBitCast(ctx
->builder
, a
->lod
, coord_type
, "");
1775 overload
[num_overloads
++] = sample
? ".f32" : ".i32";
1777 args
[num_args
++] = a
->resource
;
1779 args
[num_args
++] = a
->sampler
;
1780 args
[num_args
++] = LLVMConstInt(ctx
->i1
, a
->unorm
, false);
1783 args
[num_args
++] = ctx
->i32_0
; /* texfailctrl */
1784 args
[num_args
++] = LLVMConstInt(ctx
->i32
, a
->cache_policy
, false);
1787 const char *atomic_subop
= "";
1788 switch (a
->opcode
) {
1789 case ac_image_sample
: name
= "sample"; break;
1790 case ac_image_gather4
: name
= "gather4"; break;
1791 case ac_image_load
: name
= "load"; break;
1792 case ac_image_load_mip
: name
= "load.mip"; break;
1793 case ac_image_store
: name
= "store"; break;
1794 case ac_image_store_mip
: name
= "store.mip"; break;
1795 case ac_image_atomic
:
1797 atomic_subop
= get_atomic_name(a
->atomic
);
1799 case ac_image_atomic_cmpswap
:
1801 atomic_subop
= "cmpswap";
1803 case ac_image_get_lod
: name
= "getlod"; break;
1804 case ac_image_get_resinfo
: name
= "getresinfo"; break;
1805 default: unreachable("invalid image opcode");
1808 const char *dimname
;
1810 case ac_image_1d
: dimname
= "1d"; break;
1811 case ac_image_2d
: dimname
= "2d"; break;
1812 case ac_image_3d
: dimname
= "3d"; break;
1813 case ac_image_cube
: dimname
= "cube"; break;
1814 case ac_image_1darray
: dimname
= "1darray"; break;
1815 case ac_image_2darray
: dimname
= "2darray"; break;
1816 case ac_image_2dmsaa
: dimname
= "2dmsaa"; break;
1817 case ac_image_2darraymsaa
: dimname
= "2darraymsaa"; break;
1818 default: unreachable("invalid dim");
1822 a
->lod
&& (a
->opcode
== ac_image_sample
|| a
->opcode
== ac_image_gather4
);
1824 snprintf(intr_name
, sizeof(intr_name
),
1825 "llvm.amdgcn.image.%s%s" /* base name */
1826 "%s%s%s" /* sample/gather modifiers */
1827 ".%s.%s%s%s%s", /* dimension and type overloads */
1829 a
->compare
? ".c" : "",
1832 a
->derivs
[0] ? ".d" :
1833 a
->level_zero
? ".lz" : "",
1834 a
->offset
? ".o" : "",
1836 atomic
? "i32" : "v4f32",
1837 overload
[0], overload
[1], overload
[2]);
1842 else if (a
->opcode
== ac_image_store
|| a
->opcode
== ac_image_store_mip
)
1847 LLVMValueRef result
=
1848 ac_build_intrinsic(ctx
, intr_name
, retty
, args
, num_args
,
1850 if (!sample
&& retty
== ctx
->v4f32
) {
1851 result
= LLVMBuildBitCast(ctx
->builder
, result
,
1857 LLVMValueRef
ac_build_cvt_pkrtz_f16(struct ac_llvm_context
*ctx
,
1858 LLVMValueRef args
[2])
1861 LLVMVectorType(LLVMHalfTypeInContext(ctx
->context
), 2);
1863 return ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pkrtz", v2f16
,
1864 args
, 2, AC_FUNC_ATTR_READNONE
);
1867 LLVMValueRef
ac_build_cvt_pknorm_i16(struct ac_llvm_context
*ctx
,
1868 LLVMValueRef args
[2])
1871 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.i16",
1872 ctx
->v2i16
, args
, 2,
1873 AC_FUNC_ATTR_READNONE
);
1874 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1877 LLVMValueRef
ac_build_cvt_pknorm_u16(struct ac_llvm_context
*ctx
,
1878 LLVMValueRef args
[2])
1881 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pknorm.u16",
1882 ctx
->v2i16
, args
, 2,
1883 AC_FUNC_ATTR_READNONE
);
1884 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1887 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1888 LLVMValueRef
ac_build_cvt_pk_i16(struct ac_llvm_context
*ctx
,
1889 LLVMValueRef args
[2], unsigned bits
, bool hi
)
1891 assert(bits
== 8 || bits
== 10 || bits
== 16);
1893 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
1894 bits
== 8 ? 127 : bits
== 10 ? 511 : 32767, 0);
1895 LLVMValueRef min_rgb
= LLVMConstInt(ctx
->i32
,
1896 bits
== 8 ? -128 : bits
== 10 ? -512 : -32768, 0);
1897 LLVMValueRef max_alpha
=
1898 bits
!= 10 ? max_rgb
: ctx
->i32_1
;
1899 LLVMValueRef min_alpha
=
1900 bits
!= 10 ? min_rgb
: LLVMConstInt(ctx
->i32
, -2, 0);
1904 for (int i
= 0; i
< 2; i
++) {
1905 bool alpha
= hi
&& i
== 1;
1906 args
[i
] = ac_build_imin(ctx
, args
[i
],
1907 alpha
? max_alpha
: max_rgb
);
1908 args
[i
] = ac_build_imax(ctx
, args
[i
],
1909 alpha
? min_alpha
: min_rgb
);
1914 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.i16",
1915 ctx
->v2i16
, args
, 2,
1916 AC_FUNC_ATTR_READNONE
);
1917 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1920 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1921 LLVMValueRef
ac_build_cvt_pk_u16(struct ac_llvm_context
*ctx
,
1922 LLVMValueRef args
[2], unsigned bits
, bool hi
)
1924 assert(bits
== 8 || bits
== 10 || bits
== 16);
1926 LLVMValueRef max_rgb
= LLVMConstInt(ctx
->i32
,
1927 bits
== 8 ? 255 : bits
== 10 ? 1023 : 65535, 0);
1928 LLVMValueRef max_alpha
=
1929 bits
!= 10 ? max_rgb
: LLVMConstInt(ctx
->i32
, 3, 0);
1933 for (int i
= 0; i
< 2; i
++) {
1934 bool alpha
= hi
&& i
== 1;
1935 args
[i
] = ac_build_umin(ctx
, args
[i
],
1936 alpha
? max_alpha
: max_rgb
);
1941 ac_build_intrinsic(ctx
, "llvm.amdgcn.cvt.pk.u16",
1942 ctx
->v2i16
, args
, 2,
1943 AC_FUNC_ATTR_READNONE
);
1944 return LLVMBuildBitCast(ctx
->builder
, res
, ctx
->i32
, "");
1947 LLVMValueRef
ac_build_wqm_vote(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
1949 return ac_build_intrinsic(ctx
, "llvm.amdgcn.wqm.vote", ctx
->i1
,
1950 &i1
, 1, AC_FUNC_ATTR_READNONE
);
1953 void ac_build_kill_if_false(struct ac_llvm_context
*ctx
, LLVMValueRef i1
)
1955 ac_build_intrinsic(ctx
, "llvm.amdgcn.kill", ctx
->voidt
,
1959 LLVMValueRef
ac_build_bfe(struct ac_llvm_context
*ctx
, LLVMValueRef input
,
1960 LLVMValueRef offset
, LLVMValueRef width
,
1963 LLVMValueRef args
[] = {
1969 return ac_build_intrinsic(ctx
,
1970 is_signed
? "llvm.amdgcn.sbfe.i32" :
1971 "llvm.amdgcn.ubfe.i32",
1973 AC_FUNC_ATTR_READNONE
);
1976 LLVMValueRef
ac_build_imad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
1977 LLVMValueRef s1
, LLVMValueRef s2
)
1979 return LLVMBuildAdd(ctx
->builder
,
1980 LLVMBuildMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
1983 LLVMValueRef
ac_build_fmad(struct ac_llvm_context
*ctx
, LLVMValueRef s0
,
1984 LLVMValueRef s1
, LLVMValueRef s2
)
1986 return LLVMBuildFAdd(ctx
->builder
,
1987 LLVMBuildFMul(ctx
->builder
, s0
, s1
, ""), s2
, "");
1990 void ac_build_waitcnt(struct ac_llvm_context
*ctx
, unsigned simm16
)
1992 LLVMValueRef args
[1] = {
1993 LLVMConstInt(ctx
->i32
, simm16
, false),
1995 ac_build_intrinsic(ctx
, "llvm.amdgcn.s.waitcnt",
1996 ctx
->voidt
, args
, 1, 0);
1999 LLVMValueRef
ac_build_fract(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2005 if (bitsize
== 32) {
2006 intr
= "llvm.floor.f32";
2009 intr
= "llvm.floor.f64";
2013 LLVMValueRef params
[] = {
2016 LLVMValueRef floor
= ac_build_intrinsic(ctx
, intr
, type
, params
, 1,
2017 AC_FUNC_ATTR_READNONE
);
2018 return LLVMBuildFSub(ctx
->builder
, src0
, floor
, "");
2021 LLVMValueRef
ac_build_isign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2024 LLVMValueRef cmp
, val
, zero
, one
;
2044 unreachable(!"invalid bitsize");
2048 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, src0
, zero
, "");
2049 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2050 cmp
= LLVMBuildICmp(ctx
->builder
, LLVMIntSGE
, val
, zero
, "");
2051 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstInt(type
, -1, true), "");
2055 LLVMValueRef
ac_build_fsign(struct ac_llvm_context
*ctx
, LLVMValueRef src0
,
2058 LLVMValueRef cmp
, val
, zero
, one
;
2061 if (bitsize
== 32) {
2071 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGT
, src0
, zero
, "");
2072 val
= LLVMBuildSelect(ctx
->builder
, cmp
, one
, src0
, "");
2073 cmp
= LLVMBuildFCmp(ctx
->builder
, LLVMRealOGE
, val
, zero
, "");
2074 val
= LLVMBuildSelect(ctx
->builder
, cmp
, val
, LLVMConstReal(type
, -1.0), "");
2078 LLVMValueRef
ac_build_bit_count(struct ac_llvm_context
*ctx
, LLVMValueRef src0
)
2080 LLVMValueRef result
;
2083 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2087 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i64", ctx
->i64
,
2088 (LLVMValueRef
[]) { src0
}, 1,
2089 AC_FUNC_ATTR_READNONE
);
2091 result
= LLVMBuildTrunc(ctx
->builder
, result
, ctx
->i32
, "");
2094 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i32", ctx
->i32
,
2095 (LLVMValueRef
[]) { src0
}, 1,
2096 AC_FUNC_ATTR_READNONE
);
2099 result
= ac_build_intrinsic(ctx
, "llvm.ctpop.i16", ctx
->i16
,
2100 (LLVMValueRef
[]) { src0
}, 1,
2101 AC_FUNC_ATTR_READNONE
);
2104 unreachable(!"invalid bitsize");
2111 LLVMValueRef
ac_build_bitfield_reverse(struct ac_llvm_context
*ctx
,
2114 LLVMValueRef result
;
2117 bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2121 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i32", ctx
->i32
,
2122 (LLVMValueRef
[]) { src0
}, 1,
2123 AC_FUNC_ATTR_READNONE
);
2126 result
= ac_build_intrinsic(ctx
, "llvm.bitreverse.i16", ctx
->i16
,
2127 (LLVMValueRef
[]) { src0
}, 1,
2128 AC_FUNC_ATTR_READNONE
);
2131 unreachable(!"invalid bitsize");
2138 #define AC_EXP_TARGET 0
2139 #define AC_EXP_ENABLED_CHANNELS 1
2140 #define AC_EXP_OUT0 2
2148 struct ac_vs_exp_chan
2152 enum ac_ir_type type
;
2155 struct ac_vs_exp_inst
{
2158 struct ac_vs_exp_chan chan
[4];
2161 struct ac_vs_exports
{
2163 struct ac_vs_exp_inst exp
[VARYING_SLOT_MAX
];
2166 /* Return true if the PARAM export has been eliminated. */
2167 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset
,
2168 uint32_t num_outputs
,
2169 struct ac_vs_exp_inst
*exp
)
2171 unsigned i
, default_val
; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2172 bool is_zero
[4] = {}, is_one
[4] = {};
2174 for (i
= 0; i
< 4; i
++) {
2175 /* It's a constant expression. Undef outputs are eliminated too. */
2176 if (exp
->chan
[i
].type
== AC_IR_UNDEF
) {
2179 } else if (exp
->chan
[i
].type
== AC_IR_CONST
) {
2180 if (exp
->chan
[i
].const_float
== 0)
2182 else if (exp
->chan
[i
].const_float
== 1)
2185 return false; /* other constant */
2190 /* Only certain combinations of 0 and 1 can be eliminated. */
2191 if (is_zero
[0] && is_zero
[1] && is_zero
[2])
2192 default_val
= is_zero
[3] ? 0 : 1;
2193 else if (is_one
[0] && is_one
[1] && is_one
[2])
2194 default_val
= is_zero
[3] ? 2 : 3;
2198 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2199 LLVMInstructionEraseFromParent(exp
->inst
);
2201 /* Change OFFSET to DEFAULT_VAL. */
2202 for (i
= 0; i
< num_outputs
; i
++) {
2203 if (vs_output_param_offset
[i
] == exp
->offset
) {
2204 vs_output_param_offset
[i
] =
2205 AC_EXP_PARAM_DEFAULT_VAL_0000
+ default_val
;
2212 static bool ac_eliminate_duplicated_output(struct ac_llvm_context
*ctx
,
2213 uint8_t *vs_output_param_offset
,
2214 uint32_t num_outputs
,
2215 struct ac_vs_exports
*processed
,
2216 struct ac_vs_exp_inst
*exp
)
2218 unsigned p
, copy_back_channels
= 0;
2220 /* See if the output is already in the list of processed outputs.
2221 * The LLVMValueRef comparison relies on SSA.
2223 for (p
= 0; p
< processed
->num
; p
++) {
2224 bool different
= false;
2226 for (unsigned j
= 0; j
< 4; j
++) {
2227 struct ac_vs_exp_chan
*c1
= &processed
->exp
[p
].chan
[j
];
2228 struct ac_vs_exp_chan
*c2
= &exp
->chan
[j
];
2230 /* Treat undef as a match. */
2231 if (c2
->type
== AC_IR_UNDEF
)
2234 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2235 * and consider the instruction duplicated.
2237 if (c1
->type
== AC_IR_UNDEF
) {
2238 copy_back_channels
|= 1 << j
;
2242 /* Test whether the channels are not equal. */
2243 if (c1
->type
!= c2
->type
||
2244 (c1
->type
== AC_IR_CONST
&&
2245 c1
->const_float
!= c2
->const_float
) ||
2246 (c1
->type
== AC_IR_VALUE
&&
2247 c1
->value
!= c2
->value
)) {
2255 copy_back_channels
= 0;
2257 if (p
== processed
->num
)
2260 /* If a match was found, but the matching export has undef where the new
2261 * one has a normal value, copy the normal value to the undef channel.
2263 struct ac_vs_exp_inst
*match
= &processed
->exp
[p
];
2265 /* Get current enabled channels mask. */
2266 LLVMValueRef arg
= LLVMGetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
);
2267 unsigned enabled_channels
= LLVMConstIntGetZExtValue(arg
);
2269 while (copy_back_channels
) {
2270 unsigned chan
= u_bit_scan(©_back_channels
);
2272 assert(match
->chan
[chan
].type
== AC_IR_UNDEF
);
2273 LLVMSetOperand(match
->inst
, AC_EXP_OUT0
+ chan
,
2274 exp
->chan
[chan
].value
);
2275 match
->chan
[chan
] = exp
->chan
[chan
];
2277 /* Update number of enabled channels because the original mask
2278 * is not always 0xf.
2280 enabled_channels
|= (1 << chan
);
2281 LLVMSetOperand(match
->inst
, AC_EXP_ENABLED_CHANNELS
,
2282 LLVMConstInt(ctx
->i32
, enabled_channels
, 0));
2285 /* The PARAM export is duplicated. Kill it. */
2286 LLVMInstructionEraseFromParent(exp
->inst
);
2288 /* Change OFFSET to the matching export. */
2289 for (unsigned i
= 0; i
< num_outputs
; i
++) {
2290 if (vs_output_param_offset
[i
] == exp
->offset
) {
2291 vs_output_param_offset
[i
] = match
->offset
;
2298 void ac_optimize_vs_outputs(struct ac_llvm_context
*ctx
,
2299 LLVMValueRef main_fn
,
2300 uint8_t *vs_output_param_offset
,
2301 uint32_t num_outputs
,
2302 uint8_t *num_param_exports
)
2304 LLVMBasicBlockRef bb
;
2305 bool removed_any
= false;
2306 struct ac_vs_exports exports
;
2310 /* Process all LLVM instructions. */
2311 bb
= LLVMGetFirstBasicBlock(main_fn
);
2313 LLVMValueRef inst
= LLVMGetFirstInstruction(bb
);
2316 LLVMValueRef cur
= inst
;
2317 inst
= LLVMGetNextInstruction(inst
);
2318 struct ac_vs_exp_inst exp
;
2320 if (LLVMGetInstructionOpcode(cur
) != LLVMCall
)
2323 LLVMValueRef callee
= ac_llvm_get_called_value(cur
);
2325 if (!ac_llvm_is_function(callee
))
2328 const char *name
= LLVMGetValueName(callee
);
2329 unsigned num_args
= LLVMCountParams(callee
);
2331 /* Check if this is an export instruction. */
2332 if ((num_args
!= 9 && num_args
!= 8) ||
2333 (strcmp(name
, "llvm.SI.export") &&
2334 strcmp(name
, "llvm.amdgcn.exp.f32")))
2337 LLVMValueRef arg
= LLVMGetOperand(cur
, AC_EXP_TARGET
);
2338 unsigned target
= LLVMConstIntGetZExtValue(arg
);
2340 if (target
< V_008DFC_SQ_EXP_PARAM
)
2343 target
-= V_008DFC_SQ_EXP_PARAM
;
2345 /* Parse the instruction. */
2346 memset(&exp
, 0, sizeof(exp
));
2347 exp
.offset
= target
;
2350 for (unsigned i
= 0; i
< 4; i
++) {
2351 LLVMValueRef v
= LLVMGetOperand(cur
, AC_EXP_OUT0
+ i
);
2353 exp
.chan
[i
].value
= v
;
2355 if (LLVMIsUndef(v
)) {
2356 exp
.chan
[i
].type
= AC_IR_UNDEF
;
2357 } else if (LLVMIsAConstantFP(v
)) {
2358 LLVMBool loses_info
;
2359 exp
.chan
[i
].type
= AC_IR_CONST
;
2360 exp
.chan
[i
].const_float
=
2361 LLVMConstRealGetDouble(v
, &loses_info
);
2363 exp
.chan
[i
].type
= AC_IR_VALUE
;
2367 /* Eliminate constant and duplicated PARAM exports. */
2368 if (ac_eliminate_const_output(vs_output_param_offset
,
2369 num_outputs
, &exp
) ||
2370 ac_eliminate_duplicated_output(ctx
,
2371 vs_output_param_offset
,
2372 num_outputs
, &exports
,
2376 exports
.exp
[exports
.num
++] = exp
;
2379 bb
= LLVMGetNextBasicBlock(bb
);
2382 /* Remove holes in export memory due to removed PARAM exports.
2383 * This is done by renumbering all PARAM exports.
2386 uint8_t old_offset
[VARYING_SLOT_MAX
];
2389 /* Make a copy of the offsets. We need the old version while
2390 * we are modifying some of them. */
2391 memcpy(old_offset
, vs_output_param_offset
,
2392 sizeof(old_offset
));
2394 for (i
= 0; i
< exports
.num
; i
++) {
2395 unsigned offset
= exports
.exp
[i
].offset
;
2397 /* Update vs_output_param_offset. Multiple outputs can
2398 * have the same offset.
2400 for (out
= 0; out
< num_outputs
; out
++) {
2401 if (old_offset
[out
] == offset
)
2402 vs_output_param_offset
[out
] = i
;
2405 /* Change the PARAM offset in the instruction. */
2406 LLVMSetOperand(exports
.exp
[i
].inst
, AC_EXP_TARGET
,
2407 LLVMConstInt(ctx
->i32
,
2408 V_008DFC_SQ_EXP_PARAM
+ i
, 0));
2410 *num_param_exports
= exports
.num
;
2414 void ac_init_exec_full_mask(struct ac_llvm_context
*ctx
)
2416 LLVMValueRef full_mask
= LLVMConstInt(ctx
->i64
, ~0ull, 0);
2417 ac_build_intrinsic(ctx
,
2418 "llvm.amdgcn.init.exec", ctx
->voidt
,
2419 &full_mask
, 1, AC_FUNC_ATTR_CONVERGENT
);
2422 void ac_declare_lds_as_pointer(struct ac_llvm_context
*ctx
)
2424 unsigned lds_size
= ctx
->chip_class
>= CIK
? 65536 : 32768;
2425 ctx
->lds
= LLVMBuildIntToPtr(ctx
->builder
, ctx
->i32_0
,
2426 LLVMPointerType(LLVMArrayType(ctx
->i32
, lds_size
/ 4), AC_ADDR_SPACE_LDS
),
2430 LLVMValueRef
ac_lds_load(struct ac_llvm_context
*ctx
,
2431 LLVMValueRef dw_addr
)
2433 return ac_build_load(ctx
, ctx
->lds
, dw_addr
);
2436 void ac_lds_store(struct ac_llvm_context
*ctx
,
2437 LLVMValueRef dw_addr
,
2440 value
= ac_to_integer(ctx
, value
);
2441 ac_build_indexed_store(ctx
, ctx
->lds
,
2445 LLVMValueRef
ac_find_lsb(struct ac_llvm_context
*ctx
,
2446 LLVMTypeRef dst_type
,
2449 unsigned src0_bitsize
= ac_get_elem_bits(ctx
, LLVMTypeOf(src0
));
2450 const char *intrin_name
;
2454 switch (src0_bitsize
) {
2456 intrin_name
= "llvm.cttz.i64";
2461 intrin_name
= "llvm.cttz.i32";
2466 intrin_name
= "llvm.cttz.i16";
2471 unreachable(!"invalid bitsize");
2474 LLVMValueRef params
[2] = {
2477 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
2478 * add special code to check for x=0. The reason is that
2479 * the LLVM behavior for x=0 is different from what we
2480 * need here. However, LLVM also assumes that ffs(x) is
2481 * in [0, 31], but GLSL expects that ffs(0) = -1, so
2482 * a conditional assignment to handle 0 is still required.
2484 * The hardware already implements the correct behavior.
2489 LLVMValueRef lsb
= ac_build_intrinsic(ctx
, intrin_name
, type
,
2491 AC_FUNC_ATTR_READNONE
);
2493 if (src0_bitsize
== 64) {
2494 lsb
= LLVMBuildTrunc(ctx
->builder
, lsb
, ctx
->i32
, "");
2497 /* TODO: We need an intrinsic to skip this conditional. */
2498 /* Check for zero: */
2499 return LLVMBuildSelect(ctx
->builder
, LLVMBuildICmp(ctx
->builder
,
2502 LLVMConstInt(ctx
->i32
, -1, 0), lsb
, "");
2505 LLVMTypeRef
ac_array_in_const_addr_space(LLVMTypeRef elem_type
)
2507 return LLVMPointerType(LLVMArrayType(elem_type
, 0),
2508 AC_ADDR_SPACE_CONST
);
2511 LLVMTypeRef
ac_array_in_const32_addr_space(LLVMTypeRef elem_type
)
2513 return LLVMPointerType(LLVMArrayType(elem_type
, 0),
2514 AC_ADDR_SPACE_CONST_32BIT
);
2517 static struct ac_llvm_flow
*
2518 get_current_flow(struct ac_llvm_context
*ctx
)
2520 if (ctx
->flow_depth
> 0)
2521 return &ctx
->flow
[ctx
->flow_depth
- 1];
2525 static struct ac_llvm_flow
*
2526 get_innermost_loop(struct ac_llvm_context
*ctx
)
2528 for (unsigned i
= ctx
->flow_depth
; i
> 0; --i
) {
2529 if (ctx
->flow
[i
- 1].loop_entry_block
)
2530 return &ctx
->flow
[i
- 1];
2535 static struct ac_llvm_flow
*
2536 push_flow(struct ac_llvm_context
*ctx
)
2538 struct ac_llvm_flow
*flow
;
2540 if (ctx
->flow_depth
>= ctx
->flow_depth_max
) {
2541 unsigned new_max
= MAX2(ctx
->flow_depth
<< 1,
2542 AC_LLVM_INITIAL_CF_DEPTH
);
2544 ctx
->flow
= realloc(ctx
->flow
, new_max
* sizeof(*ctx
->flow
));
2545 ctx
->flow_depth_max
= new_max
;
2548 flow
= &ctx
->flow
[ctx
->flow_depth
];
2551 flow
->next_block
= NULL
;
2552 flow
->loop_entry_block
= NULL
;
2556 static void set_basicblock_name(LLVMBasicBlockRef bb
, const char *base
,
2560 snprintf(buf
, sizeof(buf
), "%s%d", base
, label_id
);
2561 LLVMSetValueName(LLVMBasicBlockAsValue(bb
), buf
);
2564 /* Append a basic block at the level of the parent flow.
2566 static LLVMBasicBlockRef
append_basic_block(struct ac_llvm_context
*ctx
,
2569 assert(ctx
->flow_depth
>= 1);
2571 if (ctx
->flow_depth
>= 2) {
2572 struct ac_llvm_flow
*flow
= &ctx
->flow
[ctx
->flow_depth
- 2];
2574 return LLVMInsertBasicBlockInContext(ctx
->context
,
2575 flow
->next_block
, name
);
2578 LLVMValueRef main_fn
=
2579 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx
->builder
));
2580 return LLVMAppendBasicBlockInContext(ctx
->context
, main_fn
, name
);
2583 /* Emit a branch to the given default target for the current block if
2584 * applicable -- that is, if the current block does not already contain a
2585 * branch from a break or continue.
2587 static void emit_default_branch(LLVMBuilderRef builder
,
2588 LLVMBasicBlockRef target
)
2590 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder
)))
2591 LLVMBuildBr(builder
, target
);
2594 void ac_build_bgnloop(struct ac_llvm_context
*ctx
, int label_id
)
2596 struct ac_llvm_flow
*flow
= push_flow(ctx
);
2597 flow
->loop_entry_block
= append_basic_block(ctx
, "LOOP");
2598 flow
->next_block
= append_basic_block(ctx
, "ENDLOOP");
2599 set_basicblock_name(flow
->loop_entry_block
, "loop", label_id
);
2600 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
2601 LLVMPositionBuilderAtEnd(ctx
->builder
, flow
->loop_entry_block
);
2604 void ac_build_break(struct ac_llvm_context
*ctx
)
2606 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
2607 LLVMBuildBr(ctx
->builder
, flow
->next_block
);
2610 void ac_build_continue(struct ac_llvm_context
*ctx
)
2612 struct ac_llvm_flow
*flow
= get_innermost_loop(ctx
);
2613 LLVMBuildBr(ctx
->builder
, flow
->loop_entry_block
);
2616 void ac_build_else(struct ac_llvm_context
*ctx
, int label_id
)
2618 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
2619 LLVMBasicBlockRef endif_block
;
2621 assert(!current_branch
->loop_entry_block
);
2623 endif_block
= append_basic_block(ctx
, "ENDIF");
2624 emit_default_branch(ctx
->builder
, endif_block
);
2626 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
2627 set_basicblock_name(current_branch
->next_block
, "else", label_id
);
2629 current_branch
->next_block
= endif_block
;
2632 void ac_build_endif(struct ac_llvm_context
*ctx
, int label_id
)
2634 struct ac_llvm_flow
*current_branch
= get_current_flow(ctx
);
2636 assert(!current_branch
->loop_entry_block
);
2638 emit_default_branch(ctx
->builder
, current_branch
->next_block
);
2639 LLVMPositionBuilderAtEnd(ctx
->builder
, current_branch
->next_block
);
2640 set_basicblock_name(current_branch
->next_block
, "endif", label_id
);
2645 void ac_build_endloop(struct ac_llvm_context
*ctx
, int label_id
)
2647 struct ac_llvm_flow
*current_loop
= get_current_flow(ctx
);
2649 assert(current_loop
->loop_entry_block
);
2651 emit_default_branch(ctx
->builder
, current_loop
->loop_entry_block
);
2653 LLVMPositionBuilderAtEnd(ctx
->builder
, current_loop
->next_block
);
2654 set_basicblock_name(current_loop
->next_block
, "endloop", label_id
);
2658 void ac_build_ifcc(struct ac_llvm_context
*ctx
, LLVMValueRef cond
, int label_id
)
2660 struct ac_llvm_flow
*flow
= push_flow(ctx
);
2661 LLVMBasicBlockRef if_block
;
2663 if_block
= append_basic_block(ctx
, "IF");
2664 flow
->next_block
= append_basic_block(ctx
, "ELSE");
2665 set_basicblock_name(if_block
, "if", label_id
);
2666 LLVMBuildCondBr(ctx
->builder
, cond
, if_block
, flow
->next_block
);
2667 LLVMPositionBuilderAtEnd(ctx
->builder
, if_block
);
2670 void ac_build_if(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
2673 LLVMValueRef cond
= LLVMBuildFCmp(ctx
->builder
, LLVMRealUNE
,
2674 value
, ctx
->f32_0
, "");
2675 ac_build_ifcc(ctx
, cond
, label_id
);
2678 void ac_build_uif(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
2681 LLVMValueRef cond
= LLVMBuildICmp(ctx
->builder
, LLVMIntNE
,
2682 ac_to_integer(ctx
, value
),
2684 ac_build_ifcc(ctx
, cond
, label_id
);
2687 LLVMValueRef
ac_build_alloca_undef(struct ac_llvm_context
*ac
, LLVMTypeRef type
,
2690 LLVMBuilderRef builder
= ac
->builder
;
2691 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(builder
);
2692 LLVMValueRef function
= LLVMGetBasicBlockParent(current_block
);
2693 LLVMBasicBlockRef first_block
= LLVMGetEntryBasicBlock(function
);
2694 LLVMValueRef first_instr
= LLVMGetFirstInstruction(first_block
);
2695 LLVMBuilderRef first_builder
= LLVMCreateBuilderInContext(ac
->context
);
2699 LLVMPositionBuilderBefore(first_builder
, first_instr
);
2701 LLVMPositionBuilderAtEnd(first_builder
, first_block
);
2704 res
= LLVMBuildAlloca(first_builder
, type
, name
);
2705 LLVMDisposeBuilder(first_builder
);
2709 LLVMValueRef
ac_build_alloca(struct ac_llvm_context
*ac
,
2710 LLVMTypeRef type
, const char *name
)
2712 LLVMValueRef ptr
= ac_build_alloca_undef(ac
, type
, name
);
2713 LLVMBuildStore(ac
->builder
, LLVMConstNull(type
), ptr
);
2717 LLVMValueRef
ac_cast_ptr(struct ac_llvm_context
*ctx
, LLVMValueRef ptr
,
2720 int addr_space
= LLVMGetPointerAddressSpace(LLVMTypeOf(ptr
));
2721 return LLVMBuildBitCast(ctx
->builder
, ptr
,
2722 LLVMPointerType(type
, addr_space
), "");
2725 LLVMValueRef
ac_trim_vector(struct ac_llvm_context
*ctx
, LLVMValueRef value
,
2728 unsigned num_components
= ac_get_llvm_num_components(value
);
2729 if (count
== num_components
)
2732 LLVMValueRef masks
[MAX2(count
, 2)];
2733 masks
[0] = ctx
->i32_0
;
2734 masks
[1] = ctx
->i32_1
;
2735 for (unsigned i
= 2; i
< count
; i
++)
2736 masks
[i
] = LLVMConstInt(ctx
->i32
, i
, false);
2739 return LLVMBuildExtractElement(ctx
->builder
, value
, masks
[0],
2742 LLVMValueRef swizzle
= LLVMConstVector(masks
, count
);
2743 return LLVMBuildShuffleVector(ctx
->builder
, value
, value
, swizzle
, "");
2746 LLVMValueRef
ac_unpack_param(struct ac_llvm_context
*ctx
, LLVMValueRef param
,
2747 unsigned rshift
, unsigned bitwidth
)
2749 LLVMValueRef value
= param
;
2751 value
= LLVMBuildLShr(ctx
->builder
, value
,
2752 LLVMConstInt(ctx
->i32
, rshift
, false), "");
2754 if (rshift
+ bitwidth
< 32) {
2755 unsigned mask
= (1 << bitwidth
) - 1;
2756 value
= LLVMBuildAnd(ctx
->builder
, value
,
2757 LLVMConstInt(ctx
->i32
, mask
, false), "");
2762 /* Adjust the sample index according to FMASK.
2764 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2765 * which is the identity mapping. Each nibble says which physical sample
2766 * should be fetched to get that sample.
2768 * For example, 0x11111100 means there are only 2 samples stored and
2769 * the second sample covers 3/4 of the pixel. When reading samples 0
2770 * and 1, return physical sample 0 (determined by the first two 0s
2771 * in FMASK), otherwise return physical sample 1.
2773 * The sample index should be adjusted as follows:
2774 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
2776 void ac_apply_fmask_to_sample(struct ac_llvm_context
*ac
, LLVMValueRef fmask
,
2777 LLVMValueRef
*addr
, bool is_array_tex
)
2779 struct ac_image_args fmask_load
= {};
2780 fmask_load
.opcode
= ac_image_load
;
2781 fmask_load
.resource
= fmask
;
2782 fmask_load
.dmask
= 0xf;
2783 fmask_load
.dim
= is_array_tex
? ac_image_2darray
: ac_image_2d
;
2785 fmask_load
.coords
[0] = addr
[0];
2786 fmask_load
.coords
[1] = addr
[1];
2788 fmask_load
.coords
[2] = addr
[2];
2790 LLVMValueRef fmask_value
= ac_build_image_opcode(ac
, &fmask_load
);
2791 fmask_value
= LLVMBuildExtractElement(ac
->builder
, fmask_value
,
2794 /* Apply the formula. */
2795 unsigned sample_chan
= is_array_tex
? 3 : 2;
2796 LLVMValueRef final_sample
;
2797 final_sample
= LLVMBuildMul(ac
->builder
, addr
[sample_chan
],
2798 LLVMConstInt(ac
->i32
, 4, 0), "");
2799 final_sample
= LLVMBuildLShr(ac
->builder
, fmask_value
, final_sample
, "");
2800 /* Mask the sample index by 0x7, because 0x8 means an unknown value
2801 * with EQAA, so those will map to 0. */
2802 final_sample
= LLVMBuildAnd(ac
->builder
, final_sample
,
2803 LLVMConstInt(ac
->i32
, 0x7, 0), "");
2805 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2806 * resource descriptor is 0 (invalid).
2809 tmp
= LLVMBuildBitCast(ac
->builder
, fmask
, ac
->v8i32
, "");
2810 tmp
= LLVMBuildExtractElement(ac
->builder
, tmp
, ac
->i32_1
, "");
2811 tmp
= LLVMBuildICmp(ac
->builder
, LLVMIntNE
, tmp
, ac
->i32_0
, "");
2813 /* Replace the MSAA sample index. */
2814 addr
[sample_chan
] = LLVMBuildSelect(ac
->builder
, tmp
, final_sample
,
2815 addr
[sample_chan
], "");
2819 _ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
2821 ac_build_optimization_barrier(ctx
, &src
);
2822 return ac_build_intrinsic(ctx
,
2823 lane
== NULL
? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
2824 LLVMTypeOf(src
), (LLVMValueRef
[]) {
2826 lane
== NULL
? 1 : 2,
2827 AC_FUNC_ATTR_READNONE
|
2828 AC_FUNC_ATTR_CONVERGENT
);
2832 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
2835 * @param lane - id of the lane or NULL for the first active lane
2836 * @return value of the lane
2839 ac_build_readlane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef lane
)
2841 LLVMTypeRef src_type
= LLVMTypeOf(src
);
2842 src
= ac_to_integer(ctx
, src
);
2843 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
2847 ret
= _ac_build_readlane(ctx
, src
, lane
);
2849 assert(bits
% 32 == 0);
2850 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
2851 LLVMValueRef src_vector
=
2852 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
2853 ret
= LLVMGetUndef(vec_type
);
2854 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
2855 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
2856 LLVMConstInt(ctx
->i32
, i
, 0), "");
2857 LLVMValueRef ret_comp
= _ac_build_readlane(ctx
, src
, lane
);
2858 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
, ret_comp
,
2859 LLVMConstInt(ctx
->i32
, i
, 0), "");
2862 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
2866 ac_build_writelane(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef value
, LLVMValueRef lane
)
2868 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
2870 LLVMValueRef pred
= LLVMBuildICmp(ctx
->builder
, LLVMIntEQ
, lane
,
2871 ac_get_thread_id(ctx
), "");
2872 return LLVMBuildSelect(ctx
->builder
, pred
, value
, src
, "");
2876 ac_build_mbcnt(struct ac_llvm_context
*ctx
, LLVMValueRef mask
)
2878 LLVMValueRef mask_vec
= LLVMBuildBitCast(ctx
->builder
, mask
,
2879 LLVMVectorType(ctx
->i32
, 2),
2881 LLVMValueRef mask_lo
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
2883 LLVMValueRef mask_hi
= LLVMBuildExtractElement(ctx
->builder
, mask_vec
,
2886 ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.lo", ctx
->i32
,
2887 (LLVMValueRef
[]) { mask_lo
, ctx
->i32_0
},
2888 2, AC_FUNC_ATTR_READNONE
);
2889 val
= ac_build_intrinsic(ctx
, "llvm.amdgcn.mbcnt.hi", ctx
->i32
,
2890 (LLVMValueRef
[]) { mask_hi
, val
},
2891 2, AC_FUNC_ATTR_READNONE
);
2896 _dpp_quad_perm
= 0x000,
2897 _dpp_row_sl
= 0x100,
2898 _dpp_row_sr
= 0x110,
2899 _dpp_row_rr
= 0x120,
2904 dpp_row_mirror
= 0x140,
2905 dpp_row_half_mirror
= 0x141,
2906 dpp_row_bcast15
= 0x142,
2907 dpp_row_bcast31
= 0x143
2910 static inline enum dpp_ctrl
2911 dpp_quad_perm(unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
2913 assert(lane0
< 4 && lane1
< 4 && lane2
< 4 && lane3
< 4);
2914 return _dpp_quad_perm
| lane0
| (lane1
<< 2) | (lane2
<< 4) | (lane3
<< 6);
2917 static inline enum dpp_ctrl
2918 dpp_row_sl(unsigned amount
)
2920 assert(amount
> 0 && amount
< 16);
2921 return _dpp_row_sl
| amount
;
2924 static inline enum dpp_ctrl
2925 dpp_row_sr(unsigned amount
)
2927 assert(amount
> 0 && amount
< 16);
2928 return _dpp_row_sr
| amount
;
2932 _ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
2933 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
2936 return ac_build_intrinsic(ctx
, "llvm.amdgcn.update.dpp.i32",
2940 LLVMConstInt(ctx
->i32
, dpp_ctrl
, 0),
2941 LLVMConstInt(ctx
->i32
, row_mask
, 0),
2942 LLVMConstInt(ctx
->i32
, bank_mask
, 0),
2943 LLVMConstInt(ctx
->i1
, bound_ctrl
, 0) },
2944 6, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
2948 ac_build_dpp(struct ac_llvm_context
*ctx
, LLVMValueRef old
, LLVMValueRef src
,
2949 enum dpp_ctrl dpp_ctrl
, unsigned row_mask
, unsigned bank_mask
,
2952 LLVMTypeRef src_type
= LLVMTypeOf(src
);
2953 src
= ac_to_integer(ctx
, src
);
2954 old
= ac_to_integer(ctx
, old
);
2955 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
2958 ret
= _ac_build_dpp(ctx
, old
, src
, dpp_ctrl
, row_mask
,
2959 bank_mask
, bound_ctrl
);
2961 assert(bits
% 32 == 0);
2962 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
2963 LLVMValueRef src_vector
=
2964 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
2965 LLVMValueRef old_vector
=
2966 LLVMBuildBitCast(ctx
->builder
, old
, vec_type
, "");
2967 ret
= LLVMGetUndef(vec_type
);
2968 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
2969 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
2970 LLVMConstInt(ctx
->i32
, i
,
2972 old
= LLVMBuildExtractElement(ctx
->builder
, old_vector
,
2973 LLVMConstInt(ctx
->i32
, i
,
2975 LLVMValueRef ret_comp
= _ac_build_dpp(ctx
, old
, src
,
2980 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
2982 LLVMConstInt(ctx
->i32
, i
,
2986 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
2989 static inline unsigned
2990 ds_pattern_bitmode(unsigned and_mask
, unsigned or_mask
, unsigned xor_mask
)
2992 assert(and_mask
< 32 && or_mask
< 32 && xor_mask
< 32);
2993 return and_mask
| (or_mask
<< 5) | (xor_mask
<< 10);
2997 _ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
2999 return ac_build_intrinsic(ctx
, "llvm.amdgcn.ds.swizzle",
3000 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3001 src
, LLVMConstInt(ctx
->i32
, mask
, 0) },
3002 2, AC_FUNC_ATTR_READNONE
| AC_FUNC_ATTR_CONVERGENT
);
3006 ac_build_ds_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, unsigned mask
)
3008 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3009 src
= ac_to_integer(ctx
, src
);
3010 unsigned bits
= LLVMGetIntTypeWidth(LLVMTypeOf(src
));
3013 ret
= _ac_build_ds_swizzle(ctx
, src
, mask
);
3015 assert(bits
% 32 == 0);
3016 LLVMTypeRef vec_type
= LLVMVectorType(ctx
->i32
, bits
/ 32);
3017 LLVMValueRef src_vector
=
3018 LLVMBuildBitCast(ctx
->builder
, src
, vec_type
, "");
3019 ret
= LLVMGetUndef(vec_type
);
3020 for (unsigned i
= 0; i
< bits
/ 32; i
++) {
3021 src
= LLVMBuildExtractElement(ctx
->builder
, src_vector
,
3022 LLVMConstInt(ctx
->i32
, i
,
3024 LLVMValueRef ret_comp
= _ac_build_ds_swizzle(ctx
, src
,
3026 ret
= LLVMBuildInsertElement(ctx
->builder
, ret
,
3028 LLVMConstInt(ctx
->i32
, i
,
3032 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3036 ac_build_wwm(struct ac_llvm_context
*ctx
, LLVMValueRef src
)
3038 char name
[32], type
[8];
3039 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3040 snprintf(name
, sizeof(name
), "llvm.amdgcn.wwm.%s", type
);
3041 return ac_build_intrinsic(ctx
, name
, LLVMTypeOf(src
),
3042 (LLVMValueRef
[]) { src
}, 1,
3043 AC_FUNC_ATTR_READNONE
);
3047 ac_build_set_inactive(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3048 LLVMValueRef inactive
)
3050 char name
[33], type
[8];
3051 LLVMTypeRef src_type
= LLVMTypeOf(src
);
3052 src
= ac_to_integer(ctx
, src
);
3053 inactive
= ac_to_integer(ctx
, inactive
);
3054 ac_build_type_name_for_intr(LLVMTypeOf(src
), type
, sizeof(type
));
3055 snprintf(name
, sizeof(name
), "llvm.amdgcn.set.inactive.%s", type
);
3057 ac_build_intrinsic(ctx
, name
,
3058 LLVMTypeOf(src
), (LLVMValueRef
[]) {
3060 AC_FUNC_ATTR_READNONE
|
3061 AC_FUNC_ATTR_CONVERGENT
);
3062 return LLVMBuildBitCast(ctx
->builder
, ret
, src_type
, "");
3066 get_reduction_identity(struct ac_llvm_context
*ctx
, nir_op op
, unsigned type_size
)
3068 if (type_size
== 4) {
3070 case nir_op_iadd
: return ctx
->i32_0
;
3071 case nir_op_fadd
: return ctx
->f32_0
;
3072 case nir_op_imul
: return ctx
->i32_1
;
3073 case nir_op_fmul
: return ctx
->f32_1
;
3074 case nir_op_imin
: return LLVMConstInt(ctx
->i32
, INT32_MAX
, 0);
3075 case nir_op_umin
: return LLVMConstInt(ctx
->i32
, UINT32_MAX
, 0);
3076 case nir_op_fmin
: return LLVMConstReal(ctx
->f32
, INFINITY
);
3077 case nir_op_imax
: return LLVMConstInt(ctx
->i32
, INT32_MIN
, 0);
3078 case nir_op_umax
: return ctx
->i32_0
;
3079 case nir_op_fmax
: return LLVMConstReal(ctx
->f32
, -INFINITY
);
3080 case nir_op_iand
: return LLVMConstInt(ctx
->i32
, -1, 0);
3081 case nir_op_ior
: return ctx
->i32_0
;
3082 case nir_op_ixor
: return ctx
->i32_0
;
3084 unreachable("bad reduction intrinsic");
3086 } else { /* type_size == 64bit */
3088 case nir_op_iadd
: return ctx
->i64_0
;
3089 case nir_op_fadd
: return ctx
->f64_0
;
3090 case nir_op_imul
: return ctx
->i64_1
;
3091 case nir_op_fmul
: return ctx
->f64_1
;
3092 case nir_op_imin
: return LLVMConstInt(ctx
->i64
, INT64_MAX
, 0);
3093 case nir_op_umin
: return LLVMConstInt(ctx
->i64
, UINT64_MAX
, 0);
3094 case nir_op_fmin
: return LLVMConstReal(ctx
->f64
, INFINITY
);
3095 case nir_op_imax
: return LLVMConstInt(ctx
->i64
, INT64_MIN
, 0);
3096 case nir_op_umax
: return ctx
->i64_0
;
3097 case nir_op_fmax
: return LLVMConstReal(ctx
->f64
, -INFINITY
);
3098 case nir_op_iand
: return LLVMConstInt(ctx
->i64
, -1, 0);
3099 case nir_op_ior
: return ctx
->i64_0
;
3100 case nir_op_ixor
: return ctx
->i64_0
;
3102 unreachable("bad reduction intrinsic");
3108 ac_build_alu_op(struct ac_llvm_context
*ctx
, LLVMValueRef lhs
, LLVMValueRef rhs
, nir_op op
)
3110 bool _64bit
= ac_get_type_size(LLVMTypeOf(lhs
)) == 8;
3112 case nir_op_iadd
: return LLVMBuildAdd(ctx
->builder
, lhs
, rhs
, "");
3113 case nir_op_fadd
: return LLVMBuildFAdd(ctx
->builder
, lhs
, rhs
, "");
3114 case nir_op_imul
: return LLVMBuildMul(ctx
->builder
, lhs
, rhs
, "");
3115 case nir_op_fmul
: return LLVMBuildFMul(ctx
->builder
, lhs
, rhs
, "");
3116 case nir_op_imin
: return LLVMBuildSelect(ctx
->builder
,
3117 LLVMBuildICmp(ctx
->builder
, LLVMIntSLT
, lhs
, rhs
, ""),
3119 case nir_op_umin
: return LLVMBuildSelect(ctx
->builder
,
3120 LLVMBuildICmp(ctx
->builder
, LLVMIntULT
, lhs
, rhs
, ""),
3122 case nir_op_fmin
: return ac_build_intrinsic(ctx
,
3123 _64bit
? "llvm.minnum.f64" : "llvm.minnum.f32",
3124 _64bit
? ctx
->f64
: ctx
->f32
,
3125 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
3126 case nir_op_imax
: return LLVMBuildSelect(ctx
->builder
,
3127 LLVMBuildICmp(ctx
->builder
, LLVMIntSGT
, lhs
, rhs
, ""),
3129 case nir_op_umax
: return LLVMBuildSelect(ctx
->builder
,
3130 LLVMBuildICmp(ctx
->builder
, LLVMIntUGT
, lhs
, rhs
, ""),
3132 case nir_op_fmax
: return ac_build_intrinsic(ctx
,
3133 _64bit
? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3134 _64bit
? ctx
->f64
: ctx
->f32
,
3135 (LLVMValueRef
[]){lhs
, rhs
}, 2, AC_FUNC_ATTR_READNONE
);
3136 case nir_op_iand
: return LLVMBuildAnd(ctx
->builder
, lhs
, rhs
, "");
3137 case nir_op_ior
: return LLVMBuildOr(ctx
->builder
, lhs
, rhs
, "");
3138 case nir_op_ixor
: return LLVMBuildXor(ctx
->builder
, lhs
, rhs
, "");
3140 unreachable("bad reduction intrinsic");
3145 * \param maxprefix specifies that the result only needs to be correct for a
3146 * prefix of this many threads
3148 * TODO: add inclusive and excluse scan functions for SI chip class.
3151 ac_build_scan(struct ac_llvm_context
*ctx
, nir_op op
, LLVMValueRef src
, LLVMValueRef identity
,
3154 LLVMValueRef result
, tmp
;
3158 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(1), 0xf, 0xf, false);
3159 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3162 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(2), 0xf, 0xf, false);
3163 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3166 tmp
= ac_build_dpp(ctx
, identity
, src
, dpp_row_sr(3), 0xf, 0xf, false);
3167 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3170 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(4), 0xf, 0xe, false);
3171 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3174 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_sr(8), 0xf, 0xc, false);
3175 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3176 if (maxprefix
<= 16)
3178 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
3179 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3180 if (maxprefix
<= 32)
3182 tmp
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
3183 result
= ac_build_alu_op(ctx
, result
, tmp
, op
);
3188 ac_build_inclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
3190 LLVMValueRef result
;
3192 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
3193 LLVMBuilderRef builder
= ctx
->builder
;
3194 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
3195 result
= ac_build_ballot(ctx
, src
);
3196 result
= ac_build_mbcnt(ctx
, result
);
3197 result
= LLVMBuildAdd(builder
, result
, src
, "");
3201 ac_build_optimization_barrier(ctx
, &src
);
3203 LLVMValueRef identity
=
3204 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
3205 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
3206 LLVMTypeOf(identity
), "");
3207 result
= ac_build_scan(ctx
, op
, result
, identity
, 64);
3209 return ac_build_wwm(ctx
, result
);
3213 ac_build_exclusive_scan(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
)
3215 LLVMValueRef result
;
3217 if (LLVMTypeOf(src
) == ctx
->i1
&& op
== nir_op_iadd
) {
3218 LLVMBuilderRef builder
= ctx
->builder
;
3219 src
= LLVMBuildZExt(builder
, src
, ctx
->i32
, "");
3220 result
= ac_build_ballot(ctx
, src
);
3221 result
= ac_build_mbcnt(ctx
, result
);
3225 ac_build_optimization_barrier(ctx
, &src
);
3227 LLVMValueRef identity
=
3228 get_reduction_identity(ctx
, op
, ac_get_type_size(LLVMTypeOf(src
)));
3229 result
= LLVMBuildBitCast(ctx
->builder
, ac_build_set_inactive(ctx
, src
, identity
),
3230 LLVMTypeOf(identity
), "");
3231 result
= ac_build_dpp(ctx
, identity
, result
, dpp_wf_sr1
, 0xf, 0xf, false);
3232 result
= ac_build_scan(ctx
, op
, result
, identity
, 64);
3234 return ac_build_wwm(ctx
, result
);
3238 ac_build_reduce(struct ac_llvm_context
*ctx
, LLVMValueRef src
, nir_op op
, unsigned cluster_size
)
3240 if (cluster_size
== 1) return src
;
3241 ac_build_optimization_barrier(ctx
, &src
);
3242 LLVMValueRef result
, swap
;
3243 LLVMValueRef identity
= get_reduction_identity(ctx
, op
,
3244 ac_get_type_size(LLVMTypeOf(src
)));
3245 result
= LLVMBuildBitCast(ctx
->builder
,
3246 ac_build_set_inactive(ctx
, src
, identity
),
3247 LLVMTypeOf(identity
), "");
3248 swap
= ac_build_quad_swizzle(ctx
, result
, 1, 0, 3, 2);
3249 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3250 if (cluster_size
== 2) return ac_build_wwm(ctx
, result
);
3252 swap
= ac_build_quad_swizzle(ctx
, result
, 2, 3, 0, 1);
3253 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3254 if (cluster_size
== 4) return ac_build_wwm(ctx
, result
);
3256 if (ctx
->chip_class
>= VI
)
3257 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_half_mirror
, 0xf, 0xf, false);
3259 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x04));
3260 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3261 if (cluster_size
== 8) return ac_build_wwm(ctx
, result
);
3263 if (ctx
->chip_class
>= VI
)
3264 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_mirror
, 0xf, 0xf, false);
3266 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x08));
3267 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3268 if (cluster_size
== 16) return ac_build_wwm(ctx
, result
);
3270 if (ctx
->chip_class
>= VI
&& cluster_size
!= 32)
3271 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast15
, 0xa, 0xf, false);
3273 swap
= ac_build_ds_swizzle(ctx
, result
, ds_pattern_bitmode(0x1f, 0, 0x10));
3274 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3275 if (cluster_size
== 32) return ac_build_wwm(ctx
, result
);
3277 if (ctx
->chip_class
>= VI
) {
3278 swap
= ac_build_dpp(ctx
, identity
, result
, dpp_row_bcast31
, 0xc, 0xf, false);
3279 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3280 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 63, 0));
3281 return ac_build_wwm(ctx
, result
);
3283 swap
= ac_build_readlane(ctx
, result
, ctx
->i32_0
);
3284 result
= ac_build_readlane(ctx
, result
, LLVMConstInt(ctx
->i32
, 32, 0));
3285 result
= ac_build_alu_op(ctx
, result
, swap
, op
);
3286 return ac_build_wwm(ctx
, result
);
3291 * "Top half" of a scan that reduces per-wave values across an entire
3294 * The source value must be present in the highest lane of the wave, and the
3295 * highest lane must be live.
3298 ac_build_wg_wavescan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3300 if (ws
->maxwaves
<= 1)
3303 const LLVMValueRef i32_63
= LLVMConstInt(ctx
->i32
, 63, false);
3304 LLVMBuilderRef builder
= ctx
->builder
;
3305 LLVMValueRef tid
= ac_get_thread_id(ctx
);
3308 tmp
= LLVMBuildICmp(builder
, LLVMIntEQ
, tid
, i32_63
, "");
3309 ac_build_ifcc(ctx
, tmp
, 1000);
3310 LLVMBuildStore(builder
, ws
->src
, LLVMBuildGEP(builder
, ws
->scratch
, &ws
->waveidx
, 1, ""));
3311 ac_build_endif(ctx
, 1000);
3315 * "Bottom half" of a scan that reduces per-wave values across an entire
3318 * The caller must place a barrier between the top and bottom halves.
3321 ac_build_wg_wavescan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3323 const LLVMTypeRef type
= LLVMTypeOf(ws
->src
);
3324 const LLVMValueRef identity
=
3325 get_reduction_identity(ctx
, ws
->op
, ac_get_type_size(type
));
3327 if (ws
->maxwaves
<= 1) {
3328 ws
->result_reduce
= ws
->src
;
3329 ws
->result_inclusive
= ws
->src
;
3330 ws
->result_exclusive
= identity
;
3333 assert(ws
->maxwaves
<= 32);
3335 LLVMBuilderRef builder
= ctx
->builder
;
3336 LLVMValueRef tid
= ac_get_thread_id(ctx
);
3337 LLVMBasicBlockRef bbs
[2];
3338 LLVMValueRef phivalues_scan
[2];
3339 LLVMValueRef tmp
, tmp2
;
3341 bbs
[0] = LLVMGetInsertBlock(builder
);
3342 phivalues_scan
[0] = LLVMGetUndef(type
);
3344 if (ws
->enable_reduce
)
3345 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->numwaves
, "");
3346 else if (ws
->enable_inclusive
)
3347 tmp
= LLVMBuildICmp(builder
, LLVMIntULE
, tid
, ws
->waveidx
, "");
3349 tmp
= LLVMBuildICmp(builder
, LLVMIntULT
, tid
, ws
->waveidx
, "");
3350 ac_build_ifcc(ctx
, tmp
, 1001);
3352 tmp
= LLVMBuildLoad(builder
, LLVMBuildGEP(builder
, ws
->scratch
, &tid
, 1, ""), "");
3354 ac_build_optimization_barrier(ctx
, &tmp
);
3356 bbs
[1] = LLVMGetInsertBlock(builder
);
3357 phivalues_scan
[1] = ac_build_scan(ctx
, ws
->op
, tmp
, identity
, ws
->maxwaves
);
3359 ac_build_endif(ctx
, 1001);
3361 const LLVMValueRef scan
= ac_build_phi(ctx
, type
, 2, phivalues_scan
, bbs
);
3363 if (ws
->enable_reduce
) {
3364 tmp
= LLVMBuildSub(builder
, ws
->numwaves
, ctx
->i32_1
, "");
3365 ws
->result_reduce
= ac_build_readlane(ctx
, scan
, tmp
);
3367 if (ws
->enable_inclusive
)
3368 ws
->result_inclusive
= ac_build_readlane(ctx
, scan
, ws
->waveidx
);
3369 if (ws
->enable_exclusive
) {
3370 tmp
= LLVMBuildSub(builder
, ws
->waveidx
, ctx
->i32_1
, "");
3371 tmp
= ac_build_readlane(ctx
, scan
, tmp
);
3372 tmp2
= LLVMBuildICmp(builder
, LLVMIntEQ
, ws
->waveidx
, ctx
->i32_0
, "");
3373 ws
->result_exclusive
= LLVMBuildSelect(builder
, tmp2
, identity
, tmp
, "");
3378 * Inclusive scan of a per-wave value across an entire workgroup.
3380 * This implies an s_barrier instruction.
3382 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
3383 * of the workgroup are live. (This requirement cannot easily be relaxed in a
3384 * useful manner because of the barrier in the algorithm.)
3387 ac_build_wg_wavescan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3389 ac_build_wg_wavescan_top(ctx
, ws
);
3390 ac_build_s_barrier(ctx
);
3391 ac_build_wg_wavescan_bottom(ctx
, ws
);
3395 * "Top half" of a scan that reduces per-thread values across an entire
3398 * All lanes must be active when this code runs.
3401 ac_build_wg_scan_top(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3403 if (ws
->enable_exclusive
) {
3404 ws
->extra
= ac_build_exclusive_scan(ctx
, ws
->src
, ws
->op
);
3405 if (LLVMTypeOf(ws
->src
) == ctx
->i1
&& ws
->op
== nir_op_iadd
)
3406 ws
->src
= LLVMBuildZExt(ctx
->builder
, ws
->src
, ctx
->i32
, "");
3407 ws
->src
= ac_build_alu_op(ctx
, ws
->extra
, ws
->src
, ws
->op
);
3409 ws
->src
= ac_build_inclusive_scan(ctx
, ws
->src
, ws
->op
);
3412 bool enable_inclusive
= ws
->enable_inclusive
;
3413 bool enable_exclusive
= ws
->enable_exclusive
;
3414 ws
->enable_inclusive
= false;
3415 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
3416 ac_build_wg_wavescan_top(ctx
, ws
);
3417 ws
->enable_inclusive
= enable_inclusive
;
3418 ws
->enable_exclusive
= enable_exclusive
;
3422 * "Bottom half" of a scan that reduces per-thread values across an entire
3425 * The caller must place a barrier between the top and bottom halves.
3428 ac_build_wg_scan_bottom(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3430 bool enable_inclusive
= ws
->enable_inclusive
;
3431 bool enable_exclusive
= ws
->enable_exclusive
;
3432 ws
->enable_inclusive
= false;
3433 ws
->enable_exclusive
= ws
->enable_exclusive
|| enable_inclusive
;
3434 ac_build_wg_wavescan_bottom(ctx
, ws
);
3435 ws
->enable_inclusive
= enable_inclusive
;
3436 ws
->enable_exclusive
= enable_exclusive
;
3438 /* ws->result_reduce is already the correct value */
3439 if (ws
->enable_inclusive
)
3440 ws
->result_inclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->src
, ws
->op
);
3441 if (ws
->enable_exclusive
)
3442 ws
->result_exclusive
= ac_build_alu_op(ctx
, ws
->result_exclusive
, ws
->extra
, ws
->op
);
3446 * A scan that reduces per-thread values across an entire workgroup.
3448 * The caller must ensure that all lanes are active when this code runs
3449 * (WWM is insufficient!), because there is an implied barrier.
3452 ac_build_wg_scan(struct ac_llvm_context
*ctx
, struct ac_wg_scan
*ws
)
3454 ac_build_wg_scan_top(ctx
, ws
);
3455 ac_build_s_barrier(ctx
);
3456 ac_build_wg_scan_bottom(ctx
, ws
);
3460 ac_build_quad_swizzle(struct ac_llvm_context
*ctx
, LLVMValueRef src
,
3461 unsigned lane0
, unsigned lane1
, unsigned lane2
, unsigned lane3
)
3463 unsigned mask
= dpp_quad_perm(lane0
, lane1
, lane2
, lane3
);
3464 if (ctx
->chip_class
>= VI
) {
3465 return ac_build_dpp(ctx
, src
, src
, mask
, 0xf, 0xf, false);
3467 return ac_build_ds_swizzle(ctx
, src
, (1 << 15) | mask
);
3472 ac_build_shuffle(struct ac_llvm_context
*ctx
, LLVMValueRef src
, LLVMValueRef index
)
3474 index
= LLVMBuildMul(ctx
->builder
, index
, LLVMConstInt(ctx
->i32
, 4, 0), "");
3475 return ac_build_intrinsic(ctx
,
3476 "llvm.amdgcn.ds.bpermute", ctx
->i32
,
3477 (LLVMValueRef
[]) {index
, src
}, 2,
3478 AC_FUNC_ATTR_READNONE
|
3479 AC_FUNC_ATTR_CONVERGENT
);