projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac: add if/loop build helpers
[mesa.git] / src / amd / common / ac_llvm_build.c
1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
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:
11 *
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.
19 *
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
22 * of the Software.
23 *
24 */
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
27
28 #include <llvm-c/Core.h>
29
30 #include "c11/threads.h"
31
32 #include <assert.h>
33 #include <stdio.h>
34
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 "sid.h"
41
42 #include "shader_enums.h"
43
44 #define AC_LLVM_INITIAL_CF_DEPTH 4
45
46 /* Data for if/else/endif and bgnloop/endloop control flow structures.
47 */
48 struct ac_llvm_flow {
49 /* Loop exit or next part of if/else/endif. */
50 LLVMBasicBlockRef next_block;
51 LLVMBasicBlockRef loop_entry_block;
52 };
53
54 /* Initialize module-independent parts of the context.
55 *
56 * The caller is responsible for initializing ctx::module and ctx::builder.
57 */
58 void
59 ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context,
60 enum chip_class chip_class, enum radeon_family family)
61 {
62 LLVMValueRef args[1];
63
64 ctx->chip_class = chip_class;
65 ctx->family = family;
66
67 ctx->context = context;
68 ctx->module = NULL;
69 ctx->builder = NULL;
70
71 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
72 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
73 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
74 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
75 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
76 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
77 ctx->intptr = HAVE_32BIT_POINTERS ? ctx->i32 : ctx->i64;
78 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
79 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
80 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
81 ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
82 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
83 ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
84 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
85 ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
86 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
87 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
88
89 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
90 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
91 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
92 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
93 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
94 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
95 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
96 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
97
98 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
99 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
100
101 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
102 "range", 5);
103
104 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
105 "invariant.load", 14);
106
107 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
108
109 args[0] = LLVMConstReal(ctx->f32, 2.5);
110 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
111
112 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
113 "amdgpu.uniform", 14);
114
115 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
116 }
117
118 void
119 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
120 {
121 free(ctx->flow);
122 ctx->flow = NULL;
123 ctx->flow_depth_max = 0;
124 }
125
126 int
127 ac_get_llvm_num_components(LLVMValueRef value)
128 {
129 LLVMTypeRef type = LLVMTypeOf(value);
130 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
131 ? LLVMGetVectorSize(type)
132 : 1;
133 return num_components;
134 }
135
136 LLVMValueRef
137 ac_llvm_extract_elem(struct ac_llvm_context *ac,
138 LLVMValueRef value,
139 int index)
140 {
141 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
142 assert(index == 0);
143 return value;
144 }
145
146 return LLVMBuildExtractElement(ac->builder, value,
147 LLVMConstInt(ac->i32, index, false), "");
148 }
149
150 int
151 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
152 {
153 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
154 type = LLVMGetElementType(type);
155
156 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
157 return LLVMGetIntTypeWidth(type);
158
159 if (type == ctx->f16)
160 return 16;
161 if (type == ctx->f32)
162 return 32;
163 if (type == ctx->f64)
164 return 64;
165
166 unreachable("Unhandled type kind in get_elem_bits");
167 }
168
169 unsigned
170 ac_get_type_size(LLVMTypeRef type)
171 {
172 LLVMTypeKind kind = LLVMGetTypeKind(type);
173
174 switch (kind) {
175 case LLVMIntegerTypeKind:
176 return LLVMGetIntTypeWidth(type) / 8;
177 case LLVMFloatTypeKind:
178 return 4;
179 case LLVMDoubleTypeKind:
180 return 8;
181 case LLVMPointerTypeKind:
182 if (LLVMGetPointerAddressSpace(type) == AC_CONST_32BIT_ADDR_SPACE)
183 return 4;
184 return 8;
185 case LLVMVectorTypeKind:
186 return LLVMGetVectorSize(type) *
187 ac_get_type_size(LLVMGetElementType(type));
188 case LLVMArrayTypeKind:
189 return LLVMGetArrayLength(type) *
190 ac_get_type_size(LLVMGetElementType(type));
191 default:
192 assert(0);
193 return 0;
194 }
195 }
196
197 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
198 {
199 if (t == ctx->f16 || t == ctx->i16)
200 return ctx->i16;
201 else if (t == ctx->f32 || t == ctx->i32)
202 return ctx->i32;
203 else if (t == ctx->f64 || t == ctx->i64)
204 return ctx->i64;
205 else
206 unreachable("Unhandled integer size");
207 }
208
209 LLVMTypeRef
210 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
211 {
212 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
213 LLVMTypeRef elem_type = LLVMGetElementType(t);
214 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
215 LLVMGetVectorSize(t));
216 }
217 return to_integer_type_scalar(ctx, t);
218 }
219
220 LLVMValueRef
221 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
222 {
223 LLVMTypeRef type = LLVMTypeOf(v);
224 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
225 }
226
227 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
228 {
229 if (t == ctx->i16 || t == ctx->f16)
230 return ctx->f16;
231 else if (t == ctx->i32 || t == ctx->f32)
232 return ctx->f32;
233 else if (t == ctx->i64 || t == ctx->f64)
234 return ctx->f64;
235 else
236 unreachable("Unhandled float size");
237 }
238
239 LLVMTypeRef
240 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
241 {
242 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
243 LLVMTypeRef elem_type = LLVMGetElementType(t);
244 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
245 LLVMGetVectorSize(t));
246 }
247 return to_float_type_scalar(ctx, t);
248 }
249
250 LLVMValueRef
251 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
252 {
253 LLVMTypeRef type = LLVMTypeOf(v);
254 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
255 }
256
257
258 LLVMValueRef
259 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
260 LLVMTypeRef return_type, LLVMValueRef *params,
261 unsigned param_count, unsigned attrib_mask)
262 {
263 LLVMValueRef function, call;
264 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
265
266 function = LLVMGetNamedFunction(ctx->module, name);
267 if (!function) {
268 LLVMTypeRef param_types[32], function_type;
269 unsigned i;
270
271 assert(param_count <= 32);
272
273 for (i = 0; i < param_count; ++i) {
274 assert(params[i]);
275 param_types[i] = LLVMTypeOf(params[i]);
276 }
277 function_type =
278 LLVMFunctionType(return_type, param_types, param_count, 0);
279 function = LLVMAddFunction(ctx->module, name, function_type);
280
281 LLVMSetFunctionCallConv(function, LLVMCCallConv);
282 LLVMSetLinkage(function, LLVMExternalLinkage);
283
284 if (!set_callsite_attrs)
285 ac_add_func_attributes(ctx->context, function, attrib_mask);
286 }
287
288 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
289 if (set_callsite_attrs)
290 ac_add_func_attributes(ctx->context, call, attrib_mask);
291 return call;
292 }
293
294 /**
295 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
296 * intrinsic names).
297 */
298 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
299 {
300 LLVMTypeRef elem_type = type;
301
302 assert(bufsize >= 8);
303
304 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
305 int ret = snprintf(buf, bufsize, "v%u",
306 LLVMGetVectorSize(type));
307 if (ret < 0) {
308 char *type_name = LLVMPrintTypeToString(type);
309 fprintf(stderr, "Error building type name for: %s\n",
310 type_name);
311 return;
312 }
313 elem_type = LLVMGetElementType(type);
314 buf += ret;
315 bufsize -= ret;
316 }
317 switch (LLVMGetTypeKind(elem_type)) {
318 default: break;
319 case LLVMIntegerTypeKind:
320 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
321 break;
322 case LLVMFloatTypeKind:
323 snprintf(buf, bufsize, "f32");
324 break;
325 case LLVMDoubleTypeKind:
326 snprintf(buf, bufsize, "f64");
327 break;
328 }
329 }
330
331 /**
332 * Helper function that builds an LLVM IR PHI node and immediately adds
333 * incoming edges.
334 */
335 LLVMValueRef
336 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
337 unsigned count_incoming, LLVMValueRef *values,
338 LLVMBasicBlockRef *blocks)
339 {
340 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
341 LLVMAddIncoming(phi, values, blocks, count_incoming);
342 return phi;
343 }
344
345 /* Prevent optimizations (at least of memory accesses) across the current
346 * point in the program by emitting empty inline assembly that is marked as
347 * having side effects.
348 *
349 * Optionally, a value can be passed through the inline assembly to prevent
350 * LLVM from hoisting calls to ReadNone functions.
351 */
352 void
353 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
354 LLVMValueRef *pvgpr)
355 {
356 static int counter = 0;
357
358 LLVMBuilderRef builder = ctx->builder;
359 char code[16];
360
361 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
362
363 if (!pvgpr) {
364 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
365 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
366 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
367 } else {
368 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
369 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
370 LLVMValueRef vgpr = *pvgpr;
371 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
372 unsigned vgpr_size = ac_get_type_size(vgpr_type);
373 LLVMValueRef vgpr0;
374
375 assert(vgpr_size % 4 == 0);
376
377 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
378 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
379 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
380 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
381 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
382
383 *pvgpr = vgpr;
384 }
385 }
386
387 LLVMValueRef
388 ac_build_shader_clock(struct ac_llvm_context *ctx)
389 {
390 LLVMValueRef tmp = ac_build_intrinsic(ctx, "llvm.readcyclecounter",
391 ctx->i64, NULL, 0, 0);
392 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
393 }
394
395 LLVMValueRef
396 ac_build_ballot(struct ac_llvm_context *ctx,
397 LLVMValueRef value)
398 {
399 LLVMValueRef args[3] = {
400 value,
401 ctx->i32_0,
402 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
403 };
404
405 /* We currently have no other way to prevent LLVM from lifting the icmp
406 * calls to a dominating basic block.
407 */
408 ac_build_optimization_barrier(ctx, &args[0]);
409
410 if (LLVMTypeOf(args[0]) != ctx->i32)
411 args[0] = LLVMBuildBitCast(ctx->builder, args[0], ctx->i32, "");
412
413 return ac_build_intrinsic(ctx,
414 "llvm.amdgcn.icmp.i32",
415 ctx->i64, args, 3,
416 AC_FUNC_ATTR_NOUNWIND |
417 AC_FUNC_ATTR_READNONE |
418 AC_FUNC_ATTR_CONVERGENT);
419 }
420
421 LLVMValueRef
422 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
423 {
424 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
425 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
426 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
427 }
428
429 LLVMValueRef
430 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
431 {
432 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
433 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
434 LLVMConstInt(ctx->i64, 0, 0), "");
435 }
436
437 LLVMValueRef
438 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
439 {
440 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
441 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
442
443 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
444 vote_set, active_set, "");
445 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
446 vote_set,
447 LLVMConstInt(ctx->i64, 0, 0), "");
448 return LLVMBuildOr(ctx->builder, all, none, "");
449 }
450
451 LLVMValueRef
452 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
453 unsigned value_count, unsigned component)
454 {
455 LLVMValueRef vec = NULL;
456
457 if (value_count == 1) {
458 return values[component];
459 } else if (!value_count)
460 unreachable("value_count is 0");
461
462 for (unsigned i = component; i < value_count + component; i++) {
463 LLVMValueRef value = values[i];
464
465 if (i == component)
466 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
467 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
468 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
469 }
470 return vec;
471 }
472
473 LLVMValueRef
474 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
475 LLVMValueRef *values,
476 unsigned value_count,
477 unsigned value_stride,
478 bool load,
479 bool always_vector)
480 {
481 LLVMBuilderRef builder = ctx->builder;
482 LLVMValueRef vec = NULL;
483 unsigned i;
484
485 if (value_count == 1 && !always_vector) {
486 if (load)
487 return LLVMBuildLoad(builder, values[0], "");
488 return values[0];
489 } else if (!value_count)
490 unreachable("value_count is 0");
491
492 for (i = 0; i < value_count; i++) {
493 LLVMValueRef value = values[i * value_stride];
494 if (load)
495 value = LLVMBuildLoad(builder, value, "");
496
497 if (!i)
498 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
499 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
500 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
501 }
502 return vec;
503 }
504
505 LLVMValueRef
506 ac_build_gather_values(struct ac_llvm_context *ctx,
507 LLVMValueRef *values,
508 unsigned value_count)
509 {
510 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
511 }
512
513 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
514 * with undef. Extract at most num_channels components from the input.
515 */
516 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
517 LLVMValueRef value,
518 unsigned num_channels)
519 {
520 LLVMTypeRef elemtype;
521 LLVMValueRef chan[4];
522
523 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
524 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
525 num_channels = MIN2(num_channels, vec_size);
526
527 if (num_channels >= 4)
528 return value;
529
530 for (unsigned i = 0; i < num_channels; i++)
531 chan[i] = ac_llvm_extract_elem(ctx, value, i);
532
533 elemtype = LLVMGetElementType(LLVMTypeOf(value));
534 } else {
535 if (num_channels) {
536 assert(num_channels == 1);
537 chan[0] = value;
538 }
539 elemtype = LLVMTypeOf(value);
540 }
541
542 while (num_channels < 4)
543 chan[num_channels++] = LLVMGetUndef(elemtype);
544
545 return ac_build_gather_values(ctx, chan, 4);
546 }
547
548 LLVMValueRef
549 ac_build_fdiv(struct ac_llvm_context *ctx,
550 LLVMValueRef num,
551 LLVMValueRef den)
552 {
553 LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
554
555 /* Use v_rcp_f32 instead of precise division. */
556 if (!LLVMIsConstant(ret))
557 LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
558 return ret;
559 }
560
561 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
562 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
563 * already multiplied by two. id is the cube face number.
564 */
565 struct cube_selection_coords {
566 LLVMValueRef stc[2];
567 LLVMValueRef ma;
568 LLVMValueRef id;
569 };
570
571 static void
572 build_cube_intrinsic(struct ac_llvm_context *ctx,
573 LLVMValueRef in[3],
574 struct cube_selection_coords *out)
575 {
576 LLVMTypeRef f32 = ctx->f32;
577
578 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
579 f32, in, 3, AC_FUNC_ATTR_READNONE);
580 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
581 f32, in, 3, AC_FUNC_ATTR_READNONE);
582 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
583 f32, in, 3, AC_FUNC_ATTR_READNONE);
584 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
585 f32, in, 3, AC_FUNC_ATTR_READNONE);
586 }
587
588 /**
589 * Build a manual selection sequence for cube face sc/tc coordinates and
590 * major axis vector (multiplied by 2 for consistency) for the given
591 * vec3 \p coords, for the face implied by \p selcoords.
592 *
593 * For the major axis, we always adjust the sign to be in the direction of
594 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
595 * the selcoords major axis.
596 */
597 static void build_cube_select(struct ac_llvm_context *ctx,
598 const struct cube_selection_coords *selcoords,
599 const LLVMValueRef *coords,
600 LLVMValueRef *out_st,
601 LLVMValueRef *out_ma)
602 {
603 LLVMBuilderRef builder = ctx->builder;
604 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
605 LLVMValueRef is_ma_positive;
606 LLVMValueRef sgn_ma;
607 LLVMValueRef is_ma_z, is_not_ma_z;
608 LLVMValueRef is_ma_y;
609 LLVMValueRef is_ma_x;
610 LLVMValueRef sgn;
611 LLVMValueRef tmp;
612
613 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
614 selcoords->ma, LLVMConstReal(f32, 0.0), "");
615 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
616 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
617
618 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
619 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
620 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
621 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
622 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
623
624 /* Select sc */
625 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
626 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
627 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
628 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
629 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
630
631 /* Select tc */
632 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
633 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
634 LLVMConstReal(f32, -1.0), "");
635 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
636
637 /* Select ma */
638 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
639 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
640 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
641 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
642 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
643 }
644
645 void
646 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
647 bool is_deriv, bool is_array, bool is_lod,
648 LLVMValueRef *coords_arg,
649 LLVMValueRef *derivs_arg)
650 {
651
652 LLVMBuilderRef builder = ctx->builder;
653 struct cube_selection_coords selcoords;
654 LLVMValueRef coords[3];
655 LLVMValueRef invma;
656
657 if (is_array && !is_lod) {
658 LLVMValueRef tmp = coords_arg[3];
659 tmp = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &tmp, 1, 0);
660
661 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
662 *
663 * "For Array forms, the array layer used will be
664 *
665 * max(0, min(d−1, floor(layer+0.5)))
666 *
667 * where d is the depth of the texture array and layer
668 * comes from the component indicated in the tables below.
669 * Workaroudn for an issue where the layer is taken from a
670 * helper invocation which happens to fall on a different
671 * layer due to extrapolation."
672 *
673 * VI and earlier attempt to implement this in hardware by
674 * clamping the value of coords[2] = (8 * layer) + face.
675 * Unfortunately, this means that the we end up with the wrong
676 * face when clamping occurs.
677 *
678 * Clamp the layer earlier to work around the issue.
679 */
680 if (ctx->chip_class <= VI) {
681 LLVMValueRef ge0;
682 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
683 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
684 }
685
686 coords_arg[3] = tmp;
687 }
688
689 build_cube_intrinsic(ctx, coords_arg, &selcoords);
690
691 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
692 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
693 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
694
695 for (int i = 0; i < 2; ++i)
696 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
697
698 coords[2] = selcoords.id;
699
700 if (is_deriv && derivs_arg) {
701 LLVMValueRef derivs[4];
702 int axis;
703
704 /* Convert cube derivatives to 2D derivatives. */
705 for (axis = 0; axis < 2; axis++) {
706 LLVMValueRef deriv_st[2];
707 LLVMValueRef deriv_ma;
708
709 /* Transform the derivative alongside the texture
710 * coordinate. Mathematically, the correct formula is
711 * as follows. Assume we're projecting onto the +Z face
712 * and denote by dx/dh the derivative of the (original)
713 * X texture coordinate with respect to horizontal
714 * window coordinates. The projection onto the +Z face
715 * plane is:
716 *
717 * f(x,z) = x/z
718 *
719 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
720 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
721 *
722 * This motivatives the implementation below.
723 *
724 * Whether this actually gives the expected results for
725 * apps that might feed in derivatives obtained via
726 * finite differences is anyone's guess. The OpenGL spec
727 * seems awfully quiet about how textureGrad for cube
728 * maps should be handled.
729 */
730 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
731 deriv_st, &deriv_ma);
732
733 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
734
735 for (int i = 0; i < 2; ++i)
736 derivs[axis * 2 + i] =
737 LLVMBuildFSub(builder,
738 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
739 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
740 }
741
742 memcpy(derivs_arg, derivs, sizeof(derivs));
743 }
744
745 /* Shift the texture coordinate. This must be applied after the
746 * derivative calculation.
747 */
748 for (int i = 0; i < 2; ++i)
749 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
750
751 if (is_array) {
752 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
753 /* coords_arg.w component - array_index for cube arrays */
754 LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
755 coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
756 }
757
758 memcpy(coords_arg, coords, sizeof(coords));
759 }
760
761
762 LLVMValueRef
763 ac_build_fs_interp(struct ac_llvm_context *ctx,
764 LLVMValueRef llvm_chan,
765 LLVMValueRef attr_number,
766 LLVMValueRef params,
767 LLVMValueRef i,
768 LLVMValueRef j)
769 {
770 LLVMValueRef args[5];
771 LLVMValueRef p1;
772
773 args[0] = i;
774 args[1] = llvm_chan;
775 args[2] = attr_number;
776 args[3] = params;
777
778 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
779 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
780
781 args[0] = p1;
782 args[1] = j;
783 args[2] = llvm_chan;
784 args[3] = attr_number;
785 args[4] = params;
786
787 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
788 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
789 }
790
791 LLVMValueRef
792 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
793 LLVMValueRef parameter,
794 LLVMValueRef llvm_chan,
795 LLVMValueRef attr_number,
796 LLVMValueRef params)
797 {
798 LLVMValueRef args[4];
799
800 args[0] = parameter;
801 args[1] = llvm_chan;
802 args[2] = attr_number;
803 args[3] = params;
804
805 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
806 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
807 }
808
809 LLVMValueRef
810 ac_build_gep0(struct ac_llvm_context *ctx,
811 LLVMValueRef base_ptr,
812 LLVMValueRef index)
813 {
814 LLVMValueRef indices[2] = {
815 LLVMConstInt(ctx->i32, 0, 0),
816 index,
817 };
818 return LLVMBuildGEP(ctx->builder, base_ptr,
819 indices, 2, "");
820 }
821
822 void
823 ac_build_indexed_store(struct ac_llvm_context *ctx,
824 LLVMValueRef base_ptr, LLVMValueRef index,
825 LLVMValueRef value)
826 {
827 LLVMBuildStore(ctx->builder, value,
828 ac_build_gep0(ctx, base_ptr, index));
829 }
830
831 /**
832 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
833 * It's equivalent to doing a load from &base_ptr[index].
834 *
835 * \param base_ptr Where the array starts.
836 * \param index The element index into the array.
837 * \param uniform Whether the base_ptr and index can be assumed to be
838 * dynamically uniform (i.e. load to an SGPR)
839 * \param invariant Whether the load is invariant (no other opcodes affect it)
840 */
841 static LLVMValueRef
842 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
843 LLVMValueRef index, bool uniform, bool invariant)
844 {
845 LLVMValueRef pointer, result;
846
847 pointer = ac_build_gep0(ctx, base_ptr, index);
848 if (uniform)
849 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
850 result = LLVMBuildLoad(ctx->builder, pointer, "");
851 if (invariant)
852 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
853 return result;
854 }
855
856 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
857 LLVMValueRef index)
858 {
859 return ac_build_load_custom(ctx, base_ptr, index, false, false);
860 }
861
862 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
863 LLVMValueRef base_ptr, LLVMValueRef index)
864 {
865 return ac_build_load_custom(ctx, base_ptr, index, false, true);
866 }
867
868 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
869 LLVMValueRef base_ptr, LLVMValueRef index)
870 {
871 return ac_build_load_custom(ctx, base_ptr, index, true, true);
872 }
873
874 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
875 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
876 * or v4i32 (num_channels=3,4).
877 */
878 void
879 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
880 LLVMValueRef rsrc,
881 LLVMValueRef vdata,
882 unsigned num_channels,
883 LLVMValueRef voffset,
884 LLVMValueRef soffset,
885 unsigned inst_offset,
886 bool glc,
887 bool slc,
888 bool writeonly_memory,
889 bool swizzle_enable_hint)
890 {
891 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
892 * (voffset is swizzled, but soffset isn't swizzled).
893 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
894 */
895 if (!swizzle_enable_hint) {
896 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
897 * intrinsics. */
898 if (num_channels == 3) {
899 LLVMValueRef v[3], v01;
900
901 for (int i = 0; i < 3; i++) {
902 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
903 LLVMConstInt(ctx->i32, i, 0), "");
904 }
905 v01 = ac_build_gather_values(ctx, v, 2);
906
907 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
908 soffset, inst_offset, glc, slc,
909 writeonly_memory, swizzle_enable_hint);
910 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
911 soffset, inst_offset + 8,
912 glc, slc,
913 writeonly_memory, swizzle_enable_hint);
914 return;
915 }
916
917 unsigned func = CLAMP(num_channels, 1, 3) - 1;
918 static const char *types[] = {"f32", "v2f32", "v4f32"};
919 char name[256];
920 LLVMValueRef offset = soffset;
921
922 if (inst_offset)
923 offset = LLVMBuildAdd(ctx->builder, offset,
924 LLVMConstInt(ctx->i32, inst_offset, 0), "");
925 if (voffset)
926 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
927
928 LLVMValueRef args[] = {
929 ac_to_float(ctx, vdata),
930 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
931 LLVMConstInt(ctx->i32, 0, 0),
932 offset,
933 LLVMConstInt(ctx->i1, glc, 0),
934 LLVMConstInt(ctx->i1, slc, 0),
935 };
936
937 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
938 types[func]);
939
940 ac_build_intrinsic(ctx, name, ctx->voidt,
941 args, ARRAY_SIZE(args),
942 writeonly_memory ?
943 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
944 AC_FUNC_ATTR_WRITEONLY);
945 return;
946 }
947
948 static unsigned dfmt[] = {
949 V_008F0C_BUF_DATA_FORMAT_32,
950 V_008F0C_BUF_DATA_FORMAT_32_32,
951 V_008F0C_BUF_DATA_FORMAT_32_32_32,
952 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
953 };
954 assert(num_channels >= 1 && num_channels <= 4);
955
956 LLVMValueRef args[] = {
957 rsrc,
958 vdata,
959 LLVMConstInt(ctx->i32, num_channels, 0),
960 voffset ? voffset : LLVMGetUndef(ctx->i32),
961 soffset,
962 LLVMConstInt(ctx->i32, inst_offset, 0),
963 LLVMConstInt(ctx->i32, dfmt[num_channels - 1], 0),
964 LLVMConstInt(ctx->i32, V_008F0C_BUF_NUM_FORMAT_UINT, 0),
965 LLVMConstInt(ctx->i32, voffset != NULL, 0),
966 LLVMConstInt(ctx->i32, 0, 0), /* idxen */
967 LLVMConstInt(ctx->i32, glc, 0),
968 LLVMConstInt(ctx->i32, slc, 0),
969 LLVMConstInt(ctx->i32, 0, 0), /* tfe*/
970 };
971
972 /* The instruction offset field has 12 bits */
973 assert(voffset || inst_offset < (1 << 12));
974
975 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
976 unsigned func = CLAMP(num_channels, 1, 3) - 1;
977 const char *types[] = {"i32", "v2i32", "v4i32"};
978 char name[256];
979 snprintf(name, sizeof(name), "llvm.SI.tbuffer.store.%s", types[func]);
980
981 ac_build_intrinsic(ctx, name, ctx->voidt,
982 args, ARRAY_SIZE(args),
983 AC_FUNC_ATTR_LEGACY);
984 }
985
986 static LLVMValueRef
987 ac_build_buffer_load_common(struct ac_llvm_context *ctx,
988 LLVMValueRef rsrc,
989 LLVMValueRef vindex,
990 LLVMValueRef voffset,
991 unsigned num_channels,
992 bool glc,
993 bool slc,
994 bool can_speculate,
995 bool use_format)
996 {
997 LLVMValueRef args[] = {
998 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
999 vindex ? vindex : LLVMConstInt(ctx->i32, 0, 0),
1000 voffset,
1001 LLVMConstInt(ctx->i1, glc, 0),
1002 LLVMConstInt(ctx->i1, slc, 0)
1003 };
1004 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1005
1006 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1007 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1008 char name[256];
1009
1010 if (use_format) {
1011 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
1012 type_names[func]);
1013 } else {
1014 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1015 type_names[func]);
1016 }
1017
1018 return ac_build_intrinsic(ctx, name, types[func], args,
1019 ARRAY_SIZE(args),
1020 ac_get_load_intr_attribs(can_speculate));
1021 }
1022
1023 LLVMValueRef
1024 ac_build_buffer_load(struct ac_llvm_context *ctx,
1025 LLVMValueRef rsrc,
1026 int num_channels,
1027 LLVMValueRef vindex,
1028 LLVMValueRef voffset,
1029 LLVMValueRef soffset,
1030 unsigned inst_offset,
1031 unsigned glc,
1032 unsigned slc,
1033 bool can_speculate,
1034 bool allow_smem)
1035 {
1036 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1037 if (voffset)
1038 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1039 if (soffset)
1040 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1041
1042 /* TODO: VI and later generations can use SMEM with GLC=1.*/
1043 if (allow_smem && !glc && !slc) {
1044 assert(vindex == NULL);
1045
1046 LLVMValueRef result[8];
1047
1048 for (int i = 0; i < num_channels; i++) {
1049 if (i) {
1050 offset = LLVMBuildAdd(ctx->builder, offset,
1051 LLVMConstInt(ctx->i32, 4, 0), "");
1052 }
1053 LLVMValueRef args[2] = {rsrc, offset};
1054 result[i] = ac_build_intrinsic(ctx, "llvm.SI.load.const.v4i32",
1055 ctx->f32, args, 2,
1056 AC_FUNC_ATTR_READNONE |
1057 AC_FUNC_ATTR_LEGACY);
1058 }
1059 if (num_channels == 1)
1060 return result[0];
1061
1062 if (num_channels == 3)
1063 result[num_channels++] = LLVMGetUndef(ctx->f32);
1064 return ac_build_gather_values(ctx, result, num_channels);
1065 }
1066
1067 return ac_build_buffer_load_common(ctx, rsrc, vindex, offset,
1068 num_channels, glc, slc,
1069 can_speculate, false);
1070 }
1071
1072 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1073 LLVMValueRef rsrc,
1074 LLVMValueRef vindex,
1075 LLVMValueRef voffset,
1076 unsigned num_channels,
1077 bool glc,
1078 bool can_speculate)
1079 {
1080 return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1081 num_channels, glc, false,
1082 can_speculate, true);
1083 }
1084
1085 /**
1086 * Set range metadata on an instruction. This can only be used on load and
1087 * call instructions. If you know an instruction can only produce the values
1088 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1089 * \p lo is the minimum value inclusive.
1090 * \p hi is the maximum value exclusive.
1091 */
1092 static void set_range_metadata(struct ac_llvm_context *ctx,
1093 LLVMValueRef value, unsigned lo, unsigned hi)
1094 {
1095 LLVMValueRef range_md, md_args[2];
1096 LLVMTypeRef type = LLVMTypeOf(value);
1097 LLVMContextRef context = LLVMGetTypeContext(type);
1098
1099 md_args[0] = LLVMConstInt(type, lo, false);
1100 md_args[1] = LLVMConstInt(type, hi, false);
1101 range_md = LLVMMDNodeInContext(context, md_args, 2);
1102 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1103 }
1104
1105 LLVMValueRef
1106 ac_get_thread_id(struct ac_llvm_context *ctx)
1107 {
1108 LLVMValueRef tid;
1109
1110 LLVMValueRef tid_args[2];
1111 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
1112 tid_args[1] = LLVMConstInt(ctx->i32, 0, false);
1113 tid_args[1] = ac_build_intrinsic(ctx,
1114 "llvm.amdgcn.mbcnt.lo", ctx->i32,
1115 tid_args, 2, AC_FUNC_ATTR_READNONE);
1116
1117 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
1118 ctx->i32, tid_args,
1119 2, AC_FUNC_ATTR_READNONE);
1120 set_range_metadata(ctx, tid, 0, 64);
1121 return tid;
1122 }
1123
1124 /*
1125 * SI implements derivatives using the local data store (LDS)
1126 * All writes to the LDS happen in all executing threads at
1127 * the same time. TID is the Thread ID for the current
1128 * thread and is a value between 0 and 63, representing
1129 * the thread's position in the wavefront.
1130 *
1131 * For the pixel shader threads are grouped into quads of four pixels.
1132 * The TIDs of the pixels of a quad are:
1133 *
1134 * +------+------+
1135 * |4n + 0|4n + 1|
1136 * +------+------+
1137 * |4n + 2|4n + 3|
1138 * +------+------+
1139 *
1140 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1141 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1142 * the current pixel's column, and masking with 0xfffffffe yields the TID
1143 * of the left pixel of the current pixel's row.
1144 *
1145 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1146 * adding 2 yields the TID of the pixel below the top pixel.
1147 */
1148 LLVMValueRef
1149 ac_build_ddxy(struct ac_llvm_context *ctx,
1150 uint32_t mask,
1151 int idx,
1152 LLVMValueRef val)
1153 {
1154 LLVMValueRef tl, trbl, args[2];
1155 LLVMValueRef result;
1156
1157 if (ctx->chip_class >= VI) {
1158 LLVMValueRef thread_id, tl_tid, trbl_tid;
1159 thread_id = ac_get_thread_id(ctx);
1160
1161 tl_tid = LLVMBuildAnd(ctx->builder, thread_id,
1162 LLVMConstInt(ctx->i32, mask, false), "");
1163
1164 trbl_tid = LLVMBuildAdd(ctx->builder, tl_tid,
1165 LLVMConstInt(ctx->i32, idx, false), "");
1166
1167 args[0] = LLVMBuildMul(ctx->builder, tl_tid,
1168 LLVMConstInt(ctx->i32, 4, false), "");
1169 args[1] = val;
1170 tl = ac_build_intrinsic(ctx,
1171 "llvm.amdgcn.ds.bpermute", ctx->i32,
1172 args, 2,
1173 AC_FUNC_ATTR_READNONE |
1174 AC_FUNC_ATTR_CONVERGENT);
1175
1176 args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
1177 LLVMConstInt(ctx->i32, 4, false), "");
1178 trbl = ac_build_intrinsic(ctx,
1179 "llvm.amdgcn.ds.bpermute", ctx->i32,
1180 args, 2,
1181 AC_FUNC_ATTR_READNONE |
1182 AC_FUNC_ATTR_CONVERGENT);
1183 } else {
1184 uint32_t masks[2] = {};
1185
1186 switch (mask) {
1187 case AC_TID_MASK_TOP_LEFT:
1188 masks[0] = 0x8000;
1189 if (idx == 1)
1190 masks[1] = 0x8055;
1191 else
1192 masks[1] = 0x80aa;
1193
1194 break;
1195 case AC_TID_MASK_TOP:
1196 masks[0] = 0x8044;
1197 masks[1] = 0x80ee;
1198 break;
1199 case AC_TID_MASK_LEFT:
1200 masks[0] = 0x80a0;
1201 masks[1] = 0x80f5;
1202 break;
1203 default:
1204 assert(0);
1205 }
1206
1207 args[0] = val;
1208 args[1] = LLVMConstInt(ctx->i32, masks[0], false);
1209
1210 tl = ac_build_intrinsic(ctx,
1211 "llvm.amdgcn.ds.swizzle", ctx->i32,
1212 args, 2,
1213 AC_FUNC_ATTR_READNONE |
1214 AC_FUNC_ATTR_CONVERGENT);
1215
1216 args[1] = LLVMConstInt(ctx->i32, masks[1], false);
1217 trbl = ac_build_intrinsic(ctx,
1218 "llvm.amdgcn.ds.swizzle", ctx->i32,
1219 args, 2,
1220 AC_FUNC_ATTR_READNONE |
1221 AC_FUNC_ATTR_CONVERGENT);
1222 }
1223
1224 tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
1225 trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
1226 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
1227 return result;
1228 }
1229
1230 void
1231 ac_build_sendmsg(struct ac_llvm_context *ctx,
1232 uint32_t msg,
1233 LLVMValueRef wave_id)
1234 {
1235 LLVMValueRef args[2];
1236 args[0] = LLVMConstInt(ctx->i32, msg, false);
1237 args[1] = wave_id;
1238 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
1239 }
1240
1241 LLVMValueRef
1242 ac_build_imsb(struct ac_llvm_context *ctx,
1243 LLVMValueRef arg,
1244 LLVMTypeRef dst_type)
1245 {
1246 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
1247 dst_type, &arg, 1,
1248 AC_FUNC_ATTR_READNONE);
1249
1250 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1251 * the index from LSB. Invert it by doing "31 - msb". */
1252 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1253 msb, "");
1254
1255 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
1256 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
1257 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1258 arg, LLVMConstInt(ctx->i32, 0, 0), ""),
1259 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1260 arg, all_ones, ""), "");
1261
1262 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
1263 }
1264
1265 LLVMValueRef
1266 ac_build_umsb(struct ac_llvm_context *ctx,
1267 LLVMValueRef arg,
1268 LLVMTypeRef dst_type)
1269 {
1270 LLVMValueRef args[2] = {
1271 arg,
1272 ctx->i1true,
1273 };
1274 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.ctlz.i32",
1275 dst_type, args, ARRAY_SIZE(args),
1276 AC_FUNC_ATTR_READNONE);
1277
1278 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1279 * the index from LSB. Invert it by doing "31 - msb". */
1280 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1281 msb, "");
1282
1283 /* check for zero */
1284 return LLVMBuildSelect(ctx->builder,
1285 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg,
1286 LLVMConstInt(ctx->i32, 0, 0), ""),
1287 LLVMConstInt(ctx->i32, -1, true), msb, "");
1288 }
1289
1290 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
1291 LLVMValueRef b)
1292 {
1293 LLVMValueRef args[2] = {a, b};
1294 return ac_build_intrinsic(ctx, "llvm.minnum.f32", ctx->f32, args, 2,
1295 AC_FUNC_ATTR_READNONE);
1296 }
1297
1298 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
1299 LLVMValueRef b)
1300 {
1301 LLVMValueRef args[2] = {a, b};
1302 return ac_build_intrinsic(ctx, "llvm.maxnum.f32", ctx->f32, args, 2,
1303 AC_FUNC_ATTR_READNONE);
1304 }
1305
1306 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
1307 LLVMValueRef b)
1308 {
1309 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
1310 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1311 }
1312
1313 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
1314 LLVMValueRef b)
1315 {
1316 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
1317 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1318 }
1319
1320 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
1321 LLVMValueRef b)
1322 {
1323 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
1324 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1325 }
1326
1327 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
1328 {
1329 if (HAVE_LLVM >= 0x0500) {
1330 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, ctx->f32_0),
1331 ctx->f32_1);
1332 }
1333
1334 LLVMValueRef args[3] = {
1335 value,
1336 LLVMConstReal(ctx->f32, 0),
1337 LLVMConstReal(ctx->f32, 1),
1338 };
1339
1340 return ac_build_intrinsic(ctx, "llvm.AMDGPU.clamp.", ctx->f32, args, 3,
1341 AC_FUNC_ATTR_READNONE |
1342 AC_FUNC_ATTR_LEGACY);
1343 }
1344
1345 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
1346 {
1347 LLVMValueRef args[9];
1348
1349 if (HAVE_LLVM >= 0x0500) {
1350 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
1351 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1352
1353 if (a->compr) {
1354 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
1355 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
1356
1357 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
1358 v2i16, "");
1359 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
1360 v2i16, "");
1361 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
1362 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1363
1364 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
1365 ctx->voidt, args, 6, 0);
1366 } else {
1367 args[2] = a->out[0];
1368 args[3] = a->out[1];
1369 args[4] = a->out[2];
1370 args[5] = a->out[3];
1371 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
1372 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1373
1374 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
1375 ctx->voidt, args, 8, 0);
1376 }
1377 return;
1378 }
1379
1380 args[0] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1381 args[1] = LLVMConstInt(ctx->i32, a->valid_mask, 0);
1382 args[2] = LLVMConstInt(ctx->i32, a->done, 0);
1383 args[3] = LLVMConstInt(ctx->i32, a->target, 0);
1384 args[4] = LLVMConstInt(ctx->i32, a->compr, 0);
1385 memcpy(args + 5, a->out, sizeof(a->out[0]) * 4);
1386
1387 ac_build_intrinsic(ctx, "llvm.SI.export", ctx->voidt, args, 9,
1388 AC_FUNC_ATTR_LEGACY);
1389 }
1390
1391 void ac_build_export_null(struct ac_llvm_context *ctx)
1392 {
1393 struct ac_export_args args;
1394
1395 args.enabled_channels = 0x0; /* enabled channels */
1396 args.valid_mask = 1; /* whether the EXEC mask is valid */
1397 args.done = 1; /* DONE bit */
1398 args.target = V_008DFC_SQ_EXP_NULL;
1399 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
1400 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
1401 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
1402 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
1403 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
1404
1405 ac_build_export(ctx, &args);
1406 }
1407
1408 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
1409 struct ac_image_args *a)
1410 {
1411 LLVMValueRef args[11];
1412 unsigned num_args = 0;
1413 const char *name = NULL;
1414 char intr_name[128], type[64];
1415
1416 bool sample = a->opcode == ac_image_sample ||
1417 a->opcode == ac_image_gather4 ||
1418 a->opcode == ac_image_get_lod;
1419
1420 if (sample)
1421 args[num_args++] = ac_to_float(ctx, a->addr);
1422 else
1423 args[num_args++] = a->addr;
1424
1425 args[num_args++] = a->resource;
1426 if (sample)
1427 args[num_args++] = a->sampler;
1428 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, 0);
1429 if (sample)
1430 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, 0);
1431 args[num_args++] = ctx->i1false; /* glc */
1432 args[num_args++] = ctx->i1false; /* slc */
1433 args[num_args++] = ctx->i1false; /* lwe */
1434 args[num_args++] = LLVMConstInt(ctx->i1, a->da, 0);
1435
1436 switch (a->opcode) {
1437 case ac_image_sample:
1438 name = "llvm.amdgcn.image.sample";
1439 break;
1440 case ac_image_gather4:
1441 name = "llvm.amdgcn.image.gather4";
1442 break;
1443 case ac_image_load:
1444 name = "llvm.amdgcn.image.load";
1445 break;
1446 case ac_image_load_mip:
1447 name = "llvm.amdgcn.image.load.mip";
1448 break;
1449 case ac_image_get_lod:
1450 name = "llvm.amdgcn.image.getlod";
1451 break;
1452 case ac_image_get_resinfo:
1453 name = "llvm.amdgcn.image.getresinfo";
1454 break;
1455 default:
1456 unreachable("invalid image opcode");
1457 }
1458
1459 ac_build_type_name_for_intr(LLVMTypeOf(args[0]), type,
1460 sizeof(type));
1461
1462 snprintf(intr_name, sizeof(intr_name), "%s%s%s%s.v4f32.%s.v8i32",
1463 name,
1464 a->compare ? ".c" : "",
1465 a->bias ? ".b" :
1466 a->lod ? ".l" :
1467 a->deriv ? ".d" :
1468 a->level_zero ? ".lz" : "",
1469 a->offset ? ".o" : "",
1470 type);
1471
1472 LLVMValueRef result =
1473 ac_build_intrinsic(ctx, intr_name,
1474 ctx->v4f32, args, num_args,
1475 AC_FUNC_ATTR_READNONE);
1476 if (!sample) {
1477 result = LLVMBuildBitCast(ctx->builder, result,
1478 ctx->v4i32, "");
1479 }
1480 return result;
1481 }
1482
1483 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
1484 LLVMValueRef args[2])
1485 {
1486 if (HAVE_LLVM >= 0x0500) {
1487 LLVMTypeRef v2f16 =
1488 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
1489 LLVMValueRef res =
1490 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz",
1491 v2f16, args, 2,
1492 AC_FUNC_ATTR_READNONE);
1493 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1494 }
1495
1496 return ac_build_intrinsic(ctx, "llvm.SI.packf16", ctx->i32, args, 2,
1497 AC_FUNC_ATTR_READNONE |
1498 AC_FUNC_ATTR_LEGACY);
1499 }
1500
1501 /* Upper 16 bits must be zero. */
1502 static LLVMValueRef ac_llvm_pack_two_int16(struct ac_llvm_context *ctx,
1503 LLVMValueRef val[2])
1504 {
1505 return LLVMBuildOr(ctx->builder, val[0],
1506 LLVMBuildShl(ctx->builder, val[1],
1507 LLVMConstInt(ctx->i32, 16, 0),
1508 ""), "");
1509 }
1510
1511 /* Upper 16 bits are ignored and will be dropped. */
1512 static LLVMValueRef ac_llvm_pack_two_int32_as_int16(struct ac_llvm_context *ctx,
1513 LLVMValueRef val[2])
1514 {
1515 LLVMValueRef v[2] = {
1516 LLVMBuildAnd(ctx->builder, val[0],
1517 LLVMConstInt(ctx->i32, 0xffff, 0), ""),
1518 val[1],
1519 };
1520 return ac_llvm_pack_two_int16(ctx, v);
1521 }
1522
1523 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
1524 LLVMValueRef args[2])
1525 {
1526 if (HAVE_LLVM >= 0x0600) {
1527 LLVMValueRef res =
1528 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
1529 ctx->v2i16, args, 2,
1530 AC_FUNC_ATTR_READNONE);
1531 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1532 }
1533
1534 LLVMValueRef val[2];
1535
1536 for (int chan = 0; chan < 2; chan++) {
1537 /* Clamp between [-1, 1]. */
1538 val[chan] = ac_build_fmin(ctx, args[chan], ctx->f32_1);
1539 val[chan] = ac_build_fmax(ctx, val[chan], LLVMConstReal(ctx->f32, -1));
1540 /* Convert to a signed integer in [-32767, 32767]. */
1541 val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
1542 LLVMConstReal(ctx->f32, 32767), "");
1543 /* If positive, add 0.5, else add -0.5. */
1544 val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
1545 LLVMBuildSelect(ctx->builder,
1546 LLVMBuildFCmp(ctx->builder, LLVMRealOGE,
1547 val[chan], ctx->f32_0, ""),
1548 LLVMConstReal(ctx->f32, 0.5),
1549 LLVMConstReal(ctx->f32, -0.5), ""), "");
1550 val[chan] = LLVMBuildFPToSI(ctx->builder, val[chan], ctx->i32, "");
1551 }
1552 return ac_llvm_pack_two_int32_as_int16(ctx, val);
1553 }
1554
1555 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
1556 LLVMValueRef args[2])
1557 {
1558 if (HAVE_LLVM >= 0x0600) {
1559 LLVMValueRef res =
1560 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
1561 ctx->v2i16, args, 2,
1562 AC_FUNC_ATTR_READNONE);
1563 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1564 }
1565
1566 LLVMValueRef val[2];
1567
1568 for (int chan = 0; chan < 2; chan++) {
1569 val[chan] = ac_build_clamp(ctx, args[chan]);
1570 val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
1571 LLVMConstReal(ctx->f32, 65535), "");
1572 val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
1573 LLVMConstReal(ctx->f32, 0.5), "");
1574 val[chan] = LLVMBuildFPToUI(ctx->builder, val[chan],
1575 ctx->i32, "");
1576 }
1577 return ac_llvm_pack_two_int32_as_int16(ctx, val);
1578 }
1579
1580 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1581 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
1582 LLVMValueRef args[2], unsigned bits, bool hi)
1583 {
1584 assert(bits == 8 || bits == 10 || bits == 16);
1585
1586 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1587 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
1588 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
1589 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
1590 LLVMValueRef max_alpha =
1591 bits != 10 ? max_rgb : ctx->i32_1;
1592 LLVMValueRef min_alpha =
1593 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
1594 bool has_intrinsic = HAVE_LLVM >= 0x0600;
1595
1596 /* Clamp. */
1597 if (!has_intrinsic || bits != 16) {
1598 for (int i = 0; i < 2; i++) {
1599 bool alpha = hi && i == 1;
1600 args[i] = ac_build_imin(ctx, args[i],
1601 alpha ? max_alpha : max_rgb);
1602 args[i] = ac_build_imax(ctx, args[i],
1603 alpha ? min_alpha : min_rgb);
1604 }
1605 }
1606
1607 if (has_intrinsic) {
1608 LLVMValueRef res =
1609 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
1610 ctx->v2i16, args, 2,
1611 AC_FUNC_ATTR_READNONE);
1612 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1613 }
1614
1615 return ac_llvm_pack_two_int32_as_int16(ctx, args);
1616 }
1617
1618 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1619 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
1620 LLVMValueRef args[2], unsigned bits, bool hi)
1621 {
1622 assert(bits == 8 || bits == 10 || bits == 16);
1623
1624 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1625 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
1626 LLVMValueRef max_alpha =
1627 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
1628 bool has_intrinsic = HAVE_LLVM >= 0x0600;
1629
1630 /* Clamp. */
1631 if (!has_intrinsic || bits != 16) {
1632 for (int i = 0; i < 2; i++) {
1633 bool alpha = hi && i == 1;
1634 args[i] = ac_build_umin(ctx, args[i],
1635 alpha ? max_alpha : max_rgb);
1636 }
1637 }
1638
1639 if (has_intrinsic) {
1640 LLVMValueRef res =
1641 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
1642 ctx->v2i16, args, 2,
1643 AC_FUNC_ATTR_READNONE);
1644 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1645 }
1646
1647 return ac_llvm_pack_two_int16(ctx, args);
1648 }
1649
1650 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
1651 {
1652 assert(HAVE_LLVM >= 0x0600);
1653 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
1654 &i1, 1, AC_FUNC_ATTR_READNONE);
1655 }
1656
1657 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
1658 {
1659 if (HAVE_LLVM >= 0x0600) {
1660 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
1661 &i1, 1, 0);
1662 return;
1663 }
1664
1665 LLVMValueRef value = LLVMBuildSelect(ctx->builder, i1,
1666 LLVMConstReal(ctx->f32, 1),
1667 LLVMConstReal(ctx->f32, -1), "");
1668 ac_build_intrinsic(ctx, "llvm.AMDGPU.kill", ctx->voidt,
1669 &value, 1, AC_FUNC_ATTR_LEGACY);
1670 }
1671
1672 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
1673 LLVMValueRef offset, LLVMValueRef width,
1674 bool is_signed)
1675 {
1676 LLVMValueRef args[] = {
1677 input,
1678 offset,
1679 width,
1680 };
1681
1682 if (HAVE_LLVM >= 0x0500) {
1683 return ac_build_intrinsic(ctx,
1684 is_signed ? "llvm.amdgcn.sbfe.i32" :
1685 "llvm.amdgcn.ubfe.i32",
1686 ctx->i32, args, 3,
1687 AC_FUNC_ATTR_READNONE);
1688 }
1689
1690 return ac_build_intrinsic(ctx,
1691 is_signed ? "llvm.AMDGPU.bfe.i32" :
1692 "llvm.AMDGPU.bfe.u32",
1693 ctx->i32, args, 3,
1694 AC_FUNC_ATTR_READNONE |
1695 AC_FUNC_ATTR_LEGACY);
1696 }
1697
1698 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
1699 {
1700 LLVMValueRef args[1] = {
1701 LLVMConstInt(ctx->i32, simm16, false),
1702 };
1703 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
1704 ctx->voidt, args, 1, 0);
1705 }
1706
1707 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
1708 unsigned bitsize)
1709 {
1710 LLVMTypeRef type;
1711 char *intr;
1712
1713 if (bitsize == 32) {
1714 intr = "llvm.floor.f32";
1715 type = ctx->f32;
1716 } else {
1717 intr = "llvm.floor.f64";
1718 type = ctx->f64;
1719 }
1720
1721 LLVMValueRef params[] = {
1722 src0,
1723 };
1724 LLVMValueRef floor = ac_build_intrinsic(ctx, intr, type, params, 1,
1725 AC_FUNC_ATTR_READNONE);
1726 return LLVMBuildFSub(ctx->builder, src0, floor, "");
1727 }
1728
1729 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
1730 unsigned bitsize)
1731 {
1732 LLVMValueRef cmp, val, zero, one;
1733 LLVMTypeRef type;
1734
1735 if (bitsize == 32) {
1736 type = ctx->i32;
1737 zero = ctx->i32_0;
1738 one = ctx->i32_1;
1739 } else {
1740 type = ctx->i64;
1741 zero = ctx->i64_0;
1742 one = ctx->i64_1;
1743 }
1744
1745 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
1746 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
1747 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
1748 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
1749 return val;
1750 }
1751
1752 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
1753 unsigned bitsize)
1754 {
1755 LLVMValueRef cmp, val, zero, one;
1756 LLVMTypeRef type;
1757
1758 if (bitsize == 32) {
1759 type = ctx->f32;
1760 zero = ctx->f32_0;
1761 one = ctx->f32_1;
1762 } else {
1763 type = ctx->f64;
1764 zero = ctx->f64_0;
1765 one = ctx->f64_1;
1766 }
1767
1768 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
1769 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
1770 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
1771 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
1772 return val;
1773 }
1774
1775 void ac_get_image_intr_name(const char *base_name,
1776 LLVMTypeRef data_type,
1777 LLVMTypeRef coords_type,
1778 LLVMTypeRef rsrc_type,
1779 char *out_name, unsigned out_len)
1780 {
1781 char coords_type_name[8];
1782
1783 ac_build_type_name_for_intr(coords_type, coords_type_name,
1784 sizeof(coords_type_name));
1785
1786 char data_type_name[8];
1787 char rsrc_type_name[8];
1788
1789 ac_build_type_name_for_intr(data_type, data_type_name,
1790 sizeof(data_type_name));
1791 ac_build_type_name_for_intr(rsrc_type, rsrc_type_name,
1792 sizeof(rsrc_type_name));
1793 snprintf(out_name, out_len, "%s.%s.%s.%s", base_name,
1794 data_type_name, coords_type_name, rsrc_type_name);
1795 }
1796
1797 #define AC_EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
1798 #define AC_EXP_ENABLED_CHANNELS (HAVE_LLVM >= 0x0500 ? 1 : 0)
1799 #define AC_EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
1800
1801 enum ac_ir_type {
1802 AC_IR_UNDEF,
1803 AC_IR_CONST,
1804 AC_IR_VALUE,
1805 };
1806
1807 struct ac_vs_exp_chan
1808 {
1809 LLVMValueRef value;
1810 float const_float;
1811 enum ac_ir_type type;
1812 };
1813
1814 struct ac_vs_exp_inst {
1815 unsigned offset;
1816 LLVMValueRef inst;
1817 struct ac_vs_exp_chan chan[4];
1818 };
1819
1820 struct ac_vs_exports {
1821 unsigned num;
1822 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
1823 };
1824
1825 /* Return true if the PARAM export has been eliminated. */
1826 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
1827 uint32_t num_outputs,
1828 struct ac_vs_exp_inst *exp)
1829 {
1830 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
1831 bool is_zero[4] = {}, is_one[4] = {};
1832
1833 for (i = 0; i < 4; i++) {
1834 /* It's a constant expression. Undef outputs are eliminated too. */
1835 if (exp->chan[i].type == AC_IR_UNDEF) {
1836 is_zero[i] = true;
1837 is_one[i] = true;
1838 } else if (exp->chan[i].type == AC_IR_CONST) {
1839 if (exp->chan[i].const_float == 0)
1840 is_zero[i] = true;
1841 else if (exp->chan[i].const_float == 1)
1842 is_one[i] = true;
1843 else
1844 return false; /* other constant */
1845 } else
1846 return false;
1847 }
1848
1849 /* Only certain combinations of 0 and 1 can be eliminated. */
1850 if (is_zero[0] && is_zero[1] && is_zero[2])
1851 default_val = is_zero[3] ? 0 : 1;
1852 else if (is_one[0] && is_one[1] && is_one[2])
1853 default_val = is_zero[3] ? 2 : 3;
1854 else
1855 return false;
1856
1857 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
1858 LLVMInstructionEraseFromParent(exp->inst);
1859
1860 /* Change OFFSET to DEFAULT_VAL. */
1861 for (i = 0; i < num_outputs; i++) {
1862 if (vs_output_param_offset[i] == exp->offset) {
1863 vs_output_param_offset[i] =
1864 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
1865 break;
1866 }
1867 }
1868 return true;
1869 }
1870
1871 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
1872 uint8_t *vs_output_param_offset,
1873 uint32_t num_outputs,
1874 struct ac_vs_exports *processed,
1875 struct ac_vs_exp_inst *exp)
1876 {
1877 unsigned p, copy_back_channels = 0;
1878
1879 /* See if the output is already in the list of processed outputs.
1880 * The LLVMValueRef comparison relies on SSA.
1881 */
1882 for (p = 0; p < processed->num; p++) {
1883 bool different = false;
1884
1885 for (unsigned j = 0; j < 4; j++) {
1886 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
1887 struct ac_vs_exp_chan *c2 = &exp->chan[j];
1888
1889 /* Treat undef as a match. */
1890 if (c2->type == AC_IR_UNDEF)
1891 continue;
1892
1893 /* If c1 is undef but c2 isn't, we can copy c2 to c1
1894 * and consider the instruction duplicated.
1895 */
1896 if (c1->type == AC_IR_UNDEF) {
1897 copy_back_channels |= 1 << j;
1898 continue;
1899 }
1900
1901 /* Test whether the channels are not equal. */
1902 if (c1->type != c2->type ||
1903 (c1->type == AC_IR_CONST &&
1904 c1->const_float != c2->const_float) ||
1905 (c1->type == AC_IR_VALUE &&
1906 c1->value != c2->value)) {
1907 different = true;
1908 break;
1909 }
1910 }
1911 if (!different)
1912 break;
1913
1914 copy_back_channels = 0;
1915 }
1916 if (p == processed->num)
1917 return false;
1918
1919 /* If a match was found, but the matching export has undef where the new
1920 * one has a normal value, copy the normal value to the undef channel.
1921 */
1922 struct ac_vs_exp_inst *match = &processed->exp[p];
1923
1924 /* Get current enabled channels mask. */
1925 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
1926 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
1927
1928 while (copy_back_channels) {
1929 unsigned chan = u_bit_scan(&copy_back_channels);
1930
1931 assert(match->chan[chan].type == AC_IR_UNDEF);
1932 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
1933 exp->chan[chan].value);
1934 match->chan[chan] = exp->chan[chan];
1935
1936 /* Update number of enabled channels because the original mask
1937 * is not always 0xf.
1938 */
1939 enabled_channels |= (1 << chan);
1940 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
1941 LLVMConstInt(ctx->i32, enabled_channels, 0));
1942 }
1943
1944 /* The PARAM export is duplicated. Kill it. */
1945 LLVMInstructionEraseFromParent(exp->inst);
1946
1947 /* Change OFFSET to the matching export. */
1948 for (unsigned i = 0; i < num_outputs; i++) {
1949 if (vs_output_param_offset[i] == exp->offset) {
1950 vs_output_param_offset[i] = match->offset;
1951 break;
1952 }
1953 }
1954 return true;
1955 }
1956
1957 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
1958 LLVMValueRef main_fn,
1959 uint8_t *vs_output_param_offset,
1960 uint32_t num_outputs,
1961 uint8_t *num_param_exports)
1962 {
1963 LLVMBasicBlockRef bb;
1964 bool removed_any = false;
1965 struct ac_vs_exports exports;
1966
1967 exports.num = 0;
1968
1969 /* Process all LLVM instructions. */
1970 bb = LLVMGetFirstBasicBlock(main_fn);
1971 while (bb) {
1972 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
1973
1974 while (inst) {
1975 LLVMValueRef cur = inst;
1976 inst = LLVMGetNextInstruction(inst);
1977 struct ac_vs_exp_inst exp;
1978
1979 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
1980 continue;
1981
1982 LLVMValueRef callee = ac_llvm_get_called_value(cur);
1983
1984 if (!ac_llvm_is_function(callee))
1985 continue;
1986
1987 const char *name = LLVMGetValueName(callee);
1988 unsigned num_args = LLVMCountParams(callee);
1989
1990 /* Check if this is an export instruction. */
1991 if ((num_args != 9 && num_args != 8) ||
1992 (strcmp(name, "llvm.SI.export") &&
1993 strcmp(name, "llvm.amdgcn.exp.f32")))
1994 continue;
1995
1996 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
1997 unsigned target = LLVMConstIntGetZExtValue(arg);
1998
1999 if (target < V_008DFC_SQ_EXP_PARAM)
2000 continue;
2001
2002 target -= V_008DFC_SQ_EXP_PARAM;
2003
2004 /* Parse the instruction. */
2005 memset(&exp, 0, sizeof(exp));
2006 exp.offset = target;
2007 exp.inst = cur;
2008
2009 for (unsigned i = 0; i < 4; i++) {
2010 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
2011
2012 exp.chan[i].value = v;
2013
2014 if (LLVMIsUndef(v)) {
2015 exp.chan[i].type = AC_IR_UNDEF;
2016 } else if (LLVMIsAConstantFP(v)) {
2017 LLVMBool loses_info;
2018 exp.chan[i].type = AC_IR_CONST;
2019 exp.chan[i].const_float =
2020 LLVMConstRealGetDouble(v, &loses_info);
2021 } else {
2022 exp.chan[i].type = AC_IR_VALUE;
2023 }
2024 }
2025
2026 /* Eliminate constant and duplicated PARAM exports. */
2027 if (ac_eliminate_const_output(vs_output_param_offset,
2028 num_outputs, &exp) ||
2029 ac_eliminate_duplicated_output(ctx,
2030 vs_output_param_offset,
2031 num_outputs, &exports,
2032 &exp)) {
2033 removed_any = true;
2034 } else {
2035 exports.exp[exports.num++] = exp;
2036 }
2037 }
2038 bb = LLVMGetNextBasicBlock(bb);
2039 }
2040
2041 /* Remove holes in export memory due to removed PARAM exports.
2042 * This is done by renumbering all PARAM exports.
2043 */
2044 if (removed_any) {
2045 uint8_t old_offset[VARYING_SLOT_MAX];
2046 unsigned out, i;
2047
2048 /* Make a copy of the offsets. We need the old version while
2049 * we are modifying some of them. */
2050 memcpy(old_offset, vs_output_param_offset,
2051 sizeof(old_offset));
2052
2053 for (i = 0; i < exports.num; i++) {
2054 unsigned offset = exports.exp[i].offset;
2055
2056 /* Update vs_output_param_offset. Multiple outputs can
2057 * have the same offset.
2058 */
2059 for (out = 0; out < num_outputs; out++) {
2060 if (old_offset[out] == offset)
2061 vs_output_param_offset[out] = i;
2062 }
2063
2064 /* Change the PARAM offset in the instruction. */
2065 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
2066 LLVMConstInt(ctx->i32,
2067 V_008DFC_SQ_EXP_PARAM + i, 0));
2068 }
2069 *num_param_exports = exports.num;
2070 }
2071 }
2072
2073 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
2074 {
2075 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
2076 ac_build_intrinsic(ctx,
2077 "llvm.amdgcn.init.exec", ctx->voidt,
2078 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
2079 }
2080
2081 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
2082 {
2083 unsigned lds_size = ctx->chip_class >= CIK ? 65536 : 32768;
2084 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
2085 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_LOCAL_ADDR_SPACE),
2086 "lds");
2087 }
2088
2089 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
2090 LLVMValueRef dw_addr)
2091 {
2092 return ac_build_load(ctx, ctx->lds, dw_addr);
2093 }
2094
2095 void ac_lds_store(struct ac_llvm_context *ctx,
2096 LLVMValueRef dw_addr,
2097 LLVMValueRef value)
2098 {
2099 value = ac_to_integer(ctx, value);
2100 ac_build_indexed_store(ctx, ctx->lds,
2101 dw_addr, value);
2102 }
2103
2104 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
2105 LLVMTypeRef dst_type,
2106 LLVMValueRef src0)
2107 {
2108 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2109 const char *intrin_name;
2110 LLVMTypeRef type;
2111 LLVMValueRef zero;
2112 if (src0_bitsize == 64) {
2113 intrin_name = "llvm.cttz.i64";
2114 type = ctx->i64;
2115 zero = ctx->i64_0;
2116 } else {
2117 intrin_name = "llvm.cttz.i32";
2118 type = ctx->i32;
2119 zero = ctx->i32_0;
2120 }
2121
2122 LLVMValueRef params[2] = {
2123 src0,
2124
2125 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
2126 * add special code to check for x=0. The reason is that
2127 * the LLVM behavior for x=0 is different from what we
2128 * need here. However, LLVM also assumes that ffs(x) is
2129 * in [0, 31], but GLSL expects that ffs(0) = -1, so
2130 * a conditional assignment to handle 0 is still required.
2131 *
2132 * The hardware already implements the correct behavior.
2133 */
2134 LLVMConstInt(ctx->i1, 1, false),
2135 };
2136
2137 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
2138 params, 2,
2139 AC_FUNC_ATTR_READNONE);
2140
2141 if (src0_bitsize == 64) {
2142 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
2143 }
2144
2145 /* TODO: We need an intrinsic to skip this conditional. */
2146 /* Check for zero: */
2147 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
2148 LLVMIntEQ, src0,
2149 zero, ""),
2150 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
2151 }
2152
2153 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
2154 {
2155 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2156 AC_CONST_ADDR_SPACE);
2157 }
2158
2159 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
2160 {
2161 if (!HAVE_32BIT_POINTERS)
2162 return ac_array_in_const_addr_space(elem_type);
2163
2164 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2165 AC_CONST_32BIT_ADDR_SPACE);
2166 }
2167
2168 static struct ac_llvm_flow *
2169 get_current_flow(struct ac_llvm_context *ctx)
2170 {
2171 if (ctx->flow_depth > 0)
2172 return &ctx->flow[ctx->flow_depth - 1];
2173 return NULL;
2174 }
2175
2176 static struct ac_llvm_flow *
2177 get_innermost_loop(struct ac_llvm_context *ctx)
2178 {
2179 for (unsigned i = ctx->flow_depth; i > 0; --i) {
2180 if (ctx->flow[i - 1].loop_entry_block)
2181 return &ctx->flow[i - 1];
2182 }
2183 return NULL;
2184 }
2185
2186 static struct ac_llvm_flow *
2187 push_flow(struct ac_llvm_context *ctx)
2188 {
2189 struct ac_llvm_flow *flow;
2190
2191 if (ctx->flow_depth >= ctx->flow_depth_max) {
2192 unsigned new_max = MAX2(ctx->flow_depth << 1,
2193 AC_LLVM_INITIAL_CF_DEPTH);
2194
2195 ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
2196 ctx->flow_depth_max = new_max;
2197 }
2198
2199 flow = &ctx->flow[ctx->flow_depth];
2200 ctx->flow_depth++;
2201
2202 flow->next_block = NULL;
2203 flow->loop_entry_block = NULL;
2204 return flow;
2205 }
2206
2207 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
2208 int label_id)
2209 {
2210 char buf[32];
2211 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
2212 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
2213 }
2214
2215 /* Append a basic block at the level of the parent flow.
2216 */
2217 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
2218 const char *name)
2219 {
2220 assert(ctx->flow_depth >= 1);
2221
2222 if (ctx->flow_depth >= 2) {
2223 struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
2224
2225 return LLVMInsertBasicBlockInContext(ctx->context,
2226 flow->next_block, name);
2227 }
2228
2229 LLVMValueRef main_fn =
2230 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
2231 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
2232 }
2233
2234 /* Emit a branch to the given default target for the current block if
2235 * applicable -- that is, if the current block does not already contain a
2236 * branch from a break or continue.
2237 */
2238 static void emit_default_branch(LLVMBuilderRef builder,
2239 LLVMBasicBlockRef target)
2240 {
2241 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
2242 LLVMBuildBr(builder, target);
2243 }
2244
2245 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
2246 {
2247 struct ac_llvm_flow *flow = push_flow(ctx);
2248 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
2249 flow->next_block = append_basic_block(ctx, "ENDLOOP");
2250 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
2251 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2252 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
2253 }
2254
2255 void ac_build_break(struct ac_llvm_context *ctx)
2256 {
2257 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2258 LLVMBuildBr(ctx->builder, flow->next_block);
2259 }
2260
2261 void ac_build_continue(struct ac_llvm_context *ctx)
2262 {
2263 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2264 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2265 }
2266
2267 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
2268 {
2269 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2270 LLVMBasicBlockRef endif_block;
2271
2272 assert(!current_branch->loop_entry_block);
2273
2274 endif_block = append_basic_block(ctx, "ENDIF");
2275 emit_default_branch(ctx->builder, endif_block);
2276
2277 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2278 set_basicblock_name(current_branch->next_block, "else", label_id);
2279
2280 current_branch->next_block = endif_block;
2281 }
2282
2283 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
2284 {
2285 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2286
2287 assert(!current_branch->loop_entry_block);
2288
2289 emit_default_branch(ctx->builder, current_branch->next_block);
2290 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2291 set_basicblock_name(current_branch->next_block, "endif", label_id);
2292
2293 ctx->flow_depth--;
2294 }
2295
2296 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
2297 {
2298 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
2299
2300 assert(current_loop->loop_entry_block);
2301
2302 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
2303
2304 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
2305 set_basicblock_name(current_loop->next_block, "endloop", label_id);
2306 ctx->flow_depth--;
2307 }
2308
2309 static void if_cond_emit(struct ac_llvm_context *ctx, LLVMValueRef cond,
2310 int label_id)
2311 {
2312 struct ac_llvm_flow *flow = push_flow(ctx);
2313 LLVMBasicBlockRef if_block;
2314
2315 if_block = append_basic_block(ctx, "IF");
2316 flow->next_block = append_basic_block(ctx, "ELSE");
2317 set_basicblock_name(if_block, "if", label_id);
2318 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
2319 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
2320 }
2321
2322 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
2323 int label_id)
2324 {
2325 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
2326 value, ctx->f32_0, "");
2327 if_cond_emit(ctx, cond, label_id);
2328 }
2329
2330 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
2331 int label_id)
2332 {
2333 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
2334 ac_to_integer(ctx, value),
2335 ctx->i32_0, "");
2336 if_cond_emit(ctx, cond, label_id);
2337 }