projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
amd/common: use the dimension-aware image intrinsics on LLVM 7+
[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 "util/u_math.h"
41 #include "sid.h"
42
43 #include "shader_enums.h"
44
45 #define AC_LLVM_INITIAL_CF_DEPTH 4
46
47 /* Data for if/else/endif and bgnloop/endloop control flow structures.
48 */
49 struct ac_llvm_flow {
50 /* Loop exit or next part of if/else/endif. */
51 LLVMBasicBlockRef next_block;
52 LLVMBasicBlockRef loop_entry_block;
53 };
54
55 /* Initialize module-independent parts of the context.
56 *
57 * The caller is responsible for initializing ctx::module and ctx::builder.
58 */
59 void
60 ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context,
61 enum chip_class chip_class, enum radeon_family family)
62 {
63 LLVMValueRef args[1];
64
65 ctx->chip_class = chip_class;
66 ctx->family = family;
67
68 ctx->context = context;
69 ctx->module = NULL;
70 ctx->builder = NULL;
71
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 = HAVE_32BIT_POINTERS ? ctx->i32 : ctx->i64;
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);
89
90 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
91 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
92 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
93 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
94 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
95 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
96 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
97 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
98
99 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
100 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
101
102 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
103 "range", 5);
104
105 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
106 "invariant.load", 14);
107
108 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
109
110 args[0] = LLVMConstReal(ctx->f32, 2.5);
111 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
112
113 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
114 "amdgpu.uniform", 14);
115
116 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
117 }
118
119 void
120 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
121 {
122 free(ctx->flow);
123 ctx->flow = NULL;
124 ctx->flow_depth_max = 0;
125 }
126
127 int
128 ac_get_llvm_num_components(LLVMValueRef value)
129 {
130 LLVMTypeRef type = LLVMTypeOf(value);
131 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
132 ? LLVMGetVectorSize(type)
133 : 1;
134 return num_components;
135 }
136
137 LLVMValueRef
138 ac_llvm_extract_elem(struct ac_llvm_context *ac,
139 LLVMValueRef value,
140 int index)
141 {
142 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
143 assert(index == 0);
144 return value;
145 }
146
147 return LLVMBuildExtractElement(ac->builder, value,
148 LLVMConstInt(ac->i32, index, false), "");
149 }
150
151 int
152 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
153 {
154 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
155 type = LLVMGetElementType(type);
156
157 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
158 return LLVMGetIntTypeWidth(type);
159
160 if (type == ctx->f16)
161 return 16;
162 if (type == ctx->f32)
163 return 32;
164 if (type == ctx->f64)
165 return 64;
166
167 unreachable("Unhandled type kind in get_elem_bits");
168 }
169
170 unsigned
171 ac_get_type_size(LLVMTypeRef type)
172 {
173 LLVMTypeKind kind = LLVMGetTypeKind(type);
174
175 switch (kind) {
176 case LLVMIntegerTypeKind:
177 return LLVMGetIntTypeWidth(type) / 8;
178 case LLVMFloatTypeKind:
179 return 4;
180 case LLVMDoubleTypeKind:
181 return 8;
182 case LLVMPointerTypeKind:
183 if (LLVMGetPointerAddressSpace(type) == AC_CONST_32BIT_ADDR_SPACE)
184 return 4;
185 return 8;
186 case LLVMVectorTypeKind:
187 return LLVMGetVectorSize(type) *
188 ac_get_type_size(LLVMGetElementType(type));
189 case LLVMArrayTypeKind:
190 return LLVMGetArrayLength(type) *
191 ac_get_type_size(LLVMGetElementType(type));
192 default:
193 assert(0);
194 return 0;
195 }
196 }
197
198 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
199 {
200 if (t == ctx->f16 || t == ctx->i16)
201 return ctx->i16;
202 else if (t == ctx->f32 || t == ctx->i32)
203 return ctx->i32;
204 else if (t == ctx->f64 || t == ctx->i64)
205 return ctx->i64;
206 else
207 unreachable("Unhandled integer size");
208 }
209
210 LLVMTypeRef
211 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
212 {
213 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
214 LLVMTypeRef elem_type = LLVMGetElementType(t);
215 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
216 LLVMGetVectorSize(t));
217 }
218 return to_integer_type_scalar(ctx, t);
219 }
220
221 LLVMValueRef
222 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
223 {
224 LLVMTypeRef type = LLVMTypeOf(v);
225 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
226 }
227
228 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
229 {
230 if (t == ctx->i16 || t == ctx->f16)
231 return ctx->f16;
232 else if (t == ctx->i32 || t == ctx->f32)
233 return ctx->f32;
234 else if (t == ctx->i64 || t == ctx->f64)
235 return ctx->f64;
236 else
237 unreachable("Unhandled float size");
238 }
239
240 LLVMTypeRef
241 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
242 {
243 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
244 LLVMTypeRef elem_type = LLVMGetElementType(t);
245 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
246 LLVMGetVectorSize(t));
247 }
248 return to_float_type_scalar(ctx, t);
249 }
250
251 LLVMValueRef
252 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
253 {
254 LLVMTypeRef type = LLVMTypeOf(v);
255 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
256 }
257
258
259 LLVMValueRef
260 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
261 LLVMTypeRef return_type, LLVMValueRef *params,
262 unsigned param_count, unsigned attrib_mask)
263 {
264 LLVMValueRef function, call;
265 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
266
267 function = LLVMGetNamedFunction(ctx->module, name);
268 if (!function) {
269 LLVMTypeRef param_types[32], function_type;
270 unsigned i;
271
272 assert(param_count <= 32);
273
274 for (i = 0; i < param_count; ++i) {
275 assert(params[i]);
276 param_types[i] = LLVMTypeOf(params[i]);
277 }
278 function_type =
279 LLVMFunctionType(return_type, param_types, param_count, 0);
280 function = LLVMAddFunction(ctx->module, name, function_type);
281
282 LLVMSetFunctionCallConv(function, LLVMCCallConv);
283 LLVMSetLinkage(function, LLVMExternalLinkage);
284
285 if (!set_callsite_attrs)
286 ac_add_func_attributes(ctx->context, function, attrib_mask);
287 }
288
289 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
290 if (set_callsite_attrs)
291 ac_add_func_attributes(ctx->context, call, attrib_mask);
292 return call;
293 }
294
295 /**
296 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
297 * intrinsic names).
298 */
299 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
300 {
301 LLVMTypeRef elem_type = type;
302
303 assert(bufsize >= 8);
304
305 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
306 int ret = snprintf(buf, bufsize, "v%u",
307 LLVMGetVectorSize(type));
308 if (ret < 0) {
309 char *type_name = LLVMPrintTypeToString(type);
310 fprintf(stderr, "Error building type name for: %s\n",
311 type_name);
312 return;
313 }
314 elem_type = LLVMGetElementType(type);
315 buf += ret;
316 bufsize -= ret;
317 }
318 switch (LLVMGetTypeKind(elem_type)) {
319 default: break;
320 case LLVMIntegerTypeKind:
321 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
322 break;
323 case LLVMFloatTypeKind:
324 snprintf(buf, bufsize, "f32");
325 break;
326 case LLVMDoubleTypeKind:
327 snprintf(buf, bufsize, "f64");
328 break;
329 }
330 }
331
332 /**
333 * Helper function that builds an LLVM IR PHI node and immediately adds
334 * incoming edges.
335 */
336 LLVMValueRef
337 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
338 unsigned count_incoming, LLVMValueRef *values,
339 LLVMBasicBlockRef *blocks)
340 {
341 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
342 LLVMAddIncoming(phi, values, blocks, count_incoming);
343 return phi;
344 }
345
346 /* Prevent optimizations (at least of memory accesses) across the current
347 * point in the program by emitting empty inline assembly that is marked as
348 * having side effects.
349 *
350 * Optionally, a value can be passed through the inline assembly to prevent
351 * LLVM from hoisting calls to ReadNone functions.
352 */
353 void
354 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
355 LLVMValueRef *pvgpr)
356 {
357 static int counter = 0;
358
359 LLVMBuilderRef builder = ctx->builder;
360 char code[16];
361
362 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
363
364 if (!pvgpr) {
365 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
366 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
367 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
368 } else {
369 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
370 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
371 LLVMValueRef vgpr = *pvgpr;
372 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
373 unsigned vgpr_size = ac_get_type_size(vgpr_type);
374 LLVMValueRef vgpr0;
375
376 assert(vgpr_size % 4 == 0);
377
378 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
379 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
380 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
381 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
382 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
383
384 *pvgpr = vgpr;
385 }
386 }
387
388 LLVMValueRef
389 ac_build_shader_clock(struct ac_llvm_context *ctx)
390 {
391 LLVMValueRef tmp = ac_build_intrinsic(ctx, "llvm.readcyclecounter",
392 ctx->i64, NULL, 0, 0);
393 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
394 }
395
396 LLVMValueRef
397 ac_build_ballot(struct ac_llvm_context *ctx,
398 LLVMValueRef value)
399 {
400 LLVMValueRef args[3] = {
401 value,
402 ctx->i32_0,
403 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
404 };
405
406 /* We currently have no other way to prevent LLVM from lifting the icmp
407 * calls to a dominating basic block.
408 */
409 ac_build_optimization_barrier(ctx, &args[0]);
410
411 args[0] = ac_to_integer(ctx, args[0]);
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 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
892 * intrinsics. */
893 if (num_channels == 3) {
894 LLVMValueRef v[3], v01;
895
896 for (int i = 0; i < 3; i++) {
897 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
898 LLVMConstInt(ctx->i32, i, 0), "");
899 }
900 v01 = ac_build_gather_values(ctx, v, 2);
901
902 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
903 soffset, inst_offset, glc, slc,
904 writeonly_memory, swizzle_enable_hint);
905 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
906 soffset, inst_offset + 8,
907 glc, slc,
908 writeonly_memory, swizzle_enable_hint);
909 return;
910 }
911
912 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
913 * (voffset is swizzled, but soffset isn't swizzled).
914 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
915 */
916 if (!swizzle_enable_hint) {
917 LLVMValueRef offset = soffset;
918
919 static const char *types[] = {"f32", "v2f32", "v4f32"};
920
921 if (inst_offset)
922 offset = LLVMBuildAdd(ctx->builder, offset,
923 LLVMConstInt(ctx->i32, inst_offset, 0), "");
924 if (voffset)
925 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
926
927 LLVMValueRef args[] = {
928 ac_to_float(ctx, vdata),
929 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
930 LLVMConstInt(ctx->i32, 0, 0),
931 offset,
932 LLVMConstInt(ctx->i1, glc, 0),
933 LLVMConstInt(ctx->i1, slc, 0),
934 };
935
936 char name[256];
937 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
938 types[CLAMP(num_channels, 1, 3) - 1]);
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 const 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 static const char *types[] = {"i32", "v2i32", "v4i32"};
955 LLVMValueRef args[] = {
956 vdata,
957 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
958 LLVMConstInt(ctx->i32, 0, 0),
959 voffset ? voffset : LLVMConstInt(ctx->i32, 0, 0),
960 soffset,
961 LLVMConstInt(ctx->i32, inst_offset, 0),
962 LLVMConstInt(ctx->i32, dfmt[num_channels - 1], 0),
963 LLVMConstInt(ctx->i32, V_008F0C_BUF_NUM_FORMAT_UINT, 0),
964 LLVMConstInt(ctx->i1, glc, 0),
965 LLVMConstInt(ctx->i1, slc, 0),
966 };
967 char name[256];
968 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.store.%s",
969 types[CLAMP(num_channels, 1, 3) - 1]);
970
971 ac_build_intrinsic(ctx, name, ctx->voidt,
972 args, ARRAY_SIZE(args),
973 writeonly_memory ?
974 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
975 AC_FUNC_ATTR_WRITEONLY);
976 }
977
978 static LLVMValueRef
979 ac_build_buffer_load_common(struct ac_llvm_context *ctx,
980 LLVMValueRef rsrc,
981 LLVMValueRef vindex,
982 LLVMValueRef voffset,
983 unsigned num_channels,
984 bool glc,
985 bool slc,
986 bool can_speculate,
987 bool use_format)
988 {
989 LLVMValueRef args[] = {
990 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
991 vindex ? vindex : LLVMConstInt(ctx->i32, 0, 0),
992 voffset,
993 LLVMConstInt(ctx->i1, glc, 0),
994 LLVMConstInt(ctx->i1, slc, 0)
995 };
996 unsigned func = CLAMP(num_channels, 1, 3) - 1;
997
998 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
999 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1000 char name[256];
1001
1002 if (use_format) {
1003 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
1004 type_names[func]);
1005 } else {
1006 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1007 type_names[func]);
1008 }
1009
1010 return ac_build_intrinsic(ctx, name, types[func], args,
1011 ARRAY_SIZE(args),
1012 ac_get_load_intr_attribs(can_speculate));
1013 }
1014
1015 LLVMValueRef
1016 ac_build_buffer_load(struct ac_llvm_context *ctx,
1017 LLVMValueRef rsrc,
1018 int num_channels,
1019 LLVMValueRef vindex,
1020 LLVMValueRef voffset,
1021 LLVMValueRef soffset,
1022 unsigned inst_offset,
1023 unsigned glc,
1024 unsigned slc,
1025 bool can_speculate,
1026 bool allow_smem)
1027 {
1028 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1029 if (voffset)
1030 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1031 if (soffset)
1032 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1033
1034 /* TODO: VI and later generations can use SMEM with GLC=1.*/
1035 if (allow_smem && !glc && !slc) {
1036 assert(vindex == NULL);
1037
1038 LLVMValueRef result[8];
1039
1040 for (int i = 0; i < num_channels; i++) {
1041 if (i) {
1042 offset = LLVMBuildAdd(ctx->builder, offset,
1043 LLVMConstInt(ctx->i32, 4, 0), "");
1044 }
1045 LLVMValueRef args[2] = {rsrc, offset};
1046 result[i] = ac_build_intrinsic(ctx, "llvm.SI.load.const.v4i32",
1047 ctx->f32, args, 2,
1048 AC_FUNC_ATTR_READNONE |
1049 AC_FUNC_ATTR_LEGACY);
1050 }
1051 if (num_channels == 1)
1052 return result[0];
1053
1054 if (num_channels == 3)
1055 result[num_channels++] = LLVMGetUndef(ctx->f32);
1056 return ac_build_gather_values(ctx, result, num_channels);
1057 }
1058
1059 return ac_build_buffer_load_common(ctx, rsrc, vindex, offset,
1060 num_channels, glc, slc,
1061 can_speculate, false);
1062 }
1063
1064 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1065 LLVMValueRef rsrc,
1066 LLVMValueRef vindex,
1067 LLVMValueRef voffset,
1068 unsigned num_channels,
1069 bool glc,
1070 bool can_speculate)
1071 {
1072 return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1073 num_channels, glc, false,
1074 can_speculate, true);
1075 }
1076
1077 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
1078 LLVMValueRef rsrc,
1079 LLVMValueRef vindex,
1080 LLVMValueRef voffset,
1081 unsigned num_channels,
1082 bool glc,
1083 bool can_speculate)
1084 {
1085 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
1086 LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 1, 0), "");
1087 stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
1088
1089 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
1090 LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
1091 elem_count, stride, "");
1092
1093 LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
1094 LLVMConstInt(ctx->i32, 2, 0), "");
1095
1096 return ac_build_buffer_load_common(ctx, new_rsrc, vindex, voffset,
1097 num_channels, glc, false,
1098 can_speculate, true);
1099 }
1100
1101 /**
1102 * Set range metadata on an instruction. This can only be used on load and
1103 * call instructions. If you know an instruction can only produce the values
1104 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1105 * \p lo is the minimum value inclusive.
1106 * \p hi is the maximum value exclusive.
1107 */
1108 static void set_range_metadata(struct ac_llvm_context *ctx,
1109 LLVMValueRef value, unsigned lo, unsigned hi)
1110 {
1111 LLVMValueRef range_md, md_args[2];
1112 LLVMTypeRef type = LLVMTypeOf(value);
1113 LLVMContextRef context = LLVMGetTypeContext(type);
1114
1115 md_args[0] = LLVMConstInt(type, lo, false);
1116 md_args[1] = LLVMConstInt(type, hi, false);
1117 range_md = LLVMMDNodeInContext(context, md_args, 2);
1118 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1119 }
1120
1121 LLVMValueRef
1122 ac_get_thread_id(struct ac_llvm_context *ctx)
1123 {
1124 LLVMValueRef tid;
1125
1126 LLVMValueRef tid_args[2];
1127 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
1128 tid_args[1] = LLVMConstInt(ctx->i32, 0, false);
1129 tid_args[1] = ac_build_intrinsic(ctx,
1130 "llvm.amdgcn.mbcnt.lo", ctx->i32,
1131 tid_args, 2, AC_FUNC_ATTR_READNONE);
1132
1133 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
1134 ctx->i32, tid_args,
1135 2, AC_FUNC_ATTR_READNONE);
1136 set_range_metadata(ctx, tid, 0, 64);
1137 return tid;
1138 }
1139
1140 /*
1141 * SI implements derivatives using the local data store (LDS)
1142 * All writes to the LDS happen in all executing threads at
1143 * the same time. TID is the Thread ID for the current
1144 * thread and is a value between 0 and 63, representing
1145 * the thread's position in the wavefront.
1146 *
1147 * For the pixel shader threads are grouped into quads of four pixels.
1148 * The TIDs of the pixels of a quad are:
1149 *
1150 * +------+------+
1151 * |4n + 0|4n + 1|
1152 * +------+------+
1153 * |4n + 2|4n + 3|
1154 * +------+------+
1155 *
1156 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1157 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1158 * the current pixel's column, and masking with 0xfffffffe yields the TID
1159 * of the left pixel of the current pixel's row.
1160 *
1161 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1162 * adding 2 yields the TID of the pixel below the top pixel.
1163 */
1164 LLVMValueRef
1165 ac_build_ddxy(struct ac_llvm_context *ctx,
1166 uint32_t mask,
1167 int idx,
1168 LLVMValueRef val)
1169 {
1170 LLVMValueRef tl, trbl, args[2];
1171 LLVMValueRef result;
1172
1173 if (HAVE_LLVM >= 0x0700) {
1174 unsigned tl_lanes[4], trbl_lanes[4];
1175
1176 for (unsigned i = 0; i < 4; ++i) {
1177 tl_lanes[i] = i & mask;
1178 trbl_lanes[i] = (i & mask) + idx;
1179 }
1180
1181 tl = ac_build_quad_swizzle(ctx, val,
1182 tl_lanes[0], tl_lanes[1],
1183 tl_lanes[2], tl_lanes[3]);
1184 trbl = ac_build_quad_swizzle(ctx, val,
1185 trbl_lanes[0], trbl_lanes[1],
1186 trbl_lanes[2], trbl_lanes[3]);
1187 } else if (ctx->chip_class >= VI) {
1188 LLVMValueRef thread_id, tl_tid, trbl_tid;
1189 thread_id = ac_get_thread_id(ctx);
1190
1191 tl_tid = LLVMBuildAnd(ctx->builder, thread_id,
1192 LLVMConstInt(ctx->i32, mask, false), "");
1193
1194 trbl_tid = LLVMBuildAdd(ctx->builder, tl_tid,
1195 LLVMConstInt(ctx->i32, idx, false), "");
1196
1197 args[0] = LLVMBuildMul(ctx->builder, tl_tid,
1198 LLVMConstInt(ctx->i32, 4, false), "");
1199 args[1] = val;
1200 tl = ac_build_intrinsic(ctx,
1201 "llvm.amdgcn.ds.bpermute", ctx->i32,
1202 args, 2,
1203 AC_FUNC_ATTR_READNONE |
1204 AC_FUNC_ATTR_CONVERGENT);
1205
1206 args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
1207 LLVMConstInt(ctx->i32, 4, false), "");
1208 trbl = ac_build_intrinsic(ctx,
1209 "llvm.amdgcn.ds.bpermute", ctx->i32,
1210 args, 2,
1211 AC_FUNC_ATTR_READNONE |
1212 AC_FUNC_ATTR_CONVERGENT);
1213 } else {
1214 uint32_t masks[2] = {};
1215
1216 switch (mask) {
1217 case AC_TID_MASK_TOP_LEFT:
1218 masks[0] = 0x8000;
1219 if (idx == 1)
1220 masks[1] = 0x8055;
1221 else
1222 masks[1] = 0x80aa;
1223
1224 break;
1225 case AC_TID_MASK_TOP:
1226 masks[0] = 0x8044;
1227 masks[1] = 0x80ee;
1228 break;
1229 case AC_TID_MASK_LEFT:
1230 masks[0] = 0x80a0;
1231 masks[1] = 0x80f5;
1232 break;
1233 default:
1234 assert(0);
1235 }
1236
1237 args[0] = val;
1238 args[1] = LLVMConstInt(ctx->i32, masks[0], false);
1239
1240 tl = ac_build_intrinsic(ctx,
1241 "llvm.amdgcn.ds.swizzle", ctx->i32,
1242 args, 2,
1243 AC_FUNC_ATTR_READNONE |
1244 AC_FUNC_ATTR_CONVERGENT);
1245
1246 args[1] = LLVMConstInt(ctx->i32, masks[1], false);
1247 trbl = ac_build_intrinsic(ctx,
1248 "llvm.amdgcn.ds.swizzle", ctx->i32,
1249 args, 2,
1250 AC_FUNC_ATTR_READNONE |
1251 AC_FUNC_ATTR_CONVERGENT);
1252 }
1253
1254 tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
1255 trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
1256 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
1257
1258 if (HAVE_LLVM >= 0x0700) {
1259 result = ac_build_intrinsic(ctx,
1260 "llvm.amdgcn.wqm.f32", ctx->f32,
1261 &result, 1, 0);
1262 }
1263
1264 return result;
1265 }
1266
1267 void
1268 ac_build_sendmsg(struct ac_llvm_context *ctx,
1269 uint32_t msg,
1270 LLVMValueRef wave_id)
1271 {
1272 LLVMValueRef args[2];
1273 args[0] = LLVMConstInt(ctx->i32, msg, false);
1274 args[1] = wave_id;
1275 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
1276 }
1277
1278 LLVMValueRef
1279 ac_build_imsb(struct ac_llvm_context *ctx,
1280 LLVMValueRef arg,
1281 LLVMTypeRef dst_type)
1282 {
1283 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
1284 dst_type, &arg, 1,
1285 AC_FUNC_ATTR_READNONE);
1286
1287 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1288 * the index from LSB. Invert it by doing "31 - msb". */
1289 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1290 msb, "");
1291
1292 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
1293 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
1294 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1295 arg, LLVMConstInt(ctx->i32, 0, 0), ""),
1296 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1297 arg, all_ones, ""), "");
1298
1299 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
1300 }
1301
1302 LLVMValueRef
1303 ac_build_umsb(struct ac_llvm_context *ctx,
1304 LLVMValueRef arg,
1305 LLVMTypeRef dst_type)
1306 {
1307 const char *intrin_name;
1308 LLVMTypeRef type;
1309 LLVMValueRef highest_bit;
1310 LLVMValueRef zero;
1311
1312 if (ac_get_elem_bits(ctx, LLVMTypeOf(arg)) == 64) {
1313 intrin_name = "llvm.ctlz.i64";
1314 type = ctx->i64;
1315 highest_bit = LLVMConstInt(ctx->i64, 63, false);
1316 zero = ctx->i64_0;
1317 } else {
1318 intrin_name = "llvm.ctlz.i32";
1319 type = ctx->i32;
1320 highest_bit = LLVMConstInt(ctx->i32, 31, false);
1321 zero = ctx->i32_0;
1322 }
1323
1324 LLVMValueRef params[2] = {
1325 arg,
1326 ctx->i1true,
1327 };
1328
1329 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
1330 params, 2,
1331 AC_FUNC_ATTR_READNONE);
1332
1333 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1334 * the index from LSB. Invert it by doing "31 - msb". */
1335 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
1336 msb = LLVMBuildTruncOrBitCast(ctx->builder, msb, ctx->i32, "");
1337
1338 /* check for zero */
1339 return LLVMBuildSelect(ctx->builder,
1340 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
1341 LLVMConstInt(ctx->i32, -1, true), msb, "");
1342 }
1343
1344 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
1345 LLVMValueRef b)
1346 {
1347 LLVMValueRef args[2] = {a, b};
1348 return ac_build_intrinsic(ctx, "llvm.minnum.f32", ctx->f32, args, 2,
1349 AC_FUNC_ATTR_READNONE);
1350 }
1351
1352 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
1353 LLVMValueRef b)
1354 {
1355 LLVMValueRef args[2] = {a, b};
1356 return ac_build_intrinsic(ctx, "llvm.maxnum.f32", ctx->f32, args, 2,
1357 AC_FUNC_ATTR_READNONE);
1358 }
1359
1360 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
1361 LLVMValueRef b)
1362 {
1363 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
1364 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1365 }
1366
1367 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
1368 LLVMValueRef b)
1369 {
1370 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
1371 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1372 }
1373
1374 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
1375 LLVMValueRef b)
1376 {
1377 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
1378 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1379 }
1380
1381 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
1382 {
1383 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, ctx->f32_0),
1384 ctx->f32_1);
1385 }
1386
1387 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
1388 {
1389 LLVMValueRef args[9];
1390
1391 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
1392 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1393
1394 if (a->compr) {
1395 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
1396 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
1397
1398 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
1399 v2i16, "");
1400 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
1401 v2i16, "");
1402 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
1403 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1404
1405 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
1406 ctx->voidt, args, 6, 0);
1407 } else {
1408 args[2] = a->out[0];
1409 args[3] = a->out[1];
1410 args[4] = a->out[2];
1411 args[5] = a->out[3];
1412 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
1413 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1414
1415 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
1416 ctx->voidt, args, 8, 0);
1417 }
1418 }
1419
1420 void ac_build_export_null(struct ac_llvm_context *ctx)
1421 {
1422 struct ac_export_args args;
1423
1424 args.enabled_channels = 0x0; /* enabled channels */
1425 args.valid_mask = 1; /* whether the EXEC mask is valid */
1426 args.done = 1; /* DONE bit */
1427 args.target = V_008DFC_SQ_EXP_NULL;
1428 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
1429 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
1430 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
1431 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
1432 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
1433
1434 ac_build_export(ctx, &args);
1435 }
1436
1437 static unsigned ac_num_coords(enum ac_image_dim dim)
1438 {
1439 switch (dim) {
1440 case ac_image_1d:
1441 return 1;
1442 case ac_image_2d:
1443 case ac_image_1darray:
1444 return 2;
1445 case ac_image_3d:
1446 case ac_image_cube:
1447 case ac_image_2darray:
1448 case ac_image_2dmsaa:
1449 return 3;
1450 case ac_image_2darraymsaa:
1451 return 4;
1452 default:
1453 unreachable("ac_num_coords: bad dim");
1454 }
1455 }
1456
1457 static unsigned ac_num_derivs(enum ac_image_dim dim)
1458 {
1459 switch (dim) {
1460 case ac_image_1d:
1461 case ac_image_1darray:
1462 return 2;
1463 case ac_image_2d:
1464 case ac_image_2darray:
1465 case ac_image_cube:
1466 return 4;
1467 case ac_image_3d:
1468 return 6;
1469 case ac_image_2dmsaa:
1470 case ac_image_2darraymsaa:
1471 default:
1472 unreachable("derivatives not supported");
1473 }
1474 }
1475
1476 static const char *get_atomic_name(enum ac_atomic_op op)
1477 {
1478 switch (op) {
1479 case ac_atomic_swap: return "swap";
1480 case ac_atomic_add: return "add";
1481 case ac_atomic_sub: return "sub";
1482 case ac_atomic_smin: return "smin";
1483 case ac_atomic_umin: return "umin";
1484 case ac_atomic_smax: return "smax";
1485 case ac_atomic_umax: return "umax";
1486 case ac_atomic_and: return "and";
1487 case ac_atomic_or: return "or";
1488 case ac_atomic_xor: return "xor";
1489 }
1490 unreachable("bad atomic op");
1491 }
1492
1493 /* LLVM 6 and older */
1494 static LLVMValueRef ac_build_image_opcode_llvm6(struct ac_llvm_context *ctx,
1495 struct ac_image_args *a)
1496 {
1497 LLVMValueRef args[16];
1498 LLVMTypeRef retty = ctx->v4f32;
1499 const char *name = NULL;
1500 const char *atomic_subop = "";
1501 char intr_name[128], coords_type[64];
1502
1503 bool sample = a->opcode == ac_image_sample ||
1504 a->opcode == ac_image_gather4 ||
1505 a->opcode == ac_image_get_lod;
1506 bool atomic = a->opcode == ac_image_atomic ||
1507 a->opcode == ac_image_atomic_cmpswap;
1508 bool da = a->dim == ac_image_cube ||
1509 a->dim == ac_image_1darray ||
1510 a->dim == ac_image_2darray ||
1511 a->dim == ac_image_2darraymsaa;
1512 if (a->opcode == ac_image_get_lod)
1513 da = false;
1514
1515 unsigned num_coords =
1516 a->opcode != ac_image_get_resinfo ? ac_num_coords(a->dim) : 0;
1517 LLVMValueRef addr;
1518 unsigned num_addr = 0;
1519
1520 if (a->opcode == ac_image_get_lod) {
1521 switch (a->dim) {
1522 case ac_image_1darray:
1523 num_coords = 1;
1524 break;
1525 case ac_image_2darray:
1526 case ac_image_cube:
1527 num_coords = 2;
1528 break;
1529 default:
1530 break;
1531 }
1532 }
1533
1534 if (a->offset)
1535 args[num_addr++] = ac_to_integer(ctx, a->offset);
1536 if (a->bias)
1537 args[num_addr++] = ac_to_integer(ctx, a->bias);
1538 if (a->compare)
1539 args[num_addr++] = ac_to_integer(ctx, a->compare);
1540 if (a->derivs[0]) {
1541 unsigned num_derivs = ac_num_derivs(a->dim);
1542 for (unsigned i = 0; i < num_derivs; ++i)
1543 args[num_addr++] = ac_to_integer(ctx, a->derivs[i]);
1544 }
1545 for (unsigned i = 0; i < num_coords; ++i)
1546 args[num_addr++] = ac_to_integer(ctx, a->coords[i]);
1547 if (a->lod)
1548 args[num_addr++] = ac_to_integer(ctx, a->lod);
1549
1550 unsigned pad_goal = util_next_power_of_two(num_addr);
1551 while (num_addr < pad_goal)
1552 args[num_addr++] = LLVMGetUndef(ctx->i32);
1553
1554 addr = ac_build_gather_values(ctx, args, num_addr);
1555
1556 unsigned num_args = 0;
1557 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
1558 args[num_args++] = a->data[0];
1559 if (a->opcode == ac_image_atomic_cmpswap)
1560 args[num_args++] = a->data[1];
1561 }
1562
1563 unsigned coords_arg = num_args;
1564 if (sample)
1565 args[num_args++] = ac_to_float(ctx, addr);
1566 else
1567 args[num_args++] = ac_to_integer(ctx, addr);
1568
1569 args[num_args++] = a->resource;
1570 if (sample)
1571 args[num_args++] = a->sampler;
1572 if (!atomic) {
1573 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, 0);
1574 if (sample)
1575 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, 0);
1576 args[num_args++] = a->cache_policy & ac_glc ? ctx->i1true : ctx->i1false;
1577 args[num_args++] = a->cache_policy & ac_slc ? ctx->i1true : ctx->i1false;
1578 args[num_args++] = ctx->i1false; /* lwe */
1579 args[num_args++] = LLVMConstInt(ctx->i1, da, 0);
1580 } else {
1581 args[num_args++] = ctx->i1false; /* r128 */
1582 args[num_args++] = LLVMConstInt(ctx->i1, da, 0);
1583 args[num_args++] = a->cache_policy & ac_slc ? ctx->i1true : ctx->i1false;
1584 }
1585
1586 switch (a->opcode) {
1587 case ac_image_sample:
1588 name = "llvm.amdgcn.image.sample";
1589 break;
1590 case ac_image_gather4:
1591 name = "llvm.amdgcn.image.gather4";
1592 break;
1593 case ac_image_load:
1594 name = "llvm.amdgcn.image.load";
1595 break;
1596 case ac_image_load_mip:
1597 name = "llvm.amdgcn.image.load.mip";
1598 break;
1599 case ac_image_store:
1600 name = "llvm.amdgcn.image.store";
1601 retty = ctx->voidt;
1602 break;
1603 case ac_image_store_mip:
1604 name = "llvm.amdgcn.image.store.mip";
1605 retty = ctx->voidt;
1606 break;
1607 case ac_image_atomic:
1608 case ac_image_atomic_cmpswap:
1609 name = "llvm.amdgcn.image.atomic.";
1610 retty = ctx->i32;
1611 if (a->opcode == ac_image_atomic_cmpswap) {
1612 atomic_subop = "cmpswap";
1613 } else {
1614 atomic_subop = get_atomic_name(a->atomic);
1615 }
1616 break;
1617 case ac_image_get_lod:
1618 name = "llvm.amdgcn.image.getlod";
1619 break;
1620 case ac_image_get_resinfo:
1621 name = "llvm.amdgcn.image.getresinfo";
1622 break;
1623 default:
1624 unreachable("invalid image opcode");
1625 }
1626
1627 ac_build_type_name_for_intr(LLVMTypeOf(args[coords_arg]), coords_type,
1628 sizeof(coords_type));
1629
1630 if (atomic) {
1631 snprintf(intr_name, sizeof(intr_name), "llvm.amdgcn.image.atomic.%s.%s",
1632 atomic_subop, coords_type);
1633 } else {
1634 bool lod_suffix =
1635 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
1636
1637 snprintf(intr_name, sizeof(intr_name), "%s%s%s%s.v4f32.%s.v8i32",
1638 name,
1639 a->compare ? ".c" : "",
1640 a->bias ? ".b" :
1641 lod_suffix ? ".l" :
1642 a->derivs[0] ? ".d" :
1643 a->level_zero ? ".lz" : "",
1644 a->offset ? ".o" : "",
1645 coords_type);
1646 }
1647
1648 LLVMValueRef result =
1649 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
1650 a->attributes);
1651 if (!sample && retty == ctx->v4f32) {
1652 result = LLVMBuildBitCast(ctx->builder, result,
1653 ctx->v4i32, "");
1654 }
1655 return result;
1656 }
1657
1658 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
1659 struct ac_image_args *a)
1660 {
1661 const char *overload[3] = { "", "", "" };
1662 unsigned num_overloads = 0;
1663 LLVMValueRef args[18];
1664 unsigned num_args = 0;
1665
1666 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
1667 !a->level_zero);
1668 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
1669 a->opcode != ac_image_store_mip) ||
1670 a->lod);
1671 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
1672 (!a->compare && !a->offset));
1673 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
1674 a->opcode == ac_image_get_lod) ||
1675 !a->bias);
1676 assert((a->bias ? 1 : 0) +
1677 (a->lod ? 1 : 0) +
1678 (a->level_zero ? 1 : 0) +
1679 (a->derivs[0] ? 1 : 0) <= 1);
1680
1681 if (HAVE_LLVM < 0x0700)
1682 return ac_build_image_opcode_llvm6(ctx, a);
1683
1684 bool sample = a->opcode == ac_image_sample ||
1685 a->opcode == ac_image_gather4 ||
1686 a->opcode == ac_image_get_lod;
1687 bool atomic = a->opcode == ac_image_atomic ||
1688 a->opcode == ac_image_atomic_cmpswap;
1689 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
1690
1691 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
1692 args[num_args++] = a->data[0];
1693 if (a->opcode == ac_image_atomic_cmpswap)
1694 args[num_args++] = a->data[1];
1695 }
1696
1697 if (!atomic)
1698 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
1699
1700 if (a->offset)
1701 args[num_args++] = ac_to_integer(ctx, a->offset);
1702 if (a->bias) {
1703 args[num_args++] = ac_to_float(ctx, a->bias);
1704 overload[num_overloads++] = ".f32";
1705 }
1706 if (a->compare)
1707 args[num_args++] = ac_to_float(ctx, a->compare);
1708 if (a->derivs[0]) {
1709 unsigned count = ac_num_derivs(a->dim);
1710 for (unsigned i = 0; i < count; ++i)
1711 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
1712 overload[num_overloads++] = ".f32";
1713 }
1714 unsigned num_coords =
1715 a->opcode != ac_image_get_resinfo ? ac_num_coords(a->dim) : 0;
1716 for (unsigned i = 0; i < num_coords; ++i)
1717 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
1718 if (a->lod)
1719 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
1720 overload[num_overloads++] = sample ? ".f32" : ".i32";
1721
1722 args[num_args++] = a->resource;
1723 if (sample) {
1724 args[num_args++] = a->sampler;
1725 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
1726 }
1727
1728 args[num_args++] = ctx->i32_0; /* texfailctrl */
1729 args[num_args++] = LLVMConstInt(ctx->i32, a->cache_policy, false);
1730
1731 const char *name;
1732 const char *atomic_subop = "";
1733 switch (a->opcode) {
1734 case ac_image_sample: name = "sample"; break;
1735 case ac_image_gather4: name = "gather4"; break;
1736 case ac_image_load: name = "load"; break;
1737 case ac_image_load_mip: name = "load.mip"; break;
1738 case ac_image_store: name = "store"; break;
1739 case ac_image_store_mip: name = "store.mip"; break;
1740 case ac_image_atomic:
1741 name = "atomic.";
1742 atomic_subop = get_atomic_name(a->atomic);
1743 break;
1744 case ac_image_atomic_cmpswap:
1745 name = "atomic.";
1746 atomic_subop = "cmpswap";
1747 break;
1748 case ac_image_get_lod: name = "getlod"; break;
1749 case ac_image_get_resinfo: name = "getresinfo"; break;
1750 default: unreachable("invalid image opcode");
1751 }
1752
1753 const char *dimname;
1754 switch (a->dim) {
1755 case ac_image_1d: dimname = "1d"; break;
1756 case ac_image_2d: dimname = "2d"; break;
1757 case ac_image_3d: dimname = "3d"; break;
1758 case ac_image_cube: dimname = "cube"; break;
1759 case ac_image_1darray: dimname = "1darray"; break;
1760 case ac_image_2darray: dimname = "2darray"; break;
1761 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
1762 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
1763 default: unreachable("invalid dim");
1764 }
1765
1766 bool lod_suffix =
1767 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
1768 char intr_name[96];
1769 snprintf(intr_name, sizeof(intr_name),
1770 "llvm.amdgcn.image.%s%s" /* base name */
1771 "%s%s%s" /* sample/gather modifiers */
1772 ".%s.%s%s%s%s", /* dimension and type overloads */
1773 name, atomic_subop,
1774 a->compare ? ".c" : "",
1775 a->bias ? ".b" :
1776 lod_suffix ? ".l" :
1777 a->derivs[0] ? ".d" :
1778 a->level_zero ? ".lz" : "",
1779 a->offset ? ".o" : "",
1780 dimname,
1781 atomic ? "i32" : "v4f32",
1782 overload[0], overload[1], overload[2]);
1783
1784 LLVMTypeRef retty;
1785 if (atomic)
1786 retty = ctx->i32;
1787 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
1788 retty = ctx->voidt;
1789 else
1790 retty = ctx->v4f32;
1791
1792 LLVMValueRef result =
1793 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
1794 a->attributes);
1795 if (!sample && retty == ctx->v4f32) {
1796 result = LLVMBuildBitCast(ctx->builder, result,
1797 ctx->v4i32, "");
1798 }
1799 return result;
1800 }
1801
1802 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
1803 LLVMValueRef args[2])
1804 {
1805 LLVMTypeRef v2f16 =
1806 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
1807 LLVMValueRef res =
1808 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz",
1809 v2f16, args, 2,
1810 AC_FUNC_ATTR_READNONE);
1811 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1812 }
1813
1814 /* Upper 16 bits must be zero. */
1815 static LLVMValueRef ac_llvm_pack_two_int16(struct ac_llvm_context *ctx,
1816 LLVMValueRef val[2])
1817 {
1818 return LLVMBuildOr(ctx->builder, val[0],
1819 LLVMBuildShl(ctx->builder, val[1],
1820 LLVMConstInt(ctx->i32, 16, 0),
1821 ""), "");
1822 }
1823
1824 /* Upper 16 bits are ignored and will be dropped. */
1825 static LLVMValueRef ac_llvm_pack_two_int32_as_int16(struct ac_llvm_context *ctx,
1826 LLVMValueRef val[2])
1827 {
1828 LLVMValueRef v[2] = {
1829 LLVMBuildAnd(ctx->builder, val[0],
1830 LLVMConstInt(ctx->i32, 0xffff, 0), ""),
1831 val[1],
1832 };
1833 return ac_llvm_pack_two_int16(ctx, v);
1834 }
1835
1836 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
1837 LLVMValueRef args[2])
1838 {
1839 if (HAVE_LLVM >= 0x0600) {
1840 LLVMValueRef res =
1841 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
1842 ctx->v2i16, args, 2,
1843 AC_FUNC_ATTR_READNONE);
1844 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1845 }
1846
1847 LLVMValueRef val[2];
1848
1849 for (int chan = 0; chan < 2; chan++) {
1850 /* Clamp between [-1, 1]. */
1851 val[chan] = ac_build_fmin(ctx, args[chan], ctx->f32_1);
1852 val[chan] = ac_build_fmax(ctx, val[chan], LLVMConstReal(ctx->f32, -1));
1853 /* Convert to a signed integer in [-32767, 32767]. */
1854 val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
1855 LLVMConstReal(ctx->f32, 32767), "");
1856 /* If positive, add 0.5, else add -0.5. */
1857 val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
1858 LLVMBuildSelect(ctx->builder,
1859 LLVMBuildFCmp(ctx->builder, LLVMRealOGE,
1860 val[chan], ctx->f32_0, ""),
1861 LLVMConstReal(ctx->f32, 0.5),
1862 LLVMConstReal(ctx->f32, -0.5), ""), "");
1863 val[chan] = LLVMBuildFPToSI(ctx->builder, val[chan], ctx->i32, "");
1864 }
1865 return ac_llvm_pack_two_int32_as_int16(ctx, val);
1866 }
1867
1868 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
1869 LLVMValueRef args[2])
1870 {
1871 if (HAVE_LLVM >= 0x0600) {
1872 LLVMValueRef res =
1873 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
1874 ctx->v2i16, args, 2,
1875 AC_FUNC_ATTR_READNONE);
1876 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1877 }
1878
1879 LLVMValueRef val[2];
1880
1881 for (int chan = 0; chan < 2; chan++) {
1882 val[chan] = ac_build_clamp(ctx, args[chan]);
1883 val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
1884 LLVMConstReal(ctx->f32, 65535), "");
1885 val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
1886 LLVMConstReal(ctx->f32, 0.5), "");
1887 val[chan] = LLVMBuildFPToUI(ctx->builder, val[chan],
1888 ctx->i32, "");
1889 }
1890 return ac_llvm_pack_two_int32_as_int16(ctx, val);
1891 }
1892
1893 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1894 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
1895 LLVMValueRef args[2], unsigned bits, bool hi)
1896 {
1897 assert(bits == 8 || bits == 10 || bits == 16);
1898
1899 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1900 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
1901 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
1902 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
1903 LLVMValueRef max_alpha =
1904 bits != 10 ? max_rgb : ctx->i32_1;
1905 LLVMValueRef min_alpha =
1906 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
1907 bool has_intrinsic = HAVE_LLVM >= 0x0600;
1908
1909 /* Clamp. */
1910 if (!has_intrinsic || bits != 16) {
1911 for (int i = 0; i < 2; i++) {
1912 bool alpha = hi && i == 1;
1913 args[i] = ac_build_imin(ctx, args[i],
1914 alpha ? max_alpha : max_rgb);
1915 args[i] = ac_build_imax(ctx, args[i],
1916 alpha ? min_alpha : min_rgb);
1917 }
1918 }
1919
1920 if (has_intrinsic) {
1921 LLVMValueRef res =
1922 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
1923 ctx->v2i16, args, 2,
1924 AC_FUNC_ATTR_READNONE);
1925 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1926 }
1927
1928 return ac_llvm_pack_two_int32_as_int16(ctx, args);
1929 }
1930
1931 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1932 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
1933 LLVMValueRef args[2], unsigned bits, bool hi)
1934 {
1935 assert(bits == 8 || bits == 10 || bits == 16);
1936
1937 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1938 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
1939 LLVMValueRef max_alpha =
1940 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
1941 bool has_intrinsic = HAVE_LLVM >= 0x0600;
1942
1943 /* Clamp. */
1944 if (!has_intrinsic || bits != 16) {
1945 for (int i = 0; i < 2; i++) {
1946 bool alpha = hi && i == 1;
1947 args[i] = ac_build_umin(ctx, args[i],
1948 alpha ? max_alpha : max_rgb);
1949 }
1950 }
1951
1952 if (has_intrinsic) {
1953 LLVMValueRef res =
1954 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
1955 ctx->v2i16, args, 2,
1956 AC_FUNC_ATTR_READNONE);
1957 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1958 }
1959
1960 return ac_llvm_pack_two_int16(ctx, args);
1961 }
1962
1963 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
1964 {
1965 assert(HAVE_LLVM >= 0x0600);
1966 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
1967 &i1, 1, AC_FUNC_ATTR_READNONE);
1968 }
1969
1970 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
1971 {
1972 if (HAVE_LLVM >= 0x0600) {
1973 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
1974 &i1, 1, 0);
1975 return;
1976 }
1977
1978 LLVMValueRef value = LLVMBuildSelect(ctx->builder, i1,
1979 LLVMConstReal(ctx->f32, 1),
1980 LLVMConstReal(ctx->f32, -1), "");
1981 ac_build_intrinsic(ctx, "llvm.AMDGPU.kill", ctx->voidt,
1982 &value, 1, AC_FUNC_ATTR_LEGACY);
1983 }
1984
1985 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
1986 LLVMValueRef offset, LLVMValueRef width,
1987 bool is_signed)
1988 {
1989 LLVMValueRef args[] = {
1990 input,
1991 offset,
1992 width,
1993 };
1994
1995 return ac_build_intrinsic(ctx,
1996 is_signed ? "llvm.amdgcn.sbfe.i32" :
1997 "llvm.amdgcn.ubfe.i32",
1998 ctx->i32, args, 3,
1999 AC_FUNC_ATTR_READNONE);
2000 }
2001
2002 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
2003 {
2004 LLVMValueRef args[1] = {
2005 LLVMConstInt(ctx->i32, simm16, false),
2006 };
2007 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2008 ctx->voidt, args, 1, 0);
2009 }
2010
2011 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2012 unsigned bitsize)
2013 {
2014 LLVMTypeRef type;
2015 char *intr;
2016
2017 if (bitsize == 32) {
2018 intr = "llvm.floor.f32";
2019 type = ctx->f32;
2020 } else {
2021 intr = "llvm.floor.f64";
2022 type = ctx->f64;
2023 }
2024
2025 LLVMValueRef params[] = {
2026 src0,
2027 };
2028 LLVMValueRef floor = ac_build_intrinsic(ctx, intr, type, params, 1,
2029 AC_FUNC_ATTR_READNONE);
2030 return LLVMBuildFSub(ctx->builder, src0, floor, "");
2031 }
2032
2033 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2034 unsigned bitsize)
2035 {
2036 LLVMValueRef cmp, val, zero, one;
2037 LLVMTypeRef type;
2038
2039 if (bitsize == 32) {
2040 type = ctx->i32;
2041 zero = ctx->i32_0;
2042 one = ctx->i32_1;
2043 } else {
2044 type = ctx->i64;
2045 zero = ctx->i64_0;
2046 one = ctx->i64_1;
2047 }
2048
2049 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2050 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2051 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2052 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2053 return val;
2054 }
2055
2056 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2057 unsigned bitsize)
2058 {
2059 LLVMValueRef cmp, val, zero, one;
2060 LLVMTypeRef type;
2061
2062 if (bitsize == 32) {
2063 type = ctx->f32;
2064 zero = ctx->f32_0;
2065 one = ctx->f32_1;
2066 } else {
2067 type = ctx->f64;
2068 zero = ctx->f64_0;
2069 one = ctx->f64_1;
2070 }
2071
2072 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2073 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2074 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2075 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2076 return val;
2077 }
2078
2079 #define AC_EXP_TARGET 0
2080 #define AC_EXP_ENABLED_CHANNELS 1
2081 #define AC_EXP_OUT0 2
2082
2083 enum ac_ir_type {
2084 AC_IR_UNDEF,
2085 AC_IR_CONST,
2086 AC_IR_VALUE,
2087 };
2088
2089 struct ac_vs_exp_chan
2090 {
2091 LLVMValueRef value;
2092 float const_float;
2093 enum ac_ir_type type;
2094 };
2095
2096 struct ac_vs_exp_inst {
2097 unsigned offset;
2098 LLVMValueRef inst;
2099 struct ac_vs_exp_chan chan[4];
2100 };
2101
2102 struct ac_vs_exports {
2103 unsigned num;
2104 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2105 };
2106
2107 /* Return true if the PARAM export has been eliminated. */
2108 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2109 uint32_t num_outputs,
2110 struct ac_vs_exp_inst *exp)
2111 {
2112 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2113 bool is_zero[4] = {}, is_one[4] = {};
2114
2115 for (i = 0; i < 4; i++) {
2116 /* It's a constant expression. Undef outputs are eliminated too. */
2117 if (exp->chan[i].type == AC_IR_UNDEF) {
2118 is_zero[i] = true;
2119 is_one[i] = true;
2120 } else if (exp->chan[i].type == AC_IR_CONST) {
2121 if (exp->chan[i].const_float == 0)
2122 is_zero[i] = true;
2123 else if (exp->chan[i].const_float == 1)
2124 is_one[i] = true;
2125 else
2126 return false; /* other constant */
2127 } else
2128 return false;
2129 }
2130
2131 /* Only certain combinations of 0 and 1 can be eliminated. */
2132 if (is_zero[0] && is_zero[1] && is_zero[2])
2133 default_val = is_zero[3] ? 0 : 1;
2134 else if (is_one[0] && is_one[1] && is_one[2])
2135 default_val = is_zero[3] ? 2 : 3;
2136 else
2137 return false;
2138
2139 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2140 LLVMInstructionEraseFromParent(exp->inst);
2141
2142 /* Change OFFSET to DEFAULT_VAL. */
2143 for (i = 0; i < num_outputs; i++) {
2144 if (vs_output_param_offset[i] == exp->offset) {
2145 vs_output_param_offset[i] =
2146 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
2147 break;
2148 }
2149 }
2150 return true;
2151 }
2152
2153 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
2154 uint8_t *vs_output_param_offset,
2155 uint32_t num_outputs,
2156 struct ac_vs_exports *processed,
2157 struct ac_vs_exp_inst *exp)
2158 {
2159 unsigned p, copy_back_channels = 0;
2160
2161 /* See if the output is already in the list of processed outputs.
2162 * The LLVMValueRef comparison relies on SSA.
2163 */
2164 for (p = 0; p < processed->num; p++) {
2165 bool different = false;
2166
2167 for (unsigned j = 0; j < 4; j++) {
2168 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
2169 struct ac_vs_exp_chan *c2 = &exp->chan[j];
2170
2171 /* Treat undef as a match. */
2172 if (c2->type == AC_IR_UNDEF)
2173 continue;
2174
2175 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2176 * and consider the instruction duplicated.
2177 */
2178 if (c1->type == AC_IR_UNDEF) {
2179 copy_back_channels |= 1 << j;
2180 continue;
2181 }
2182
2183 /* Test whether the channels are not equal. */
2184 if (c1->type != c2->type ||
2185 (c1->type == AC_IR_CONST &&
2186 c1->const_float != c2->const_float) ||
2187 (c1->type == AC_IR_VALUE &&
2188 c1->value != c2->value)) {
2189 different = true;
2190 break;
2191 }
2192 }
2193 if (!different)
2194 break;
2195
2196 copy_back_channels = 0;
2197 }
2198 if (p == processed->num)
2199 return false;
2200
2201 /* If a match was found, but the matching export has undef where the new
2202 * one has a normal value, copy the normal value to the undef channel.
2203 */
2204 struct ac_vs_exp_inst *match = &processed->exp[p];
2205
2206 /* Get current enabled channels mask. */
2207 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
2208 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
2209
2210 while (copy_back_channels) {
2211 unsigned chan = u_bit_scan(&copy_back_channels);
2212
2213 assert(match->chan[chan].type == AC_IR_UNDEF);
2214 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
2215 exp->chan[chan].value);
2216 match->chan[chan] = exp->chan[chan];
2217
2218 /* Update number of enabled channels because the original mask
2219 * is not always 0xf.
2220 */
2221 enabled_channels |= (1 << chan);
2222 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
2223 LLVMConstInt(ctx->i32, enabled_channels, 0));
2224 }
2225
2226 /* The PARAM export is duplicated. Kill it. */
2227 LLVMInstructionEraseFromParent(exp->inst);
2228
2229 /* Change OFFSET to the matching export. */
2230 for (unsigned i = 0; i < num_outputs; i++) {
2231 if (vs_output_param_offset[i] == exp->offset) {
2232 vs_output_param_offset[i] = match->offset;
2233 break;
2234 }
2235 }
2236 return true;
2237 }
2238
2239 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
2240 LLVMValueRef main_fn,
2241 uint8_t *vs_output_param_offset,
2242 uint32_t num_outputs,
2243 uint8_t *num_param_exports)
2244 {
2245 LLVMBasicBlockRef bb;
2246 bool removed_any = false;
2247 struct ac_vs_exports exports;
2248
2249 exports.num = 0;
2250
2251 /* Process all LLVM instructions. */
2252 bb = LLVMGetFirstBasicBlock(main_fn);
2253 while (bb) {
2254 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
2255
2256 while (inst) {
2257 LLVMValueRef cur = inst;
2258 inst = LLVMGetNextInstruction(inst);
2259 struct ac_vs_exp_inst exp;
2260
2261 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
2262 continue;
2263
2264 LLVMValueRef callee = ac_llvm_get_called_value(cur);
2265
2266 if (!ac_llvm_is_function(callee))
2267 continue;
2268
2269 const char *name = LLVMGetValueName(callee);
2270 unsigned num_args = LLVMCountParams(callee);
2271
2272 /* Check if this is an export instruction. */
2273 if ((num_args != 9 && num_args != 8) ||
2274 (strcmp(name, "llvm.SI.export") &&
2275 strcmp(name, "llvm.amdgcn.exp.f32")))
2276 continue;
2277
2278 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
2279 unsigned target = LLVMConstIntGetZExtValue(arg);
2280
2281 if (target < V_008DFC_SQ_EXP_PARAM)
2282 continue;
2283
2284 target -= V_008DFC_SQ_EXP_PARAM;
2285
2286 /* Parse the instruction. */
2287 memset(&exp, 0, sizeof(exp));
2288 exp.offset = target;
2289 exp.inst = cur;
2290
2291 for (unsigned i = 0; i < 4; i++) {
2292 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
2293
2294 exp.chan[i].value = v;
2295
2296 if (LLVMIsUndef(v)) {
2297 exp.chan[i].type = AC_IR_UNDEF;
2298 } else if (LLVMIsAConstantFP(v)) {
2299 LLVMBool loses_info;
2300 exp.chan[i].type = AC_IR_CONST;
2301 exp.chan[i].const_float =
2302 LLVMConstRealGetDouble(v, &loses_info);
2303 } else {
2304 exp.chan[i].type = AC_IR_VALUE;
2305 }
2306 }
2307
2308 /* Eliminate constant and duplicated PARAM exports. */
2309 if (ac_eliminate_const_output(vs_output_param_offset,
2310 num_outputs, &exp) ||
2311 ac_eliminate_duplicated_output(ctx,
2312 vs_output_param_offset,
2313 num_outputs, &exports,
2314 &exp)) {
2315 removed_any = true;
2316 } else {
2317 exports.exp[exports.num++] = exp;
2318 }
2319 }
2320 bb = LLVMGetNextBasicBlock(bb);
2321 }
2322
2323 /* Remove holes in export memory due to removed PARAM exports.
2324 * This is done by renumbering all PARAM exports.
2325 */
2326 if (removed_any) {
2327 uint8_t old_offset[VARYING_SLOT_MAX];
2328 unsigned out, i;
2329
2330 /* Make a copy of the offsets. We need the old version while
2331 * we are modifying some of them. */
2332 memcpy(old_offset, vs_output_param_offset,
2333 sizeof(old_offset));
2334
2335 for (i = 0; i < exports.num; i++) {
2336 unsigned offset = exports.exp[i].offset;
2337
2338 /* Update vs_output_param_offset. Multiple outputs can
2339 * have the same offset.
2340 */
2341 for (out = 0; out < num_outputs; out++) {
2342 if (old_offset[out] == offset)
2343 vs_output_param_offset[out] = i;
2344 }
2345
2346 /* Change the PARAM offset in the instruction. */
2347 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
2348 LLVMConstInt(ctx->i32,
2349 V_008DFC_SQ_EXP_PARAM + i, 0));
2350 }
2351 *num_param_exports = exports.num;
2352 }
2353 }
2354
2355 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
2356 {
2357 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
2358 ac_build_intrinsic(ctx,
2359 "llvm.amdgcn.init.exec", ctx->voidt,
2360 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
2361 }
2362
2363 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
2364 {
2365 unsigned lds_size = ctx->chip_class >= CIK ? 65536 : 32768;
2366 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
2367 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_LOCAL_ADDR_SPACE),
2368 "lds");
2369 }
2370
2371 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
2372 LLVMValueRef dw_addr)
2373 {
2374 return ac_build_load(ctx, ctx->lds, dw_addr);
2375 }
2376
2377 void ac_lds_store(struct ac_llvm_context *ctx,
2378 LLVMValueRef dw_addr,
2379 LLVMValueRef value)
2380 {
2381 value = ac_to_integer(ctx, value);
2382 ac_build_indexed_store(ctx, ctx->lds,
2383 dw_addr, value);
2384 }
2385
2386 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
2387 LLVMTypeRef dst_type,
2388 LLVMValueRef src0)
2389 {
2390 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2391 const char *intrin_name;
2392 LLVMTypeRef type;
2393 LLVMValueRef zero;
2394 if (src0_bitsize == 64) {
2395 intrin_name = "llvm.cttz.i64";
2396 type = ctx->i64;
2397 zero = ctx->i64_0;
2398 } else {
2399 intrin_name = "llvm.cttz.i32";
2400 type = ctx->i32;
2401 zero = ctx->i32_0;
2402 }
2403
2404 LLVMValueRef params[2] = {
2405 src0,
2406
2407 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
2408 * add special code to check for x=0. The reason is that
2409 * the LLVM behavior for x=0 is different from what we
2410 * need here. However, LLVM also assumes that ffs(x) is
2411 * in [0, 31], but GLSL expects that ffs(0) = -1, so
2412 * a conditional assignment to handle 0 is still required.
2413 *
2414 * The hardware already implements the correct behavior.
2415 */
2416 LLVMConstInt(ctx->i1, 1, false),
2417 };
2418
2419 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
2420 params, 2,
2421 AC_FUNC_ATTR_READNONE);
2422
2423 if (src0_bitsize == 64) {
2424 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
2425 }
2426
2427 /* TODO: We need an intrinsic to skip this conditional. */
2428 /* Check for zero: */
2429 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
2430 LLVMIntEQ, src0,
2431 zero, ""),
2432 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
2433 }
2434
2435 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
2436 {
2437 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2438 AC_CONST_ADDR_SPACE);
2439 }
2440
2441 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
2442 {
2443 if (!HAVE_32BIT_POINTERS)
2444 return ac_array_in_const_addr_space(elem_type);
2445
2446 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2447 AC_CONST_32BIT_ADDR_SPACE);
2448 }
2449
2450 static struct ac_llvm_flow *
2451 get_current_flow(struct ac_llvm_context *ctx)
2452 {
2453 if (ctx->flow_depth > 0)
2454 return &ctx->flow[ctx->flow_depth - 1];
2455 return NULL;
2456 }
2457
2458 static struct ac_llvm_flow *
2459 get_innermost_loop(struct ac_llvm_context *ctx)
2460 {
2461 for (unsigned i = ctx->flow_depth; i > 0; --i) {
2462 if (ctx->flow[i - 1].loop_entry_block)
2463 return &ctx->flow[i - 1];
2464 }
2465 return NULL;
2466 }
2467
2468 static struct ac_llvm_flow *
2469 push_flow(struct ac_llvm_context *ctx)
2470 {
2471 struct ac_llvm_flow *flow;
2472
2473 if (ctx->flow_depth >= ctx->flow_depth_max) {
2474 unsigned new_max = MAX2(ctx->flow_depth << 1,
2475 AC_LLVM_INITIAL_CF_DEPTH);
2476
2477 ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
2478 ctx->flow_depth_max = new_max;
2479 }
2480
2481 flow = &ctx->flow[ctx->flow_depth];
2482 ctx->flow_depth++;
2483
2484 flow->next_block = NULL;
2485 flow->loop_entry_block = NULL;
2486 return flow;
2487 }
2488
2489 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
2490 int label_id)
2491 {
2492 char buf[32];
2493 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
2494 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
2495 }
2496
2497 /* Append a basic block at the level of the parent flow.
2498 */
2499 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
2500 const char *name)
2501 {
2502 assert(ctx->flow_depth >= 1);
2503
2504 if (ctx->flow_depth >= 2) {
2505 struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
2506
2507 return LLVMInsertBasicBlockInContext(ctx->context,
2508 flow->next_block, name);
2509 }
2510
2511 LLVMValueRef main_fn =
2512 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
2513 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
2514 }
2515
2516 /* Emit a branch to the given default target for the current block if
2517 * applicable -- that is, if the current block does not already contain a
2518 * branch from a break or continue.
2519 */
2520 static void emit_default_branch(LLVMBuilderRef builder,
2521 LLVMBasicBlockRef target)
2522 {
2523 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
2524 LLVMBuildBr(builder, target);
2525 }
2526
2527 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
2528 {
2529 struct ac_llvm_flow *flow = push_flow(ctx);
2530 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
2531 flow->next_block = append_basic_block(ctx, "ENDLOOP");
2532 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
2533 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2534 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
2535 }
2536
2537 void ac_build_break(struct ac_llvm_context *ctx)
2538 {
2539 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2540 LLVMBuildBr(ctx->builder, flow->next_block);
2541 }
2542
2543 void ac_build_continue(struct ac_llvm_context *ctx)
2544 {
2545 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2546 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2547 }
2548
2549 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
2550 {
2551 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2552 LLVMBasicBlockRef endif_block;
2553
2554 assert(!current_branch->loop_entry_block);
2555
2556 endif_block = append_basic_block(ctx, "ENDIF");
2557 emit_default_branch(ctx->builder, endif_block);
2558
2559 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2560 set_basicblock_name(current_branch->next_block, "else", label_id);
2561
2562 current_branch->next_block = endif_block;
2563 }
2564
2565 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
2566 {
2567 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2568
2569 assert(!current_branch->loop_entry_block);
2570
2571 emit_default_branch(ctx->builder, current_branch->next_block);
2572 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2573 set_basicblock_name(current_branch->next_block, "endif", label_id);
2574
2575 ctx->flow_depth--;
2576 }
2577
2578 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
2579 {
2580 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
2581
2582 assert(current_loop->loop_entry_block);
2583
2584 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
2585
2586 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
2587 set_basicblock_name(current_loop->next_block, "endloop", label_id);
2588 ctx->flow_depth--;
2589 }
2590
2591 static void if_cond_emit(struct ac_llvm_context *ctx, LLVMValueRef cond,
2592 int label_id)
2593 {
2594 struct ac_llvm_flow *flow = push_flow(ctx);
2595 LLVMBasicBlockRef if_block;
2596
2597 if_block = append_basic_block(ctx, "IF");
2598 flow->next_block = append_basic_block(ctx, "ELSE");
2599 set_basicblock_name(if_block, "if", label_id);
2600 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
2601 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
2602 }
2603
2604 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
2605 int label_id)
2606 {
2607 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
2608 value, ctx->f32_0, "");
2609 if_cond_emit(ctx, cond, label_id);
2610 }
2611
2612 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
2613 int label_id)
2614 {
2615 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
2616 ac_to_integer(ctx, value),
2617 ctx->i32_0, "");
2618 if_cond_emit(ctx, cond, label_id);
2619 }
2620
2621 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac, LLVMTypeRef type,
2622 const char *name)
2623 {
2624 LLVMBuilderRef builder = ac->builder;
2625 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
2626 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
2627 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
2628 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
2629 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
2630 LLVMValueRef res;
2631
2632 if (first_instr) {
2633 LLVMPositionBuilderBefore(first_builder, first_instr);
2634 } else {
2635 LLVMPositionBuilderAtEnd(first_builder, first_block);
2636 }
2637
2638 res = LLVMBuildAlloca(first_builder, type, name);
2639 LLVMBuildStore(builder, LLVMConstNull(type), res);
2640
2641 LLVMDisposeBuilder(first_builder);
2642
2643 return res;
2644 }
2645
2646 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac,
2647 LLVMTypeRef type, const char *name)
2648 {
2649 LLVMValueRef ptr = ac_build_alloca(ac, type, name);
2650 LLVMBuildStore(ac->builder, LLVMGetUndef(type), ptr);
2651 return ptr;
2652 }
2653
2654 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
2655 LLVMTypeRef type)
2656 {
2657 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
2658 return LLVMBuildBitCast(ctx->builder, ptr,
2659 LLVMPointerType(type, addr_space), "");
2660 }
2661
2662 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
2663 unsigned count)
2664 {
2665 unsigned num_components = ac_get_llvm_num_components(value);
2666 if (count == num_components)
2667 return value;
2668
2669 LLVMValueRef masks[] = {
2670 LLVMConstInt(ctx->i32, 0, false), LLVMConstInt(ctx->i32, 1, false),
2671 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false)};
2672
2673 if (count == 1)
2674 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
2675 "");
2676
2677 LLVMValueRef swizzle = LLVMConstVector(masks, count);
2678 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
2679 }
2680
2681 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
2682 unsigned rshift, unsigned bitwidth)
2683 {
2684 LLVMValueRef value = param;
2685 if (rshift)
2686 value = LLVMBuildLShr(ctx->builder, value,
2687 LLVMConstInt(ctx->i32, rshift, false), "");
2688
2689 if (rshift + bitwidth < 32) {
2690 unsigned mask = (1 << bitwidth) - 1;
2691 value = LLVMBuildAnd(ctx->builder, value,
2692 LLVMConstInt(ctx->i32, mask, false), "");
2693 }
2694 return value;
2695 }
2696
2697 /* Adjust the sample index according to FMASK.
2698 *
2699 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2700 * which is the identity mapping. Each nibble says which physical sample
2701 * should be fetched to get that sample.
2702 *
2703 * For example, 0x11111100 means there are only 2 samples stored and
2704 * the second sample covers 3/4 of the pixel. When reading samples 0
2705 * and 1, return physical sample 0 (determined by the first two 0s
2706 * in FMASK), otherwise return physical sample 1.
2707 *
2708 * The sample index should be adjusted as follows:
2709 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
2710 */
2711 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
2712 LLVMValueRef *addr, bool is_array_tex)
2713 {
2714 struct ac_image_args fmask_load = {};
2715 fmask_load.opcode = ac_image_load;
2716 fmask_load.resource = fmask;
2717 fmask_load.dmask = 0xf;
2718 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
2719
2720 fmask_load.coords[0] = addr[0];
2721 fmask_load.coords[1] = addr[1];
2722 if (is_array_tex)
2723 fmask_load.coords[2] = addr[2];
2724
2725 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
2726 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
2727 ac->i32_0, "");
2728
2729 /* Apply the formula. */
2730 unsigned sample_chan = is_array_tex ? 3 : 2;
2731 LLVMValueRef final_sample;
2732 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
2733 LLVMConstInt(ac->i32, 4, 0), "");
2734 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
2735 final_sample = LLVMBuildAnd(ac->builder, final_sample,
2736 LLVMConstInt(ac->i32, 0xF, 0), "");
2737
2738 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2739 * resource descriptor is 0 (invalid),
2740 */
2741 LLVMValueRef tmp;
2742 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
2743 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
2744 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
2745
2746 /* Replace the MSAA sample index. */
2747 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
2748 addr[sample_chan], "");
2749 }
2750
2751 static LLVMValueRef
2752 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
2753 {
2754 ac_build_optimization_barrier(ctx, &src);
2755 return ac_build_intrinsic(ctx,
2756 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
2757 LLVMTypeOf(src), (LLVMValueRef []) {
2758 src, lane },
2759 lane == NULL ? 1 : 2,
2760 AC_FUNC_ATTR_READNONE |
2761 AC_FUNC_ATTR_CONVERGENT);
2762 }
2763
2764 /**
2765 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
2766 * @param ctx
2767 * @param src
2768 * @param lane - id of the lane or NULL for the first active lane
2769 * @return value of the lane
2770 */
2771 LLVMValueRef
2772 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
2773 {
2774 LLVMTypeRef src_type = LLVMTypeOf(src);
2775 src = ac_to_integer(ctx, src);
2776 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
2777 LLVMValueRef ret;
2778
2779 if (bits == 32) {
2780 ret = _ac_build_readlane(ctx, src, lane);
2781 } else {
2782 assert(bits % 32 == 0);
2783 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
2784 LLVMValueRef src_vector =
2785 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
2786 ret = LLVMGetUndef(vec_type);
2787 for (unsigned i = 0; i < bits / 32; i++) {
2788 src = LLVMBuildExtractElement(ctx->builder, src_vector,
2789 LLVMConstInt(ctx->i32, i, 0), "");
2790 LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
2791 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
2792 LLVMConstInt(ctx->i32, i, 0), "");
2793 }
2794 }
2795 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
2796 }
2797
2798 LLVMValueRef
2799 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
2800 {
2801 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
2802 */
2803 LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
2804 ac_get_thread_id(ctx), "");
2805 return LLVMBuildSelect(ctx->builder, pred, value, src, "");
2806 }
2807
2808 LLVMValueRef
2809 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
2810 {
2811 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
2812 LLVMVectorType(ctx->i32, 2),
2813 "");
2814 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
2815 ctx->i32_0, "");
2816 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
2817 ctx->i32_1, "");
2818 LLVMValueRef val =
2819 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
2820 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
2821 2, AC_FUNC_ATTR_READNONE);
2822 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
2823 (LLVMValueRef []) { mask_hi, val },
2824 2, AC_FUNC_ATTR_READNONE);
2825 return val;
2826 }
2827
2828 enum dpp_ctrl {
2829 _dpp_quad_perm = 0x000,
2830 _dpp_row_sl = 0x100,
2831 _dpp_row_sr = 0x110,
2832 _dpp_row_rr = 0x120,
2833 dpp_wf_sl1 = 0x130,
2834 dpp_wf_rl1 = 0x134,
2835 dpp_wf_sr1 = 0x138,
2836 dpp_wf_rr1 = 0x13C,
2837 dpp_row_mirror = 0x140,
2838 dpp_row_half_mirror = 0x141,
2839 dpp_row_bcast15 = 0x142,
2840 dpp_row_bcast31 = 0x143
2841 };
2842
2843 static inline enum dpp_ctrl
2844 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
2845 {
2846 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
2847 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
2848 }
2849
2850 static inline enum dpp_ctrl
2851 dpp_row_sl(unsigned amount)
2852 {
2853 assert(amount > 0 && amount < 16);
2854 return _dpp_row_sl | amount;
2855 }
2856
2857 static inline enum dpp_ctrl
2858 dpp_row_sr(unsigned amount)
2859 {
2860 assert(amount > 0 && amount < 16);
2861 return _dpp_row_sr | amount;
2862 }
2863
2864 static LLVMValueRef
2865 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
2866 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
2867 bool bound_ctrl)
2868 {
2869 return ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32",
2870 LLVMTypeOf(old),
2871 (LLVMValueRef[]) {
2872 old, src,
2873 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
2874 LLVMConstInt(ctx->i32, row_mask, 0),
2875 LLVMConstInt(ctx->i32, bank_mask, 0),
2876 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
2877 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
2878 }
2879
2880 static LLVMValueRef
2881 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
2882 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
2883 bool bound_ctrl)
2884 {
2885 LLVMTypeRef src_type = LLVMTypeOf(src);
2886 src = ac_to_integer(ctx, src);
2887 old = ac_to_integer(ctx, old);
2888 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
2889 LLVMValueRef ret;
2890 if (bits == 32) {
2891 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
2892 bank_mask, bound_ctrl);
2893 } else {
2894 assert(bits % 32 == 0);
2895 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
2896 LLVMValueRef src_vector =
2897 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
2898 LLVMValueRef old_vector =
2899 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
2900 ret = LLVMGetUndef(vec_type);
2901 for (unsigned i = 0; i < bits / 32; i++) {
2902 src = LLVMBuildExtractElement(ctx->builder, src_vector,
2903 LLVMConstInt(ctx->i32, i,
2904 0), "");
2905 old = LLVMBuildExtractElement(ctx->builder, old_vector,
2906 LLVMConstInt(ctx->i32, i,
2907 0), "");
2908 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
2909 dpp_ctrl,
2910 row_mask,
2911 bank_mask,
2912 bound_ctrl);
2913 ret = LLVMBuildInsertElement(ctx->builder, ret,
2914 ret_comp,
2915 LLVMConstInt(ctx->i32, i,
2916 0), "");
2917 }
2918 }
2919 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
2920 }
2921
2922 static inline unsigned
2923 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
2924 {
2925 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
2926 return and_mask | (or_mask << 5) | (xor_mask << 10);
2927 }
2928
2929 static LLVMValueRef
2930 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
2931 {
2932 return ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle",
2933 LLVMTypeOf(src), (LLVMValueRef []) {
2934 src, LLVMConstInt(ctx->i32, mask, 0) },
2935 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
2936 }
2937
2938 LLVMValueRef
2939 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
2940 {
2941 LLVMTypeRef src_type = LLVMTypeOf(src);
2942 src = ac_to_integer(ctx, src);
2943 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
2944 LLVMValueRef ret;
2945 if (bits == 32) {
2946 ret = _ac_build_ds_swizzle(ctx, src, mask);
2947 } else {
2948 assert(bits % 32 == 0);
2949 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
2950 LLVMValueRef src_vector =
2951 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
2952 ret = LLVMGetUndef(vec_type);
2953 for (unsigned i = 0; i < bits / 32; i++) {
2954 src = LLVMBuildExtractElement(ctx->builder, src_vector,
2955 LLVMConstInt(ctx->i32, i,
2956 0), "");
2957 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
2958 mask);
2959 ret = LLVMBuildInsertElement(ctx->builder, ret,
2960 ret_comp,
2961 LLVMConstInt(ctx->i32, i,
2962 0), "");
2963 }
2964 }
2965 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
2966 }
2967
2968 static LLVMValueRef
2969 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
2970 {
2971 char name[32], type[8];
2972 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
2973 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
2974 return ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
2975 (LLVMValueRef []) { src }, 1,
2976 AC_FUNC_ATTR_READNONE);
2977 }
2978
2979 static LLVMValueRef
2980 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
2981 LLVMValueRef inactive)
2982 {
2983 char name[32], type[8];
2984 LLVMTypeRef src_type = LLVMTypeOf(src);
2985 src = ac_to_integer(ctx, src);
2986 inactive = ac_to_integer(ctx, inactive);
2987 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
2988 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
2989 LLVMValueRef ret =
2990 ac_build_intrinsic(ctx, name,
2991 LLVMTypeOf(src), (LLVMValueRef []) {
2992 src, inactive }, 2,
2993 AC_FUNC_ATTR_READNONE |
2994 AC_FUNC_ATTR_CONVERGENT);
2995 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
2996 }
2997
2998 static LLVMValueRef
2999 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
3000 {
3001 if (type_size == 4) {
3002 switch (op) {
3003 case nir_op_iadd: return ctx->i32_0;
3004 case nir_op_fadd: return ctx->f32_0;
3005 case nir_op_imul: return ctx->i32_1;
3006 case nir_op_fmul: return ctx->f32_1;
3007 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
3008 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
3009 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
3010 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
3011 case nir_op_umax: return ctx->i32_0;
3012 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
3013 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
3014 case nir_op_ior: return ctx->i32_0;
3015 case nir_op_ixor: return ctx->i32_0;
3016 default:
3017 unreachable("bad reduction intrinsic");
3018 }
3019 } else { /* type_size == 64bit */
3020 switch (op) {
3021 case nir_op_iadd: return ctx->i64_0;
3022 case nir_op_fadd: return ctx->f64_0;
3023 case nir_op_imul: return ctx->i64_1;
3024 case nir_op_fmul: return ctx->f64_1;
3025 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
3026 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
3027 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
3028 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
3029 case nir_op_umax: return ctx->i64_0;
3030 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
3031 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
3032 case nir_op_ior: return ctx->i64_0;
3033 case nir_op_ixor: return ctx->i64_0;
3034 default:
3035 unreachable("bad reduction intrinsic");
3036 }
3037 }
3038 }
3039
3040 static LLVMValueRef
3041 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
3042 {
3043 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
3044 switch (op) {
3045 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
3046 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
3047 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
3048 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
3049 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
3050 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
3051 lhs, rhs, "");
3052 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
3053 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
3054 lhs, rhs, "");
3055 case nir_op_fmin: return ac_build_intrinsic(ctx,
3056 _64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
3057 _64bit ? ctx->f64 : ctx->f32,
3058 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3059 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
3060 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
3061 lhs, rhs, "");
3062 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
3063 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
3064 lhs, rhs, "");
3065 case nir_op_fmax: return ac_build_intrinsic(ctx,
3066 _64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3067 _64bit ? ctx->f64 : ctx->f32,
3068 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3069 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
3070 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
3071 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
3072 default:
3073 unreachable("bad reduction intrinsic");
3074 }
3075 }
3076
3077 /* TODO: add inclusive and excluse scan functions for SI chip class. */
3078 static LLVMValueRef
3079 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity)
3080 {
3081 LLVMValueRef result, tmp;
3082 result = src;
3083 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
3084 result = ac_build_alu_op(ctx, result, tmp, op);
3085 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
3086 result = ac_build_alu_op(ctx, result, tmp, op);
3087 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
3088 result = ac_build_alu_op(ctx, result, tmp, op);
3089 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
3090 result = ac_build_alu_op(ctx, result, tmp, op);
3091 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
3092 result = ac_build_alu_op(ctx, result, tmp, op);
3093 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3094 result = ac_build_alu_op(ctx, result, tmp, op);
3095 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3096 result = ac_build_alu_op(ctx, result, tmp, op);
3097 return result;
3098 }
3099
3100 LLVMValueRef
3101 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3102 {
3103 ac_build_optimization_barrier(ctx, &src);
3104 LLVMValueRef result;
3105 LLVMValueRef identity = get_reduction_identity(ctx, op,
3106 ac_get_type_size(LLVMTypeOf(src)));
3107 result = LLVMBuildBitCast(ctx->builder,
3108 ac_build_set_inactive(ctx, src, identity),
3109 LLVMTypeOf(identity), "");
3110 result = ac_build_scan(ctx, op, result, identity);
3111
3112 return ac_build_wwm(ctx, result);
3113 }
3114
3115 LLVMValueRef
3116 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3117 {
3118 ac_build_optimization_barrier(ctx, &src);
3119 LLVMValueRef result;
3120 LLVMValueRef identity = get_reduction_identity(ctx, op,
3121 ac_get_type_size(LLVMTypeOf(src)));
3122 result = LLVMBuildBitCast(ctx->builder,
3123 ac_build_set_inactive(ctx, src, identity),
3124 LLVMTypeOf(identity), "");
3125 result = ac_build_dpp(ctx, identity, result, dpp_wf_sr1, 0xf, 0xf, false);
3126 result = ac_build_scan(ctx, op, result, identity);
3127
3128 return ac_build_wwm(ctx, result);
3129 }
3130
3131 LLVMValueRef
3132 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
3133 {
3134 if (cluster_size == 1) return src;
3135 ac_build_optimization_barrier(ctx, &src);
3136 LLVMValueRef result, swap;
3137 LLVMValueRef identity = get_reduction_identity(ctx, op,
3138 ac_get_type_size(LLVMTypeOf(src)));
3139 result = LLVMBuildBitCast(ctx->builder,
3140 ac_build_set_inactive(ctx, src, identity),
3141 LLVMTypeOf(identity), "");
3142 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
3143 result = ac_build_alu_op(ctx, result, swap, op);
3144 if (cluster_size == 2) return ac_build_wwm(ctx, result);
3145
3146 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
3147 result = ac_build_alu_op(ctx, result, swap, op);
3148 if (cluster_size == 4) return ac_build_wwm(ctx, result);
3149
3150 if (ctx->chip_class >= VI)
3151 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
3152 else
3153 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
3154 result = ac_build_alu_op(ctx, result, swap, op);
3155 if (cluster_size == 8) return ac_build_wwm(ctx, result);
3156
3157 if (ctx->chip_class >= VI)
3158 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
3159 else
3160 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
3161 result = ac_build_alu_op(ctx, result, swap, op);
3162 if (cluster_size == 16) return ac_build_wwm(ctx, result);
3163
3164 if (ctx->chip_class >= VI && cluster_size != 32)
3165 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3166 else
3167 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
3168 result = ac_build_alu_op(ctx, result, swap, op);
3169 if (cluster_size == 32) return ac_build_wwm(ctx, result);
3170
3171 if (ctx->chip_class >= VI) {
3172 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3173 result = ac_build_alu_op(ctx, result, swap, op);
3174 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
3175 return ac_build_wwm(ctx, result);
3176 } else {
3177 swap = ac_build_readlane(ctx, result, ctx->i32_0);
3178 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
3179 result = ac_build_alu_op(ctx, result, swap, op);
3180 return ac_build_wwm(ctx, result);
3181 }
3182 }
3183
3184 LLVMValueRef
3185 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
3186 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3187 {
3188 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
3189 if (ctx->chip_class >= VI && HAVE_LLVM >= 0x0600) {
3190 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
3191 } else {
3192 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
3193 }
3194 }
3195
3196 LLVMValueRef
3197 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
3198 {
3199 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
3200 return ac_build_intrinsic(ctx,
3201 "llvm.amdgcn.ds.bpermute", ctx->i32,
3202 (LLVMValueRef []) {index, src}, 2,
3203 AC_FUNC_ATTR_READNONE |
3204 AC_FUNC_ATTR_CONVERGENT);
3205 }