projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac: add 16-bit float support to ac_build_alu_op()
[mesa.git] / src / amd / llvm / 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 #include <llvm/Config/llvm-config.h>
30
31 #include "c11/threads.h"
32
33 #include <assert.h>
34 #include <stdio.h>
35
36 #include "ac_llvm_util.h"
37 #include "ac_shader_util.h"
38 #include "ac_exp_param.h"
39 #include "util/bitscan.h"
40 #include "util/macros.h"
41 #include "util/u_atomic.h"
42 #include "util/u_math.h"
43 #include "sid.h"
44
45 #include "shader_enums.h"
46
47 #define AC_LLVM_INITIAL_CF_DEPTH 4
48
49 /* Data for if/else/endif and bgnloop/endloop control flow structures.
50 */
51 struct ac_llvm_flow {
52 /* Loop exit or next part of if/else/endif. */
53 LLVMBasicBlockRef next_block;
54 LLVMBasicBlockRef loop_entry_block;
55 };
56
57 /* Initialize module-independent parts of the context.
58 *
59 * The caller is responsible for initializing ctx::module and ctx::builder.
60 */
61 void
62 ac_llvm_context_init(struct ac_llvm_context *ctx,
63 struct ac_llvm_compiler *compiler,
64 enum chip_class chip_class, enum radeon_family family,
65 enum ac_float_mode float_mode, unsigned wave_size,
66 unsigned ballot_mask_bits)
67 {
68 LLVMValueRef args[1];
69
70 ctx->context = LLVMContextCreate();
71
72 ctx->chip_class = chip_class;
73 ctx->family = family;
74 ctx->wave_size = wave_size;
75 ctx->ballot_mask_bits = ballot_mask_bits;
76 ctx->float_mode = float_mode;
77 ctx->module = ac_create_module(wave_size == 32 ? compiler->tm_wave32
78 : compiler->tm,
79 ctx->context);
80 ctx->builder = ac_create_builder(ctx->context, float_mode);
81
82 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
83 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
84 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
85 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
86 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
87 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
88 ctx->intptr = ctx->i32;
89 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
90 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
91 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
92 ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
93 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
94 ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
95 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
96 ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
97 ctx->v3f32 = LLVMVectorType(ctx->f32, 3);
98 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
99 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
100 ctx->iN_wavemask = LLVMIntTypeInContext(ctx->context, ctx->wave_size);
101 ctx->iN_ballotmask = LLVMIntTypeInContext(ctx->context, ballot_mask_bits);
102
103 ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
104 ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
105 ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
106 ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
107 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
108 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
109 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
110 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
111 ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
112 ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
113 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
114 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
115 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
116 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
117
118 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
119 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
120
121 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
122 "range", 5);
123
124 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
125 "invariant.load", 14);
126
127 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
128
129 args[0] = LLVMConstReal(ctx->f32, 2.5);
130 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
131
132 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
133 "amdgpu.uniform", 14);
134
135 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
136 ctx->flow = calloc(1, sizeof(*ctx->flow));
137 }
138
139 void
140 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
141 {
142 free(ctx->flow->stack);
143 free(ctx->flow);
144 ctx->flow = NULL;
145 }
146
147 int
148 ac_get_llvm_num_components(LLVMValueRef value)
149 {
150 LLVMTypeRef type = LLVMTypeOf(value);
151 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
152 ? LLVMGetVectorSize(type)
153 : 1;
154 return num_components;
155 }
156
157 LLVMValueRef
158 ac_llvm_extract_elem(struct ac_llvm_context *ac,
159 LLVMValueRef value,
160 int index)
161 {
162 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
163 assert(index == 0);
164 return value;
165 }
166
167 return LLVMBuildExtractElement(ac->builder, value,
168 LLVMConstInt(ac->i32, index, false), "");
169 }
170
171 int
172 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
173 {
174 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
175 type = LLVMGetElementType(type);
176
177 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
178 return LLVMGetIntTypeWidth(type);
179
180 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
181 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_LDS)
182 return 32;
183 }
184
185 if (type == ctx->f16)
186 return 16;
187 if (type == ctx->f32)
188 return 32;
189 if (type == ctx->f64)
190 return 64;
191
192 unreachable("Unhandled type kind in get_elem_bits");
193 }
194
195 unsigned
196 ac_get_type_size(LLVMTypeRef type)
197 {
198 LLVMTypeKind kind = LLVMGetTypeKind(type);
199
200 switch (kind) {
201 case LLVMIntegerTypeKind:
202 return LLVMGetIntTypeWidth(type) / 8;
203 case LLVMHalfTypeKind:
204 return 2;
205 case LLVMFloatTypeKind:
206 return 4;
207 case LLVMDoubleTypeKind:
208 return 8;
209 case LLVMPointerTypeKind:
210 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
211 return 4;
212 return 8;
213 case LLVMVectorTypeKind:
214 return LLVMGetVectorSize(type) *
215 ac_get_type_size(LLVMGetElementType(type));
216 case LLVMArrayTypeKind:
217 return LLVMGetArrayLength(type) *
218 ac_get_type_size(LLVMGetElementType(type));
219 default:
220 assert(0);
221 return 0;
222 }
223 }
224
225 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
226 {
227 if (t == ctx->i8)
228 return ctx->i8;
229 else if (t == ctx->f16 || t == ctx->i16)
230 return ctx->i16;
231 else if (t == ctx->f32 || t == ctx->i32)
232 return ctx->i32;
233 else if (t == ctx->f64 || t == ctx->i64)
234 return ctx->i64;
235 else
236 unreachable("Unhandled integer size");
237 }
238
239 LLVMTypeRef
240 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
241 {
242 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
243 LLVMTypeRef elem_type = LLVMGetElementType(t);
244 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
245 LLVMGetVectorSize(t));
246 }
247 if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
248 switch (LLVMGetPointerAddressSpace(t)) {
249 case AC_ADDR_SPACE_GLOBAL:
250 return ctx->i64;
251 case AC_ADDR_SPACE_CONST_32BIT:
252 case AC_ADDR_SPACE_LDS:
253 return ctx->i32;
254 default:
255 unreachable("unhandled address space");
256 }
257 }
258 return to_integer_type_scalar(ctx, t);
259 }
260
261 LLVMValueRef
262 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
263 {
264 LLVMTypeRef type = LLVMTypeOf(v);
265 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
266 return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
267 }
268 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
269 }
270
271 LLVMValueRef
272 ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
273 {
274 LLVMTypeRef type = LLVMTypeOf(v);
275 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
276 return v;
277 return ac_to_integer(ctx, v);
278 }
279
280 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
281 {
282 if (t == ctx->i8)
283 return ctx->i8;
284 else if (t == ctx->i16 || t == ctx->f16)
285 return ctx->f16;
286 else if (t == ctx->i32 || t == ctx->f32)
287 return ctx->f32;
288 else if (t == ctx->i64 || t == ctx->f64)
289 return ctx->f64;
290 else
291 unreachable("Unhandled float size");
292 }
293
294 LLVMTypeRef
295 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
296 {
297 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
298 LLVMTypeRef elem_type = LLVMGetElementType(t);
299 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
300 LLVMGetVectorSize(t));
301 }
302 return to_float_type_scalar(ctx, t);
303 }
304
305 LLVMValueRef
306 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
307 {
308 LLVMTypeRef type = LLVMTypeOf(v);
309 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
310 }
311
312
313 LLVMValueRef
314 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
315 LLVMTypeRef return_type, LLVMValueRef *params,
316 unsigned param_count, unsigned attrib_mask)
317 {
318 LLVMValueRef function, call;
319 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
320
321 function = LLVMGetNamedFunction(ctx->module, name);
322 if (!function) {
323 LLVMTypeRef param_types[32], function_type;
324 unsigned i;
325
326 assert(param_count <= 32);
327
328 for (i = 0; i < param_count; ++i) {
329 assert(params[i]);
330 param_types[i] = LLVMTypeOf(params[i]);
331 }
332 function_type =
333 LLVMFunctionType(return_type, param_types, param_count, 0);
334 function = LLVMAddFunction(ctx->module, name, function_type);
335
336 LLVMSetFunctionCallConv(function, LLVMCCallConv);
337 LLVMSetLinkage(function, LLVMExternalLinkage);
338
339 if (!set_callsite_attrs)
340 ac_add_func_attributes(ctx->context, function, attrib_mask);
341 }
342
343 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
344 if (set_callsite_attrs)
345 ac_add_func_attributes(ctx->context, call, attrib_mask);
346 return call;
347 }
348
349 /**
350 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
351 * intrinsic names).
352 */
353 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
354 {
355 LLVMTypeRef elem_type = type;
356
357 assert(bufsize >= 8);
358
359 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
360 int ret = snprintf(buf, bufsize, "v%u",
361 LLVMGetVectorSize(type));
362 if (ret < 0) {
363 char *type_name = LLVMPrintTypeToString(type);
364 fprintf(stderr, "Error building type name for: %s\n",
365 type_name);
366 LLVMDisposeMessage(type_name);
367 return;
368 }
369 elem_type = LLVMGetElementType(type);
370 buf += ret;
371 bufsize -= ret;
372 }
373 switch (LLVMGetTypeKind(elem_type)) {
374 default: break;
375 case LLVMIntegerTypeKind:
376 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
377 break;
378 case LLVMHalfTypeKind:
379 snprintf(buf, bufsize, "f16");
380 break;
381 case LLVMFloatTypeKind:
382 snprintf(buf, bufsize, "f32");
383 break;
384 case LLVMDoubleTypeKind:
385 snprintf(buf, bufsize, "f64");
386 break;
387 }
388 }
389
390 /**
391 * Helper function that builds an LLVM IR PHI node and immediately adds
392 * incoming edges.
393 */
394 LLVMValueRef
395 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
396 unsigned count_incoming, LLVMValueRef *values,
397 LLVMBasicBlockRef *blocks)
398 {
399 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
400 LLVMAddIncoming(phi, values, blocks, count_incoming);
401 return phi;
402 }
403
404 void ac_build_s_barrier(struct ac_llvm_context *ctx)
405 {
406 ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
407 0, AC_FUNC_ATTR_CONVERGENT);
408 }
409
410 /* Prevent optimizations (at least of memory accesses) across the current
411 * point in the program by emitting empty inline assembly that is marked as
412 * having side effects.
413 *
414 * Optionally, a value can be passed through the inline assembly to prevent
415 * LLVM from hoisting calls to ReadNone functions.
416 */
417 void
418 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
419 LLVMValueRef *pvgpr)
420 {
421 static int counter = 0;
422
423 LLVMBuilderRef builder = ctx->builder;
424 char code[16];
425
426 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
427
428 if (!pvgpr) {
429 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
430 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
431 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
432 } else {
433 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
434 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
435 LLVMTypeRef type = LLVMTypeOf(*pvgpr);
436 unsigned bitsize = ac_get_elem_bits(ctx, type);
437 LLVMValueRef vgpr = *pvgpr;
438 LLVMTypeRef vgpr_type;
439 unsigned vgpr_size;
440 LLVMValueRef vgpr0;
441
442 if (bitsize < 32)
443 vgpr = LLVMBuildZExt(ctx->builder, vgpr, ctx->i32, "");
444
445 vgpr_type = LLVMTypeOf(vgpr);
446 vgpr_size = ac_get_type_size(vgpr_type);
447
448 assert(vgpr_size % 4 == 0);
449
450 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
451 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
452 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
453 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
454 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
455
456 if (bitsize < 32)
457 vgpr = LLVMBuildTrunc(builder, vgpr, type, "");
458
459 *pvgpr = vgpr;
460 }
461 }
462
463 LLVMValueRef
464 ac_build_shader_clock(struct ac_llvm_context *ctx)
465 {
466 const char *intr = LLVM_VERSION_MAJOR >= 9 && ctx->chip_class >= GFX8 ?
467 "llvm.amdgcn.s.memrealtime" : "llvm.readcyclecounter";
468 LLVMValueRef tmp = ac_build_intrinsic(ctx, intr, ctx->i64, NULL, 0, 0);
469 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
470 }
471
472 LLVMValueRef
473 ac_build_ballot(struct ac_llvm_context *ctx,
474 LLVMValueRef value)
475 {
476 const char *name;
477
478 if (LLVM_VERSION_MAJOR >= 9) {
479 if (ctx->wave_size == 64)
480 name = "llvm.amdgcn.icmp.i64.i32";
481 else
482 name = "llvm.amdgcn.icmp.i32.i32";
483 } else {
484 name = "llvm.amdgcn.icmp.i32";
485 }
486 LLVMValueRef args[3] = {
487 value,
488 ctx->i32_0,
489 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
490 };
491
492 /* We currently have no other way to prevent LLVM from lifting the icmp
493 * calls to a dominating basic block.
494 */
495 ac_build_optimization_barrier(ctx, &args[0]);
496
497 args[0] = ac_to_integer(ctx, args[0]);
498
499 return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
500 AC_FUNC_ATTR_NOUNWIND |
501 AC_FUNC_ATTR_READNONE |
502 AC_FUNC_ATTR_CONVERGENT);
503 }
504
505 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
506 LLVMValueRef value)
507 {
508 const char *name = LLVM_VERSION_MAJOR >= 9 ? "llvm.amdgcn.icmp.i64.i1" : "llvm.amdgcn.icmp.i1";
509 LLVMValueRef args[3] = {
510 value,
511 ctx->i1false,
512 LLVMConstInt(ctx->i32, LLVMIntNE, 0),
513 };
514
515 return ac_build_intrinsic(ctx, name, ctx->i64, args, 3,
516 AC_FUNC_ATTR_NOUNWIND |
517 AC_FUNC_ATTR_READNONE |
518 AC_FUNC_ATTR_CONVERGENT);
519 }
520
521 LLVMValueRef
522 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
523 {
524 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
525 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
526 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
527 }
528
529 LLVMValueRef
530 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
531 {
532 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
533 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
534 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
535 }
536
537 LLVMValueRef
538 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
539 {
540 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
541 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
542
543 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
544 vote_set, active_set, "");
545 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
546 vote_set,
547 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
548 return LLVMBuildOr(ctx->builder, all, none, "");
549 }
550
551 LLVMValueRef
552 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
553 unsigned value_count, unsigned component)
554 {
555 LLVMValueRef vec = NULL;
556
557 if (value_count == 1) {
558 return values[component];
559 } else if (!value_count)
560 unreachable("value_count is 0");
561
562 for (unsigned i = component; i < value_count + component; i++) {
563 LLVMValueRef value = values[i];
564
565 if (i == component)
566 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
567 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
568 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
569 }
570 return vec;
571 }
572
573 LLVMValueRef
574 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
575 LLVMValueRef *values,
576 unsigned value_count,
577 unsigned value_stride,
578 bool load,
579 bool always_vector)
580 {
581 LLVMBuilderRef builder = ctx->builder;
582 LLVMValueRef vec = NULL;
583 unsigned i;
584
585 if (value_count == 1 && !always_vector) {
586 if (load)
587 return LLVMBuildLoad(builder, values[0], "");
588 return values[0];
589 } else if (!value_count)
590 unreachable("value_count is 0");
591
592 for (i = 0; i < value_count; i++) {
593 LLVMValueRef value = values[i * value_stride];
594 if (load)
595 value = LLVMBuildLoad(builder, value, "");
596
597 if (!i)
598 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
599 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
600 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
601 }
602 return vec;
603 }
604
605 LLVMValueRef
606 ac_build_gather_values(struct ac_llvm_context *ctx,
607 LLVMValueRef *values,
608 unsigned value_count)
609 {
610 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
611 }
612
613 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
614 * channels with undef. Extract at most src_channels components from the input.
615 */
616 static LLVMValueRef
617 ac_build_expand(struct ac_llvm_context *ctx,
618 LLVMValueRef value,
619 unsigned src_channels,
620 unsigned dst_channels)
621 {
622 LLVMTypeRef elemtype;
623 LLVMValueRef chan[dst_channels];
624
625 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
626 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
627
628 if (src_channels == dst_channels && vec_size == dst_channels)
629 return value;
630
631 src_channels = MIN2(src_channels, vec_size);
632
633 for (unsigned i = 0; i < src_channels; i++)
634 chan[i] = ac_llvm_extract_elem(ctx, value, i);
635
636 elemtype = LLVMGetElementType(LLVMTypeOf(value));
637 } else {
638 if (src_channels) {
639 assert(src_channels == 1);
640 chan[0] = value;
641 }
642 elemtype = LLVMTypeOf(value);
643 }
644
645 for (unsigned i = src_channels; i < dst_channels; i++)
646 chan[i] = LLVMGetUndef(elemtype);
647
648 return ac_build_gather_values(ctx, chan, dst_channels);
649 }
650
651 /* Extract components [start, start + channels) from a vector.
652 */
653 LLVMValueRef
654 ac_extract_components(struct ac_llvm_context *ctx,
655 LLVMValueRef value,
656 unsigned start,
657 unsigned channels)
658 {
659 LLVMValueRef chan[channels];
660
661 for (unsigned i = 0; i < channels; i++)
662 chan[i] = ac_llvm_extract_elem(ctx, value, i + start);
663
664 return ac_build_gather_values(ctx, chan, channels);
665 }
666
667 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
668 * with undef. Extract at most num_channels components from the input.
669 */
670 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
671 LLVMValueRef value,
672 unsigned num_channels)
673 {
674 return ac_build_expand(ctx, value, num_channels, 4);
675 }
676
677 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
678 {
679 unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
680 const char *name;
681
682 if (type_size == 2)
683 name = "llvm.rint.f16";
684 else if (type_size == 4)
685 name = "llvm.rint.f32";
686 else
687 name = "llvm.rint.f64";
688
689 return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
690 AC_FUNC_ATTR_READNONE);
691 }
692
693 LLVMValueRef
694 ac_build_fdiv(struct ac_llvm_context *ctx,
695 LLVMValueRef num,
696 LLVMValueRef den)
697 {
698 /* If we do (num / den), LLVM >= 7.0 does:
699 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
700 *
701 * If we do (num * (1 / den)), LLVM does:
702 * return num * v_rcp_f32(den);
703 */
704 LLVMValueRef one = LLVMConstReal(LLVMTypeOf(num), 1.0);
705 LLVMValueRef rcp = LLVMBuildFDiv(ctx->builder, one, den, "");
706 LLVMValueRef ret = LLVMBuildFMul(ctx->builder, num, rcp, "");
707
708 /* Use v_rcp_f32 instead of precise division. */
709 if (!LLVMIsConstant(ret))
710 LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
711 return ret;
712 }
713
714 /* See fast_idiv_by_const.h. */
715 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
716 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
717 LLVMValueRef num,
718 LLVMValueRef multiplier,
719 LLVMValueRef pre_shift,
720 LLVMValueRef post_shift,
721 LLVMValueRef increment)
722 {
723 LLVMBuilderRef builder = ctx->builder;
724
725 num = LLVMBuildLShr(builder, num, pre_shift, "");
726 num = LLVMBuildMul(builder,
727 LLVMBuildZExt(builder, num, ctx->i64, ""),
728 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
729 num = LLVMBuildAdd(builder, num,
730 LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
731 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
732 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
733 return LLVMBuildLShr(builder, num, post_shift, "");
734 }
735
736 /* See fast_idiv_by_const.h. */
737 /* If num != UINT_MAX, this more efficient version can be used. */
738 /* Set: increment = util_fast_udiv_info::increment; */
739 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
740 LLVMValueRef num,
741 LLVMValueRef multiplier,
742 LLVMValueRef pre_shift,
743 LLVMValueRef post_shift,
744 LLVMValueRef increment)
745 {
746 LLVMBuilderRef builder = ctx->builder;
747
748 num = LLVMBuildLShr(builder, num, pre_shift, "");
749 num = LLVMBuildNUWAdd(builder, num, increment, "");
750 num = LLVMBuildMul(builder,
751 LLVMBuildZExt(builder, num, ctx->i64, ""),
752 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
753 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
754 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
755 return LLVMBuildLShr(builder, num, post_shift, "");
756 }
757
758 /* See fast_idiv_by_const.h. */
759 /* Both operands must fit in 31 bits and the divisor must not be 1. */
760 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
761 LLVMValueRef num,
762 LLVMValueRef multiplier,
763 LLVMValueRef post_shift)
764 {
765 LLVMBuilderRef builder = ctx->builder;
766
767 num = LLVMBuildMul(builder,
768 LLVMBuildZExt(builder, num, ctx->i64, ""),
769 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
770 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
771 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
772 return LLVMBuildLShr(builder, num, post_shift, "");
773 }
774
775 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
776 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
777 * already multiplied by two. id is the cube face number.
778 */
779 struct cube_selection_coords {
780 LLVMValueRef stc[2];
781 LLVMValueRef ma;
782 LLVMValueRef id;
783 };
784
785 static void
786 build_cube_intrinsic(struct ac_llvm_context *ctx,
787 LLVMValueRef in[3],
788 struct cube_selection_coords *out)
789 {
790 LLVMTypeRef f32 = ctx->f32;
791
792 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
793 f32, in, 3, AC_FUNC_ATTR_READNONE);
794 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
795 f32, in, 3, AC_FUNC_ATTR_READNONE);
796 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
797 f32, in, 3, AC_FUNC_ATTR_READNONE);
798 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
799 f32, in, 3, AC_FUNC_ATTR_READNONE);
800 }
801
802 /**
803 * Build a manual selection sequence for cube face sc/tc coordinates and
804 * major axis vector (multiplied by 2 for consistency) for the given
805 * vec3 \p coords, for the face implied by \p selcoords.
806 *
807 * For the major axis, we always adjust the sign to be in the direction of
808 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
809 * the selcoords major axis.
810 */
811 static void build_cube_select(struct ac_llvm_context *ctx,
812 const struct cube_selection_coords *selcoords,
813 const LLVMValueRef *coords,
814 LLVMValueRef *out_st,
815 LLVMValueRef *out_ma)
816 {
817 LLVMBuilderRef builder = ctx->builder;
818 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
819 LLVMValueRef is_ma_positive;
820 LLVMValueRef sgn_ma;
821 LLVMValueRef is_ma_z, is_not_ma_z;
822 LLVMValueRef is_ma_y;
823 LLVMValueRef is_ma_x;
824 LLVMValueRef sgn;
825 LLVMValueRef tmp;
826
827 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
828 selcoords->ma, LLVMConstReal(f32, 0.0), "");
829 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
830 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
831
832 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
833 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
834 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
835 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
836 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
837
838 /* Select sc */
839 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
840 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
841 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
842 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
843 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
844
845 /* Select tc */
846 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
847 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
848 LLVMConstReal(f32, -1.0), "");
849 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
850
851 /* Select ma */
852 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
853 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
854 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
855 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
856 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
857 }
858
859 void
860 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
861 bool is_deriv, bool is_array, bool is_lod,
862 LLVMValueRef *coords_arg,
863 LLVMValueRef *derivs_arg)
864 {
865
866 LLVMBuilderRef builder = ctx->builder;
867 struct cube_selection_coords selcoords;
868 LLVMValueRef coords[3];
869 LLVMValueRef invma;
870
871 if (is_array && !is_lod) {
872 LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
873
874 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
875 *
876 * "For Array forms, the array layer used will be
877 *
878 * max(0, min(d−1, floor(layer+0.5)))
879 *
880 * where d is the depth of the texture array and layer
881 * comes from the component indicated in the tables below.
882 * Workaroudn for an issue where the layer is taken from a
883 * helper invocation which happens to fall on a different
884 * layer due to extrapolation."
885 *
886 * GFX8 and earlier attempt to implement this in hardware by
887 * clamping the value of coords[2] = (8 * layer) + face.
888 * Unfortunately, this means that the we end up with the wrong
889 * face when clamping occurs.
890 *
891 * Clamp the layer earlier to work around the issue.
892 */
893 if (ctx->chip_class <= GFX8) {
894 LLVMValueRef ge0;
895 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
896 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
897 }
898
899 coords_arg[3] = tmp;
900 }
901
902 build_cube_intrinsic(ctx, coords_arg, &selcoords);
903
904 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
905 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
906 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
907
908 for (int i = 0; i < 2; ++i)
909 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
910
911 coords[2] = selcoords.id;
912
913 if (is_deriv && derivs_arg) {
914 LLVMValueRef derivs[4];
915 int axis;
916
917 /* Convert cube derivatives to 2D derivatives. */
918 for (axis = 0; axis < 2; axis++) {
919 LLVMValueRef deriv_st[2];
920 LLVMValueRef deriv_ma;
921
922 /* Transform the derivative alongside the texture
923 * coordinate. Mathematically, the correct formula is
924 * as follows. Assume we're projecting onto the +Z face
925 * and denote by dx/dh the derivative of the (original)
926 * X texture coordinate with respect to horizontal
927 * window coordinates. The projection onto the +Z face
928 * plane is:
929 *
930 * f(x,z) = x/z
931 *
932 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
933 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
934 *
935 * This motivatives the implementation below.
936 *
937 * Whether this actually gives the expected results for
938 * apps that might feed in derivatives obtained via
939 * finite differences is anyone's guess. The OpenGL spec
940 * seems awfully quiet about how textureGrad for cube
941 * maps should be handled.
942 */
943 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
944 deriv_st, &deriv_ma);
945
946 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
947
948 for (int i = 0; i < 2; ++i)
949 derivs[axis * 2 + i] =
950 LLVMBuildFSub(builder,
951 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
952 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
953 }
954
955 memcpy(derivs_arg, derivs, sizeof(derivs));
956 }
957
958 /* Shift the texture coordinate. This must be applied after the
959 * derivative calculation.
960 */
961 for (int i = 0; i < 2; ++i)
962 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
963
964 if (is_array) {
965 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
966 /* coords_arg.w component - array_index for cube arrays */
967 coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
968 }
969
970 memcpy(coords_arg, coords, sizeof(coords));
971 }
972
973
974 LLVMValueRef
975 ac_build_fs_interp(struct ac_llvm_context *ctx,
976 LLVMValueRef llvm_chan,
977 LLVMValueRef attr_number,
978 LLVMValueRef params,
979 LLVMValueRef i,
980 LLVMValueRef j)
981 {
982 LLVMValueRef args[5];
983 LLVMValueRef p1;
984
985 args[0] = i;
986 args[1] = llvm_chan;
987 args[2] = attr_number;
988 args[3] = params;
989
990 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
991 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
992
993 args[0] = p1;
994 args[1] = j;
995 args[2] = llvm_chan;
996 args[3] = attr_number;
997 args[4] = params;
998
999 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
1000 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
1001 }
1002
1003 LLVMValueRef
1004 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
1005 LLVMValueRef llvm_chan,
1006 LLVMValueRef attr_number,
1007 LLVMValueRef params,
1008 LLVMValueRef i,
1009 LLVMValueRef j)
1010 {
1011 LLVMValueRef args[6];
1012 LLVMValueRef p1;
1013
1014 args[0] = i;
1015 args[1] = llvm_chan;
1016 args[2] = attr_number;
1017 args[3] = ctx->i1false;
1018 args[4] = params;
1019
1020 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
1021 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
1022
1023 args[0] = p1;
1024 args[1] = j;
1025 args[2] = llvm_chan;
1026 args[3] = attr_number;
1027 args[4] = ctx->i1false;
1028 args[5] = params;
1029
1030 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
1031 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
1032 }
1033
1034 LLVMValueRef
1035 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
1036 LLVMValueRef parameter,
1037 LLVMValueRef llvm_chan,
1038 LLVMValueRef attr_number,
1039 LLVMValueRef params)
1040 {
1041 LLVMValueRef args[4];
1042
1043 args[0] = parameter;
1044 args[1] = llvm_chan;
1045 args[2] = attr_number;
1046 args[3] = params;
1047
1048 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
1049 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
1050 }
1051
1052 LLVMValueRef
1053 ac_build_gep_ptr(struct ac_llvm_context *ctx,
1054 LLVMValueRef base_ptr,
1055 LLVMValueRef index)
1056 {
1057 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1058 }
1059
1060 LLVMValueRef
1061 ac_build_gep0(struct ac_llvm_context *ctx,
1062 LLVMValueRef base_ptr,
1063 LLVMValueRef index)
1064 {
1065 LLVMValueRef indices[2] = {
1066 ctx->i32_0,
1067 index,
1068 };
1069 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1070 }
1071
1072 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
1073 LLVMValueRef index)
1074 {
1075 return LLVMBuildPointerCast(ctx->builder,
1076 LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
1077 LLVMTypeOf(ptr), "");
1078 }
1079
1080 void
1081 ac_build_indexed_store(struct ac_llvm_context *ctx,
1082 LLVMValueRef base_ptr, LLVMValueRef index,
1083 LLVMValueRef value)
1084 {
1085 LLVMBuildStore(ctx->builder, value,
1086 ac_build_gep0(ctx, base_ptr, index));
1087 }
1088
1089 /**
1090 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1091 * It's equivalent to doing a load from &base_ptr[index].
1092 *
1093 * \param base_ptr Where the array starts.
1094 * \param index The element index into the array.
1095 * \param uniform Whether the base_ptr and index can be assumed to be
1096 * dynamically uniform (i.e. load to an SGPR)
1097 * \param invariant Whether the load is invariant (no other opcodes affect it)
1098 * \param no_unsigned_wraparound
1099 * For all possible re-associations and re-distributions of an expression
1100 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1101 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1102 * does not result in an unsigned integer wraparound. This is used for
1103 * optimal code generation of 32-bit pointer arithmetic.
1104 *
1105 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1106 * integer wraparound can't be an imm offset in s_load_dword, because
1107 * the instruction performs "addr + offset" in 64 bits.
1108 *
1109 * Expected usage for bindless textures by chaining GEPs:
1110 * // possible unsigned wraparound, don't use InBounds:
1111 * ptr1 = LLVMBuildGEP(base_ptr, index);
1112 * image = load(ptr1); // becomes "s_load ptr1, 0"
1113 *
1114 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1115 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1116 */
1117 static LLVMValueRef
1118 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1119 LLVMValueRef index, bool uniform, bool invariant,
1120 bool no_unsigned_wraparound)
1121 {
1122 LLVMValueRef pointer, result;
1123
1124 if (no_unsigned_wraparound &&
1125 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1126 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
1127 else
1128 pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1129
1130 if (uniform)
1131 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1132 result = LLVMBuildLoad(ctx->builder, pointer, "");
1133 if (invariant)
1134 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1135 return result;
1136 }
1137
1138 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1139 LLVMValueRef index)
1140 {
1141 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1142 }
1143
1144 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1145 LLVMValueRef base_ptr, LLVMValueRef index)
1146 {
1147 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1148 }
1149
1150 /* This assumes that there is no unsigned integer wraparound during the address
1151 * computation, excluding all GEPs within base_ptr. */
1152 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1153 LLVMValueRef base_ptr, LLVMValueRef index)
1154 {
1155 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1156 }
1157
1158 /* See ac_build_load_custom() documentation. */
1159 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1160 LLVMValueRef base_ptr, LLVMValueRef index)
1161 {
1162 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1163 }
1164
1165 static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
1166 unsigned cache_policy)
1167 {
1168 return cache_policy |
1169 (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
1170 }
1171
1172 static void
1173 ac_build_buffer_store_common(struct ac_llvm_context *ctx,
1174 LLVMValueRef rsrc,
1175 LLVMValueRef data,
1176 LLVMValueRef vindex,
1177 LLVMValueRef voffset,
1178 LLVMValueRef soffset,
1179 unsigned num_channels,
1180 LLVMTypeRef return_channel_type,
1181 unsigned cache_policy,
1182 bool use_format,
1183 bool structurized)
1184 {
1185 LLVMValueRef args[6];
1186 int idx = 0;
1187 args[idx++] = data;
1188 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1189 if (structurized)
1190 args[idx++] = vindex ? vindex : ctx->i32_0;
1191 args[idx++] = voffset ? voffset : ctx->i32_0;
1192 args[idx++] = soffset ? soffset : ctx->i32_0;
1193 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1194 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1195 const char *indexing_kind = structurized ? "struct" : "raw";
1196 char name[256], type_name[8];
1197
1198 LLVMTypeRef type = func > 1 ? LLVMVectorType(return_channel_type, func) : return_channel_type;
1199 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1200
1201 if (use_format) {
1202 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
1203 indexing_kind, type_name);
1204 } else {
1205 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
1206 indexing_kind, type_name);
1207 }
1208
1209 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1210 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1211 }
1212
1213 void
1214 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
1215 LLVMValueRef rsrc,
1216 LLVMValueRef data,
1217 LLVMValueRef vindex,
1218 LLVMValueRef voffset,
1219 unsigned num_channels,
1220 unsigned cache_policy)
1221 {
1222 ac_build_buffer_store_common(ctx, rsrc, data, vindex,
1223 voffset, NULL, num_channels,
1224 ctx->f32, cache_policy,
1225 true, true);
1226 }
1227
1228 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1229 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1230 * or v4i32 (num_channels=3,4).
1231 */
1232 void
1233 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1234 LLVMValueRef rsrc,
1235 LLVMValueRef vdata,
1236 unsigned num_channels,
1237 LLVMValueRef voffset,
1238 LLVMValueRef soffset,
1239 unsigned inst_offset,
1240 unsigned cache_policy,
1241 bool swizzle_enable_hint)
1242 {
1243 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1244 * intrinsics. */
1245 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
1246 LLVMValueRef v[3], v01;
1247
1248 for (int i = 0; i < 3; i++) {
1249 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1250 LLVMConstInt(ctx->i32, i, 0), "");
1251 }
1252 v01 = ac_build_gather_values(ctx, v, 2);
1253
1254 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1255 soffset, inst_offset, cache_policy,
1256 swizzle_enable_hint);
1257 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1258 soffset, inst_offset + 8,
1259 cache_policy,
1260 swizzle_enable_hint);
1261 return;
1262 }
1263
1264 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1265 * (voffset is swizzled, but soffset isn't swizzled).
1266 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1267 */
1268 if (!swizzle_enable_hint) {
1269 LLVMValueRef offset = soffset;
1270
1271 if (inst_offset)
1272 offset = LLVMBuildAdd(ctx->builder, offset,
1273 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1274
1275 ac_build_buffer_store_common(ctx, rsrc, ac_to_float(ctx, vdata),
1276 ctx->i32_0, voffset, offset,
1277 num_channels, ctx->f32,
1278 cache_policy, false, false);
1279 return;
1280 }
1281
1282 static const unsigned dfmts[] = {
1283 V_008F0C_BUF_DATA_FORMAT_32,
1284 V_008F0C_BUF_DATA_FORMAT_32_32,
1285 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1286 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1287 };
1288 unsigned dfmt = dfmts[num_channels - 1];
1289 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1290 LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
1291
1292 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1293 immoffset, num_channels, dfmt, nfmt, cache_policy);
1294 }
1295
1296 static LLVMValueRef
1297 ac_build_buffer_load_common(struct ac_llvm_context *ctx,
1298 LLVMValueRef rsrc,
1299 LLVMValueRef vindex,
1300 LLVMValueRef voffset,
1301 LLVMValueRef soffset,
1302 unsigned num_channels,
1303 LLVMTypeRef channel_type,
1304 unsigned cache_policy,
1305 bool can_speculate,
1306 bool use_format,
1307 bool structurized)
1308 {
1309 LLVMValueRef args[5];
1310 int idx = 0;
1311 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1312 if (structurized)
1313 args[idx++] = vindex ? vindex : ctx->i32_0;
1314 args[idx++] = voffset ? voffset : ctx->i32_0;
1315 args[idx++] = soffset ? soffset : ctx->i32_0;
1316 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1317 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1318 const char *indexing_kind = structurized ? "struct" : "raw";
1319 char name[256], type_name[8];
1320
1321 LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
1322 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1323
1324 if (use_format) {
1325 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1326 indexing_kind, type_name);
1327 } else {
1328 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1329 indexing_kind, type_name);
1330 }
1331
1332 return ac_build_intrinsic(ctx, name, type, args, idx,
1333 ac_get_load_intr_attribs(can_speculate));
1334 }
1335
1336 LLVMValueRef
1337 ac_build_buffer_load(struct ac_llvm_context *ctx,
1338 LLVMValueRef rsrc,
1339 int num_channels,
1340 LLVMValueRef vindex,
1341 LLVMValueRef voffset,
1342 LLVMValueRef soffset,
1343 unsigned inst_offset,
1344 unsigned cache_policy,
1345 bool can_speculate,
1346 bool allow_smem)
1347 {
1348 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1349 if (voffset)
1350 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1351 if (soffset)
1352 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1353
1354 if (allow_smem && !(cache_policy & ac_slc) &&
1355 (!(cache_policy & ac_glc) || ctx->chip_class >= GFX8)) {
1356 assert(vindex == NULL);
1357
1358 LLVMValueRef result[8];
1359
1360 for (int i = 0; i < num_channels; i++) {
1361 if (i) {
1362 offset = LLVMBuildAdd(ctx->builder, offset,
1363 LLVMConstInt(ctx->i32, 4, 0), "");
1364 }
1365 LLVMValueRef args[3] = {
1366 rsrc,
1367 offset,
1368 LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
1369 };
1370 result[i] = ac_build_intrinsic(ctx,
1371 "llvm.amdgcn.s.buffer.load.f32",
1372 ctx->f32, args, 3,
1373 AC_FUNC_ATTR_READNONE);
1374 }
1375 if (num_channels == 1)
1376 return result[0];
1377
1378 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
1379 result[num_channels++] = LLVMGetUndef(ctx->f32);
1380 return ac_build_gather_values(ctx, result, num_channels);
1381 }
1382
1383 return ac_build_buffer_load_common(ctx, rsrc, vindex,
1384 offset, ctx->i32_0,
1385 num_channels, ctx->f32,
1386 cache_policy,
1387 can_speculate, false, false);
1388 }
1389
1390 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1391 LLVMValueRef rsrc,
1392 LLVMValueRef vindex,
1393 LLVMValueRef voffset,
1394 unsigned num_channels,
1395 unsigned cache_policy,
1396 bool can_speculate)
1397 {
1398 return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1399 ctx->i32_0, num_channels, ctx->f32,
1400 cache_policy, can_speculate,
1401 true, true);
1402 }
1403
1404 static LLVMValueRef
1405 ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1406 LLVMValueRef rsrc,
1407 LLVMValueRef vindex,
1408 LLVMValueRef voffset,
1409 LLVMValueRef soffset,
1410 LLVMValueRef immoffset,
1411 unsigned num_channels,
1412 unsigned dfmt,
1413 unsigned nfmt,
1414 unsigned cache_policy,
1415 bool can_speculate,
1416 bool structurized)
1417 {
1418 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1419
1420 LLVMValueRef args[6];
1421 int idx = 0;
1422 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1423 if (structurized)
1424 args[idx++] = vindex ? vindex : ctx->i32_0;
1425 args[idx++] = voffset ? voffset : ctx->i32_0;
1426 args[idx++] = soffset ? soffset : ctx->i32_0;
1427 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
1428 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1429 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1430 const char *indexing_kind = structurized ? "struct" : "raw";
1431 char name[256], type_name[8];
1432
1433 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1434 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1435
1436 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1437 indexing_kind, type_name);
1438
1439 return ac_build_intrinsic(ctx, name, type, args, idx,
1440 ac_get_load_intr_attribs(can_speculate));
1441 }
1442
1443 LLVMValueRef
1444 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1445 LLVMValueRef rsrc,
1446 LLVMValueRef vindex,
1447 LLVMValueRef voffset,
1448 LLVMValueRef soffset,
1449 LLVMValueRef immoffset,
1450 unsigned num_channels,
1451 unsigned dfmt,
1452 unsigned nfmt,
1453 unsigned cache_policy,
1454 bool can_speculate)
1455 {
1456 return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1457 immoffset, num_channels, dfmt, nfmt,
1458 cache_policy, can_speculate, true);
1459 }
1460
1461 LLVMValueRef
1462 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1463 LLVMValueRef rsrc,
1464 LLVMValueRef voffset,
1465 LLVMValueRef soffset,
1466 LLVMValueRef immoffset,
1467 unsigned num_channels,
1468 unsigned dfmt,
1469 unsigned nfmt,
1470 unsigned cache_policy,
1471 bool can_speculate)
1472 {
1473 return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1474 immoffset, num_channels, dfmt, nfmt,
1475 cache_policy, can_speculate, false);
1476 }
1477
1478 LLVMValueRef
1479 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1480 LLVMValueRef rsrc,
1481 LLVMValueRef voffset,
1482 LLVMValueRef soffset,
1483 LLVMValueRef immoffset,
1484 unsigned cache_policy)
1485 {
1486 LLVMValueRef res;
1487
1488 if (LLVM_VERSION_MAJOR >= 9) {
1489 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1490
1491 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1492 res = ac_build_buffer_load_common(ctx, rsrc, NULL,
1493 voffset, soffset,
1494 1, ctx->i16, cache_policy,
1495 false, false, false);
1496 } else {
1497 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1498 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1499
1500 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1501 immoffset, 1, dfmt, nfmt, cache_policy,
1502 false);
1503
1504 res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1505 }
1506
1507 return res;
1508 }
1509
1510 LLVMValueRef
1511 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
1512 LLVMValueRef rsrc,
1513 LLVMValueRef voffset,
1514 LLVMValueRef soffset,
1515 LLVMValueRef immoffset,
1516 unsigned cache_policy)
1517 {
1518 LLVMValueRef res;
1519
1520 if (LLVM_VERSION_MAJOR >= 9) {
1521 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1522
1523 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1524 res = ac_build_buffer_load_common(ctx, rsrc, NULL,
1525 voffset, soffset,
1526 1, ctx->i8, cache_policy,
1527 false, false, false);
1528 } else {
1529 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1530 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1531
1532 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1533 immoffset, 1, dfmt, nfmt, cache_policy,
1534 false);
1535
1536 res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
1537 }
1538
1539 return res;
1540 }
1541
1542 /**
1543 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1544 *
1545 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1546 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1547 */
1548 static LLVMValueRef
1549 ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits, unsigned mant_bits)
1550 {
1551 assert(LLVMTypeOf(src) == ctx->i32);
1552
1553 LLVMValueRef tmp;
1554 LLVMValueRef mantissa;
1555 mantissa = LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
1556
1557 /* Converting normal numbers is just a shift + correcting the exponent bias */
1558 unsigned normal_shift = 23 - mant_bits;
1559 unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
1560 LLVMValueRef shifted, normal;
1561
1562 shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
1563 normal = LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
1564
1565 /* Converting nan/inf numbers is the same, but with a different exponent update */
1566 LLVMValueRef naninf;
1567 naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
1568
1569 /* Converting denormals is the complex case: determine the leading zeros of the
1570 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1571 */
1572 LLVMValueRef denormal;
1573 LLVMValueRef params[2] = {
1574 mantissa,
1575 ctx->i1true, /* result can be undef when arg is 0 */
1576 };
1577 LLVMValueRef ctlz = ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32,
1578 params, 2, AC_FUNC_ATTR_READNONE);
1579
1580 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1581 tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
1582 denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
1583
1584 unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
1585 tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
1586 tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
1587 denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
1588
1589 /* Select the final result. */
1590 LLVMValueRef result;
1591
1592 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1593 LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
1594 result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
1595
1596 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1597 LLVMConstInt(ctx->i32, 1 << mant_bits, false), "");
1598 result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
1599
1600 tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
1601 result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
1602
1603 return ac_to_float(ctx, result);
1604 }
1605
1606 /**
1607 * Generate a fully general open coded buffer format fetch with all required
1608 * fixups suitable for vertex fetch, using non-format buffer loads.
1609 *
1610 * Some combinations of argument values have special interpretations:
1611 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1612 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1613 *
1614 * \param log_size log(size of channel in bytes)
1615 * \param num_channels number of channels (1 to 4)
1616 * \param format AC_FETCH_FORMAT_xxx value
1617 * \param reverse whether XYZ channels are reversed
1618 * \param known_aligned whether the source is known to be aligned to hardware's
1619 * effective element size for loading the given format
1620 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1621 * \param rsrc buffer resource descriptor
1622 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1623 */
1624 LLVMValueRef
1625 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
1626 unsigned log_size,
1627 unsigned num_channels,
1628 unsigned format,
1629 bool reverse,
1630 bool known_aligned,
1631 LLVMValueRef rsrc,
1632 LLVMValueRef vindex,
1633 LLVMValueRef voffset,
1634 LLVMValueRef soffset,
1635 unsigned cache_policy,
1636 bool can_speculate)
1637 {
1638 LLVMValueRef tmp;
1639 unsigned load_log_size = log_size;
1640 unsigned load_num_channels = num_channels;
1641 if (log_size == 3) {
1642 load_log_size = 2;
1643 if (format == AC_FETCH_FORMAT_FLOAT) {
1644 load_num_channels = 2 * num_channels;
1645 } else {
1646 load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
1647 }
1648 }
1649
1650 int log_recombine = 0;
1651 if (ctx->chip_class == GFX6 && !known_aligned) {
1652 /* Avoid alignment restrictions by loading one byte at a time. */
1653 load_num_channels <<= load_log_size;
1654 log_recombine = load_log_size;
1655 load_log_size = 0;
1656 } else if (load_num_channels == 2 || load_num_channels == 4) {
1657 log_recombine = -util_logbase2(load_num_channels);
1658 load_num_channels = 1;
1659 load_log_size += -log_recombine;
1660 }
1661
1662 assert(load_log_size >= 2 || LLVM_VERSION_MAJOR >= 9);
1663
1664 LLVMValueRef loads[32]; /* up to 32 bytes */
1665 for (unsigned i = 0; i < load_num_channels; ++i) {
1666 tmp = LLVMBuildAdd(ctx->builder, soffset,
1667 LLVMConstInt(ctx->i32, i << load_log_size, false), "");
1668 LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
1669 load_log_size == 1 ? ctx->i16 : ctx->i32;
1670 unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
1671 loads[i] = ac_build_buffer_load_common(
1672 ctx, rsrc, vindex, voffset, tmp,
1673 num_channels, channel_type, cache_policy,
1674 can_speculate, false, true);
1675 if (load_log_size >= 2)
1676 loads[i] = ac_to_integer(ctx, loads[i]);
1677 }
1678
1679 if (log_recombine > 0) {
1680 /* Recombine bytes if necessary (GFX6 only) */
1681 LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
1682
1683 for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
1684 LLVMValueRef accum = NULL;
1685 for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
1686 tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
1687 if (i == 0) {
1688 accum = tmp;
1689 } else {
1690 tmp = LLVMBuildShl(ctx->builder, tmp,
1691 LLVMConstInt(dst_type, 8 * i, false), "");
1692 accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
1693 }
1694 }
1695 loads[dst] = accum;
1696 }
1697 } else if (log_recombine < 0) {
1698 /* Split vectors of dwords */
1699 if (load_log_size > 2) {
1700 assert(load_num_channels == 1);
1701 LLVMValueRef loaded = loads[0];
1702 unsigned log_split = load_log_size - 2;
1703 log_recombine += log_split;
1704 load_num_channels = 1 << log_split;
1705 load_log_size = 2;
1706 for (unsigned i = 0; i < load_num_channels; ++i) {
1707 tmp = LLVMConstInt(ctx->i32, i, false);
1708 loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
1709 }
1710 }
1711
1712 /* Further split dwords and shorts if required */
1713 if (log_recombine < 0) {
1714 for (unsigned src = load_num_channels,
1715 dst = load_num_channels << -log_recombine;
1716 src > 0; --src) {
1717 unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
1718 LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
1719 LLVMValueRef loaded = loads[src - 1];
1720 LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
1721 for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
1722 tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
1723 tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
1724 loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
1725 }
1726 }
1727 }
1728 }
1729
1730 if (log_size == 3) {
1731 if (format == AC_FETCH_FORMAT_FLOAT) {
1732 for (unsigned i = 0; i < num_channels; ++i) {
1733 tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
1734 loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
1735 }
1736 } else if (format == AC_FETCH_FORMAT_FIXED) {
1737 /* 10_11_11_FLOAT */
1738 LLVMValueRef data = loads[0];
1739 LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
1740 LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
1741 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
1742 LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
1743 LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
1744
1745 loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
1746 loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
1747 loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
1748
1749 num_channels = 3;
1750 log_size = 2;
1751 format = AC_FETCH_FORMAT_FLOAT;
1752 } else {
1753 /* 2_10_10_10 data formats */
1754 LLVMValueRef data = loads[0];
1755 LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
1756 LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
1757 loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
1758 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
1759 loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1760 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
1761 loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1762 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
1763 loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
1764
1765 num_channels = 4;
1766 }
1767 }
1768
1769 if (format == AC_FETCH_FORMAT_FLOAT) {
1770 if (log_size != 2) {
1771 for (unsigned chan = 0; chan < num_channels; ++chan) {
1772 tmp = ac_to_float(ctx, loads[chan]);
1773 if (log_size == 3)
1774 tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
1775 else if (log_size == 1)
1776 tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
1777 loads[chan] = ac_to_integer(ctx, tmp);
1778 }
1779 }
1780 } else if (format == AC_FETCH_FORMAT_UINT) {
1781 if (log_size != 2) {
1782 for (unsigned chan = 0; chan < num_channels; ++chan)
1783 loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
1784 }
1785 } else if (format == AC_FETCH_FORMAT_SINT) {
1786 if (log_size != 2) {
1787 for (unsigned chan = 0; chan < num_channels; ++chan)
1788 loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
1789 }
1790 } else {
1791 bool unsign = format == AC_FETCH_FORMAT_UNORM ||
1792 format == AC_FETCH_FORMAT_USCALED ||
1793 format == AC_FETCH_FORMAT_UINT;
1794
1795 for (unsigned chan = 0; chan < num_channels; ++chan) {
1796 if (unsign) {
1797 tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
1798 } else {
1799 tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
1800 }
1801
1802 LLVMValueRef scale = NULL;
1803 if (format == AC_FETCH_FORMAT_FIXED) {
1804 assert(log_size == 2);
1805 scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
1806 } else if (format == AC_FETCH_FORMAT_UNORM) {
1807 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1808 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
1809 } else if (format == AC_FETCH_FORMAT_SNORM) {
1810 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1811 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
1812 }
1813 if (scale)
1814 tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
1815
1816 if (format == AC_FETCH_FORMAT_SNORM) {
1817 /* Clamp to [-1, 1] */
1818 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
1819 LLVMValueRef clamp =
1820 LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
1821 tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
1822 }
1823
1824 loads[chan] = ac_to_integer(ctx, tmp);
1825 }
1826 }
1827
1828 while (num_channels < 4) {
1829 if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
1830 loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
1831 } else {
1832 loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
1833 }
1834 num_channels++;
1835 }
1836
1837 if (reverse) {
1838 tmp = loads[0];
1839 loads[0] = loads[2];
1840 loads[2] = tmp;
1841 }
1842
1843 return ac_build_gather_values(ctx, loads, 4);
1844 }
1845
1846 static void
1847 ac_build_tbuffer_store(struct ac_llvm_context *ctx,
1848 LLVMValueRef rsrc,
1849 LLVMValueRef vdata,
1850 LLVMValueRef vindex,
1851 LLVMValueRef voffset,
1852 LLVMValueRef soffset,
1853 LLVMValueRef immoffset,
1854 unsigned num_channels,
1855 unsigned dfmt,
1856 unsigned nfmt,
1857 unsigned cache_policy,
1858 bool structurized)
1859 {
1860 voffset = LLVMBuildAdd(ctx->builder, voffset ? voffset : ctx->i32_0,
1861 immoffset, "");
1862
1863 LLVMValueRef args[7];
1864 int idx = 0;
1865 args[idx++] = vdata;
1866 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1867 if (structurized)
1868 args[idx++] = vindex ? vindex : ctx->i32_0;
1869 args[idx++] = voffset ? voffset : ctx->i32_0;
1870 args[idx++] = soffset ? soffset : ctx->i32_0;
1871 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
1872 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1873 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1874 const char *indexing_kind = structurized ? "struct" : "raw";
1875 char name[256], type_name[8];
1876
1877 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1878 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1879
1880 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
1881 indexing_kind, type_name);
1882
1883 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1884 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1885 }
1886
1887 void
1888 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
1889 LLVMValueRef rsrc,
1890 LLVMValueRef vdata,
1891 LLVMValueRef vindex,
1892 LLVMValueRef voffset,
1893 LLVMValueRef soffset,
1894 LLVMValueRef immoffset,
1895 unsigned num_channels,
1896 unsigned dfmt,
1897 unsigned nfmt,
1898 unsigned cache_policy)
1899 {
1900 ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
1901 immoffset, num_channels, dfmt, nfmt, cache_policy,
1902 true);
1903 }
1904
1905 void
1906 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
1907 LLVMValueRef rsrc,
1908 LLVMValueRef vdata,
1909 LLVMValueRef voffset,
1910 LLVMValueRef soffset,
1911 LLVMValueRef immoffset,
1912 unsigned num_channels,
1913 unsigned dfmt,
1914 unsigned nfmt,
1915 unsigned cache_policy)
1916 {
1917 ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
1918 immoffset, num_channels, dfmt, nfmt, cache_policy,
1919 false);
1920 }
1921
1922 void
1923 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
1924 LLVMValueRef rsrc,
1925 LLVMValueRef vdata,
1926 LLVMValueRef voffset,
1927 LLVMValueRef soffset,
1928 unsigned cache_policy)
1929 {
1930 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
1931
1932 if (LLVM_VERSION_MAJOR >= 9) {
1933 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1934 ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
1935 voffset, soffset, 1,
1936 ctx->i16, cache_policy,
1937 false, false);
1938 } else {
1939 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1940 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1941
1942 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1943
1944 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1945 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
1946 }
1947 }
1948
1949 void
1950 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
1951 LLVMValueRef rsrc,
1952 LLVMValueRef vdata,
1953 LLVMValueRef voffset,
1954 LLVMValueRef soffset,
1955 unsigned cache_policy)
1956 {
1957 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
1958
1959 if (LLVM_VERSION_MAJOR >= 9) {
1960 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1961 ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
1962 voffset, soffset, 1,
1963 ctx->i8, cache_policy,
1964 false, false);
1965 } else {
1966 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1967 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1968
1969 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1970
1971 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1972 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
1973 }
1974 }
1975 /**
1976 * Set range metadata on an instruction. This can only be used on load and
1977 * call instructions. If you know an instruction can only produce the values
1978 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1979 * \p lo is the minimum value inclusive.
1980 * \p hi is the maximum value exclusive.
1981 */
1982 static void set_range_metadata(struct ac_llvm_context *ctx,
1983 LLVMValueRef value, unsigned lo, unsigned hi)
1984 {
1985 LLVMValueRef range_md, md_args[2];
1986 LLVMTypeRef type = LLVMTypeOf(value);
1987 LLVMContextRef context = LLVMGetTypeContext(type);
1988
1989 md_args[0] = LLVMConstInt(type, lo, false);
1990 md_args[1] = LLVMConstInt(type, hi, false);
1991 range_md = LLVMMDNodeInContext(context, md_args, 2);
1992 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1993 }
1994
1995 LLVMValueRef
1996 ac_get_thread_id(struct ac_llvm_context *ctx)
1997 {
1998 LLVMValueRef tid;
1999
2000 LLVMValueRef tid_args[2];
2001 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
2002 tid_args[1] = ctx->i32_0;
2003 tid_args[1] = ac_build_intrinsic(ctx,
2004 "llvm.amdgcn.mbcnt.lo", ctx->i32,
2005 tid_args, 2, AC_FUNC_ATTR_READNONE);
2006
2007 if (ctx->wave_size == 32) {
2008 tid = tid_args[1];
2009 } else {
2010 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
2011 ctx->i32, tid_args,
2012 2, AC_FUNC_ATTR_READNONE);
2013 }
2014 set_range_metadata(ctx, tid, 0, ctx->wave_size);
2015 return tid;
2016 }
2017
2018 /*
2019 * AMD GCN implements derivatives using the local data store (LDS)
2020 * All writes to the LDS happen in all executing threads at
2021 * the same time. TID is the Thread ID for the current
2022 * thread and is a value between 0 and 63, representing
2023 * the thread's position in the wavefront.
2024 *
2025 * For the pixel shader threads are grouped into quads of four pixels.
2026 * The TIDs of the pixels of a quad are:
2027 *
2028 * +------+------+
2029 * |4n + 0|4n + 1|
2030 * +------+------+
2031 * |4n + 2|4n + 3|
2032 * +------+------+
2033 *
2034 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2035 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2036 * the current pixel's column, and masking with 0xfffffffe yields the TID
2037 * of the left pixel of the current pixel's row.
2038 *
2039 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2040 * adding 2 yields the TID of the pixel below the top pixel.
2041 */
2042 LLVMValueRef
2043 ac_build_ddxy(struct ac_llvm_context *ctx,
2044 uint32_t mask,
2045 int idx,
2046 LLVMValueRef val)
2047 {
2048 unsigned tl_lanes[4], trbl_lanes[4];
2049 char name[32], type[8];
2050 LLVMValueRef tl, trbl;
2051 LLVMTypeRef result_type;
2052 LLVMValueRef result;
2053
2054 result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
2055
2056 if (result_type == ctx->f16)
2057 val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
2058
2059 for (unsigned i = 0; i < 4; ++i) {
2060 tl_lanes[i] = i & mask;
2061 trbl_lanes[i] = (i & mask) + idx;
2062 }
2063
2064 tl = ac_build_quad_swizzle(ctx, val,
2065 tl_lanes[0], tl_lanes[1],
2066 tl_lanes[2], tl_lanes[3]);
2067 trbl = ac_build_quad_swizzle(ctx, val,
2068 trbl_lanes[0], trbl_lanes[1],
2069 trbl_lanes[2], trbl_lanes[3]);
2070
2071 if (result_type == ctx->f16) {
2072 tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
2073 trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
2074 }
2075
2076 tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
2077 trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
2078 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
2079
2080 ac_build_type_name_for_intr(result_type, type, sizeof(type));
2081 snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
2082
2083 return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
2084 }
2085
2086 void
2087 ac_build_sendmsg(struct ac_llvm_context *ctx,
2088 uint32_t msg,
2089 LLVMValueRef wave_id)
2090 {
2091 LLVMValueRef args[2];
2092 args[0] = LLVMConstInt(ctx->i32, msg, false);
2093 args[1] = wave_id;
2094 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
2095 }
2096
2097 LLVMValueRef
2098 ac_build_imsb(struct ac_llvm_context *ctx,
2099 LLVMValueRef arg,
2100 LLVMTypeRef dst_type)
2101 {
2102 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
2103 dst_type, &arg, 1,
2104 AC_FUNC_ATTR_READNONE);
2105
2106 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2107 * the index from LSB. Invert it by doing "31 - msb". */
2108 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
2109 msb, "");
2110
2111 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
2112 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
2113 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2114 arg, ctx->i32_0, ""),
2115 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2116 arg, all_ones, ""), "");
2117
2118 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
2119 }
2120
2121 LLVMValueRef
2122 ac_build_umsb(struct ac_llvm_context *ctx,
2123 LLVMValueRef arg,
2124 LLVMTypeRef dst_type)
2125 {
2126 const char *intrin_name;
2127 LLVMTypeRef type;
2128 LLVMValueRef highest_bit;
2129 LLVMValueRef zero;
2130 unsigned bitsize;
2131
2132 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
2133 switch (bitsize) {
2134 case 64:
2135 intrin_name = "llvm.ctlz.i64";
2136 type = ctx->i64;
2137 highest_bit = LLVMConstInt(ctx->i64, 63, false);
2138 zero = ctx->i64_0;
2139 break;
2140 case 32:
2141 intrin_name = "llvm.ctlz.i32";
2142 type = ctx->i32;
2143 highest_bit = LLVMConstInt(ctx->i32, 31, false);
2144 zero = ctx->i32_0;
2145 break;
2146 case 16:
2147 intrin_name = "llvm.ctlz.i16";
2148 type = ctx->i16;
2149 highest_bit = LLVMConstInt(ctx->i16, 15, false);
2150 zero = ctx->i16_0;
2151 break;
2152 case 8:
2153 intrin_name = "llvm.ctlz.i8";
2154 type = ctx->i8;
2155 highest_bit = LLVMConstInt(ctx->i8, 7, false);
2156 zero = ctx->i8_0;
2157 break;
2158 default:
2159 unreachable(!"invalid bitsize");
2160 break;
2161 }
2162
2163 LLVMValueRef params[2] = {
2164 arg,
2165 ctx->i1true,
2166 };
2167
2168 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
2169 params, 2,
2170 AC_FUNC_ATTR_READNONE);
2171
2172 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2173 * the index from LSB. Invert it by doing "31 - msb". */
2174 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
2175
2176 if (bitsize == 64) {
2177 msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
2178 } else if (bitsize < 32) {
2179 msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
2180 }
2181
2182 /* check for zero */
2183 return LLVMBuildSelect(ctx->builder,
2184 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
2185 LLVMConstInt(ctx->i32, -1, true), msb, "");
2186 }
2187
2188 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
2189 LLVMValueRef b)
2190 {
2191 char name[64];
2192 snprintf(name, sizeof(name), "llvm.minnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2193 LLVMValueRef args[2] = {a, b};
2194 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2195 AC_FUNC_ATTR_READNONE);
2196 }
2197
2198 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
2199 LLVMValueRef b)
2200 {
2201 char name[64];
2202 snprintf(name, sizeof(name), "llvm.maxnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2203 LLVMValueRef args[2] = {a, b};
2204 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2205 AC_FUNC_ATTR_READNONE);
2206 }
2207
2208 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
2209 LLVMValueRef b)
2210 {
2211 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
2212 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2213 }
2214
2215 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
2216 LLVMValueRef b)
2217 {
2218 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
2219 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2220 }
2221
2222 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
2223 LLVMValueRef b)
2224 {
2225 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
2226 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2227 }
2228
2229 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
2230 LLVMValueRef b)
2231 {
2232 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
2233 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2234 }
2235
2236 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
2237 {
2238 LLVMTypeRef t = LLVMTypeOf(value);
2239 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
2240 LLVMConstReal(t, 1.0));
2241 }
2242
2243 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
2244 {
2245 LLVMValueRef args[9];
2246
2247 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
2248 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
2249
2250 if (a->compr) {
2251 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
2252 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
2253
2254 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
2255 v2i16, "");
2256 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
2257 v2i16, "");
2258 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
2259 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2260
2261 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
2262 ctx->voidt, args, 6, 0);
2263 } else {
2264 args[2] = a->out[0];
2265 args[3] = a->out[1];
2266 args[4] = a->out[2];
2267 args[5] = a->out[3];
2268 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
2269 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2270
2271 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
2272 ctx->voidt, args, 8, 0);
2273 }
2274 }
2275
2276 void ac_build_export_null(struct ac_llvm_context *ctx)
2277 {
2278 struct ac_export_args args;
2279
2280 args.enabled_channels = 0x0; /* enabled channels */
2281 args.valid_mask = 1; /* whether the EXEC mask is valid */
2282 args.done = 1; /* DONE bit */
2283 args.target = V_008DFC_SQ_EXP_NULL;
2284 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
2285 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
2286 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
2287 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
2288 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
2289
2290 ac_build_export(ctx, &args);
2291 }
2292
2293 static unsigned ac_num_coords(enum ac_image_dim dim)
2294 {
2295 switch (dim) {
2296 case ac_image_1d:
2297 return 1;
2298 case ac_image_2d:
2299 case ac_image_1darray:
2300 return 2;
2301 case ac_image_3d:
2302 case ac_image_cube:
2303 case ac_image_2darray:
2304 case ac_image_2dmsaa:
2305 return 3;
2306 case ac_image_2darraymsaa:
2307 return 4;
2308 default:
2309 unreachable("ac_num_coords: bad dim");
2310 }
2311 }
2312
2313 static unsigned ac_num_derivs(enum ac_image_dim dim)
2314 {
2315 switch (dim) {
2316 case ac_image_1d:
2317 case ac_image_1darray:
2318 return 2;
2319 case ac_image_2d:
2320 case ac_image_2darray:
2321 case ac_image_cube:
2322 return 4;
2323 case ac_image_3d:
2324 return 6;
2325 case ac_image_2dmsaa:
2326 case ac_image_2darraymsaa:
2327 default:
2328 unreachable("derivatives not supported");
2329 }
2330 }
2331
2332 static const char *get_atomic_name(enum ac_atomic_op op)
2333 {
2334 switch (op) {
2335 case ac_atomic_swap: return "swap";
2336 case ac_atomic_add: return "add";
2337 case ac_atomic_sub: return "sub";
2338 case ac_atomic_smin: return "smin";
2339 case ac_atomic_umin: return "umin";
2340 case ac_atomic_smax: return "smax";
2341 case ac_atomic_umax: return "umax";
2342 case ac_atomic_and: return "and";
2343 case ac_atomic_or: return "or";
2344 case ac_atomic_xor: return "xor";
2345 case ac_atomic_inc_wrap: return "inc";
2346 case ac_atomic_dec_wrap: return "dec";
2347 }
2348 unreachable("bad atomic op");
2349 }
2350
2351 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
2352 struct ac_image_args *a)
2353 {
2354 const char *overload[3] = { "", "", "" };
2355 unsigned num_overloads = 0;
2356 LLVMValueRef args[18];
2357 unsigned num_args = 0;
2358 enum ac_image_dim dim = a->dim;
2359
2360 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
2361 !a->level_zero);
2362 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
2363 a->opcode != ac_image_store_mip) ||
2364 a->lod);
2365 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2366 (!a->compare && !a->offset));
2367 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2368 a->opcode == ac_image_get_lod) ||
2369 !a->bias);
2370 assert((a->bias ? 1 : 0) +
2371 (a->lod ? 1 : 0) +
2372 (a->level_zero ? 1 : 0) +
2373 (a->derivs[0] ? 1 : 0) <= 1);
2374
2375 if (a->opcode == ac_image_get_lod) {
2376 switch (dim) {
2377 case ac_image_1darray:
2378 dim = ac_image_1d;
2379 break;
2380 case ac_image_2darray:
2381 case ac_image_cube:
2382 dim = ac_image_2d;
2383 break;
2384 default:
2385 break;
2386 }
2387 }
2388
2389 bool sample = a->opcode == ac_image_sample ||
2390 a->opcode == ac_image_gather4 ||
2391 a->opcode == ac_image_get_lod;
2392 bool atomic = a->opcode == ac_image_atomic ||
2393 a->opcode == ac_image_atomic_cmpswap;
2394 bool load = a->opcode == ac_image_sample ||
2395 a->opcode == ac_image_gather4 ||
2396 a->opcode == ac_image_load ||
2397 a->opcode == ac_image_load_mip;
2398 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
2399
2400 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
2401 args[num_args++] = a->data[0];
2402 if (a->opcode == ac_image_atomic_cmpswap)
2403 args[num_args++] = a->data[1];
2404 }
2405
2406 if (!atomic)
2407 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
2408
2409 if (a->offset)
2410 args[num_args++] = ac_to_integer(ctx, a->offset);
2411 if (a->bias) {
2412 args[num_args++] = ac_to_float(ctx, a->bias);
2413 overload[num_overloads++] = ".f32";
2414 }
2415 if (a->compare)
2416 args[num_args++] = ac_to_float(ctx, a->compare);
2417 if (a->derivs[0]) {
2418 unsigned count = ac_num_derivs(dim);
2419 for (unsigned i = 0; i < count; ++i)
2420 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
2421 overload[num_overloads++] = ".f32";
2422 }
2423 unsigned num_coords =
2424 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
2425 for (unsigned i = 0; i < num_coords; ++i)
2426 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
2427 if (a->lod)
2428 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
2429 overload[num_overloads++] = sample ? ".f32" : ".i32";
2430
2431 args[num_args++] = a->resource;
2432 if (sample) {
2433 args[num_args++] = a->sampler;
2434 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
2435 }
2436
2437 args[num_args++] = ctx->i32_0; /* texfailctrl */
2438 args[num_args++] = LLVMConstInt(ctx->i32,
2439 load ? get_load_cache_policy(ctx, a->cache_policy) :
2440 a->cache_policy, false);
2441
2442 const char *name;
2443 const char *atomic_subop = "";
2444 switch (a->opcode) {
2445 case ac_image_sample: name = "sample"; break;
2446 case ac_image_gather4: name = "gather4"; break;
2447 case ac_image_load: name = "load"; break;
2448 case ac_image_load_mip: name = "load.mip"; break;
2449 case ac_image_store: name = "store"; break;
2450 case ac_image_store_mip: name = "store.mip"; break;
2451 case ac_image_atomic:
2452 name = "atomic.";
2453 atomic_subop = get_atomic_name(a->atomic);
2454 break;
2455 case ac_image_atomic_cmpswap:
2456 name = "atomic.";
2457 atomic_subop = "cmpswap";
2458 break;
2459 case ac_image_get_lod: name = "getlod"; break;
2460 case ac_image_get_resinfo: name = "getresinfo"; break;
2461 default: unreachable("invalid image opcode");
2462 }
2463
2464 const char *dimname;
2465 switch (dim) {
2466 case ac_image_1d: dimname = "1d"; break;
2467 case ac_image_2d: dimname = "2d"; break;
2468 case ac_image_3d: dimname = "3d"; break;
2469 case ac_image_cube: dimname = "cube"; break;
2470 case ac_image_1darray: dimname = "1darray"; break;
2471 case ac_image_2darray: dimname = "2darray"; break;
2472 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
2473 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
2474 default: unreachable("invalid dim");
2475 }
2476
2477 bool lod_suffix =
2478 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
2479 char intr_name[96];
2480 snprintf(intr_name, sizeof(intr_name),
2481 "llvm.amdgcn.image.%s%s" /* base name */
2482 "%s%s%s" /* sample/gather modifiers */
2483 ".%s.%s%s%s%s", /* dimension and type overloads */
2484 name, atomic_subop,
2485 a->compare ? ".c" : "",
2486 a->bias ? ".b" :
2487 lod_suffix ? ".l" :
2488 a->derivs[0] ? ".d" :
2489 a->level_zero ? ".lz" : "",
2490 a->offset ? ".o" : "",
2491 dimname,
2492 atomic ? "i32" : "v4f32",
2493 overload[0], overload[1], overload[2]);
2494
2495 LLVMTypeRef retty;
2496 if (atomic)
2497 retty = ctx->i32;
2498 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
2499 retty = ctx->voidt;
2500 else
2501 retty = ctx->v4f32;
2502
2503 LLVMValueRef result =
2504 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2505 a->attributes);
2506 if (!sample && retty == ctx->v4f32) {
2507 result = LLVMBuildBitCast(ctx->builder, result,
2508 ctx->v4i32, "");
2509 }
2510 return result;
2511 }
2512
2513 LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
2514 LLVMValueRef rsrc)
2515 {
2516 LLVMValueRef samples;
2517
2518 /* Read the samples from the descriptor directly.
2519 * Hardware doesn't have any instruction for this.
2520 */
2521 samples = LLVMBuildExtractElement(ctx->builder, rsrc,
2522 LLVMConstInt(ctx->i32, 3, 0), "");
2523 samples = LLVMBuildLShr(ctx->builder, samples,
2524 LLVMConstInt(ctx->i32, 16, 0), "");
2525 samples = LLVMBuildAnd(ctx->builder, samples,
2526 LLVMConstInt(ctx->i32, 0xf, 0), "");
2527 samples = LLVMBuildShl(ctx->builder, ctx->i32_1,
2528 samples, "");
2529 return samples;
2530 }
2531
2532 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2533 LLVMValueRef args[2])
2534 {
2535 LLVMTypeRef v2f16 =
2536 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
2537
2538 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", v2f16,
2539 args, 2, AC_FUNC_ATTR_READNONE);
2540 }
2541
2542 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2543 LLVMValueRef args[2])
2544 {
2545 LLVMValueRef res =
2546 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2547 ctx->v2i16, args, 2,
2548 AC_FUNC_ATTR_READNONE);
2549 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2550 }
2551
2552 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2553 LLVMValueRef args[2])
2554 {
2555 LLVMValueRef res =
2556 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2557 ctx->v2i16, args, 2,
2558 AC_FUNC_ATTR_READNONE);
2559 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2560 }
2561
2562 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2563 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2564 LLVMValueRef args[2], unsigned bits, bool hi)
2565 {
2566 assert(bits == 8 || bits == 10 || bits == 16);
2567
2568 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2569 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2570 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2571 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2572 LLVMValueRef max_alpha =
2573 bits != 10 ? max_rgb : ctx->i32_1;
2574 LLVMValueRef min_alpha =
2575 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2576
2577 /* Clamp. */
2578 if (bits != 16) {
2579 for (int i = 0; i < 2; i++) {
2580 bool alpha = hi && i == 1;
2581 args[i] = ac_build_imin(ctx, args[i],
2582 alpha ? max_alpha : max_rgb);
2583 args[i] = ac_build_imax(ctx, args[i],
2584 alpha ? min_alpha : min_rgb);
2585 }
2586 }
2587
2588 LLVMValueRef res =
2589 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2590 ctx->v2i16, args, 2,
2591 AC_FUNC_ATTR_READNONE);
2592 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2593 }
2594
2595 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2596 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2597 LLVMValueRef args[2], unsigned bits, bool hi)
2598 {
2599 assert(bits == 8 || bits == 10 || bits == 16);
2600
2601 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2602 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2603 LLVMValueRef max_alpha =
2604 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2605
2606 /* Clamp. */
2607 if (bits != 16) {
2608 for (int i = 0; i < 2; i++) {
2609 bool alpha = hi && i == 1;
2610 args[i] = ac_build_umin(ctx, args[i],
2611 alpha ? max_alpha : max_rgb);
2612 }
2613 }
2614
2615 LLVMValueRef res =
2616 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2617 ctx->v2i16, args, 2,
2618 AC_FUNC_ATTR_READNONE);
2619 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2620 }
2621
2622 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2623 {
2624 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2625 &i1, 1, AC_FUNC_ATTR_READNONE);
2626 }
2627
2628 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2629 {
2630 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2631 &i1, 1, 0);
2632 }
2633
2634 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2635 LLVMValueRef offset, LLVMValueRef width,
2636 bool is_signed)
2637 {
2638 LLVMValueRef args[] = {
2639 input,
2640 offset,
2641 width,
2642 };
2643
2644 return ac_build_intrinsic(ctx, is_signed ? "llvm.amdgcn.sbfe.i32" :
2645 "llvm.amdgcn.ubfe.i32",
2646 ctx->i32, args, 3, AC_FUNC_ATTR_READNONE);
2647
2648 }
2649
2650 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2651 LLVMValueRef s1, LLVMValueRef s2)
2652 {
2653 return LLVMBuildAdd(ctx->builder,
2654 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2655 }
2656
2657 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2658 LLVMValueRef s1, LLVMValueRef s2)
2659 {
2660 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
2661 if (ctx->chip_class >= GFX10) {
2662 return ac_build_intrinsic(ctx, "llvm.fma.f32", ctx->f32,
2663 (LLVMValueRef []) {s0, s1, s2}, 3,
2664 AC_FUNC_ATTR_READNONE);
2665 }
2666
2667 return LLVMBuildFAdd(ctx->builder,
2668 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2669 }
2670
2671 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags)
2672 {
2673 if (!wait_flags)
2674 return;
2675
2676 unsigned lgkmcnt = 63;
2677 unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
2678 unsigned vscnt = 63;
2679
2680 if (wait_flags & AC_WAIT_LGKM)
2681 lgkmcnt = 0;
2682 if (wait_flags & AC_WAIT_VLOAD)
2683 vmcnt = 0;
2684
2685 if (wait_flags & AC_WAIT_VSTORE) {
2686 if (ctx->chip_class >= GFX10)
2687 vscnt = 0;
2688 else
2689 vmcnt = 0;
2690 }
2691
2692 /* There is no intrinsic for vscnt(0), so use a fence. */
2693 if ((wait_flags & AC_WAIT_LGKM &&
2694 wait_flags & AC_WAIT_VLOAD &&
2695 wait_flags & AC_WAIT_VSTORE) ||
2696 vscnt == 0) {
2697 LLVMBuildFence(ctx->builder, LLVMAtomicOrderingRelease, false, "");
2698 return;
2699 }
2700
2701 unsigned simm16 = (lgkmcnt << 8) |
2702 (7 << 4) | /* expcnt */
2703 (vmcnt & 0xf) |
2704 ((vmcnt >> 4) << 14);
2705
2706 LLVMValueRef args[1] = {
2707 LLVMConstInt(ctx->i32, simm16, false),
2708 };
2709 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2710 ctx->voidt, args, 1, 0);
2711 }
2712
2713 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
2714 LLVMValueRef src1, LLVMValueRef src2,
2715 unsigned bitsize)
2716 {
2717 LLVMTypeRef type;
2718 char *intr;
2719
2720 if (bitsize == 16) {
2721 intr = "llvm.amdgcn.fmed3.f16";
2722 type = ctx->f16;
2723 } else if (bitsize == 32) {
2724 intr = "llvm.amdgcn.fmed3.f32";
2725 type = ctx->f32;
2726 } else {
2727 intr = "llvm.amdgcn.fmed3.f64";
2728 type = ctx->f64;
2729 }
2730
2731 LLVMValueRef params[] = {
2732 src0,
2733 src1,
2734 src2,
2735 };
2736 return ac_build_intrinsic(ctx, intr, type, params, 3,
2737 AC_FUNC_ATTR_READNONE);
2738 }
2739
2740 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2741 unsigned bitsize)
2742 {
2743 LLVMTypeRef type;
2744 char *intr;
2745
2746 if (bitsize == 16) {
2747 intr = "llvm.amdgcn.fract.f16";
2748 type = ctx->f16;
2749 } else if (bitsize == 32) {
2750 intr = "llvm.amdgcn.fract.f32";
2751 type = ctx->f32;
2752 } else {
2753 intr = "llvm.amdgcn.fract.f64";
2754 type = ctx->f64;
2755 }
2756
2757 LLVMValueRef params[] = {
2758 src0,
2759 };
2760 return ac_build_intrinsic(ctx, intr, type, params, 1,
2761 AC_FUNC_ATTR_READNONE);
2762 }
2763
2764 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2765 unsigned bitsize)
2766 {
2767 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2768 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2769 LLVMValueRef one = LLVMConstInt(type, 1, false);
2770
2771 LLVMValueRef cmp, val;
2772 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2773 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2774 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2775 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2776 return val;
2777 }
2778
2779 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2780 unsigned bitsize)
2781 {
2782 LLVMValueRef cmp, val, zero, one;
2783 LLVMTypeRef type;
2784
2785 if (bitsize == 16) {
2786 type = ctx->f16;
2787 zero = ctx->f16_0;
2788 one = ctx->f16_1;
2789 } else if (bitsize == 32) {
2790 type = ctx->f32;
2791 zero = ctx->f32_0;
2792 one = ctx->f32_1;
2793 } else {
2794 type = ctx->f64;
2795 zero = ctx->f64_0;
2796 one = ctx->f64_1;
2797 }
2798
2799 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2800 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2801 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2802 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2803 return val;
2804 }
2805
2806 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2807 {
2808 LLVMValueRef result;
2809 unsigned bitsize;
2810
2811 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2812
2813 switch (bitsize) {
2814 case 64:
2815 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2816 (LLVMValueRef []) { src0 }, 1,
2817 AC_FUNC_ATTR_READNONE);
2818
2819 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2820 break;
2821 case 32:
2822 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2823 (LLVMValueRef []) { src0 }, 1,
2824 AC_FUNC_ATTR_READNONE);
2825 break;
2826 case 16:
2827 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
2828 (LLVMValueRef []) { src0 }, 1,
2829 AC_FUNC_ATTR_READNONE);
2830
2831 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2832 break;
2833 case 8:
2834 result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
2835 (LLVMValueRef []) { src0 }, 1,
2836 AC_FUNC_ATTR_READNONE);
2837
2838 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2839 break;
2840 default:
2841 unreachable(!"invalid bitsize");
2842 break;
2843 }
2844
2845 return result;
2846 }
2847
2848 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
2849 LLVMValueRef src0)
2850 {
2851 LLVMValueRef result;
2852 unsigned bitsize;
2853
2854 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2855
2856 switch (bitsize) {
2857 case 64:
2858 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
2859 (LLVMValueRef []) { src0 }, 1,
2860 AC_FUNC_ATTR_READNONE);
2861
2862 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2863 break;
2864 case 32:
2865 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
2866 (LLVMValueRef []) { src0 }, 1,
2867 AC_FUNC_ATTR_READNONE);
2868 break;
2869 case 16:
2870 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
2871 (LLVMValueRef []) { src0 }, 1,
2872 AC_FUNC_ATTR_READNONE);
2873
2874 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2875 break;
2876 case 8:
2877 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
2878 (LLVMValueRef []) { src0 }, 1,
2879 AC_FUNC_ATTR_READNONE);
2880
2881 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2882 break;
2883 default:
2884 unreachable(!"invalid bitsize");
2885 break;
2886 }
2887
2888 return result;
2889 }
2890
2891 #define AC_EXP_TARGET 0
2892 #define AC_EXP_ENABLED_CHANNELS 1
2893 #define AC_EXP_OUT0 2
2894
2895 enum ac_ir_type {
2896 AC_IR_UNDEF,
2897 AC_IR_CONST,
2898 AC_IR_VALUE,
2899 };
2900
2901 struct ac_vs_exp_chan
2902 {
2903 LLVMValueRef value;
2904 float const_float;
2905 enum ac_ir_type type;
2906 };
2907
2908 struct ac_vs_exp_inst {
2909 unsigned offset;
2910 LLVMValueRef inst;
2911 struct ac_vs_exp_chan chan[4];
2912 };
2913
2914 struct ac_vs_exports {
2915 unsigned num;
2916 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2917 };
2918
2919 /* Return true if the PARAM export has been eliminated. */
2920 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2921 uint32_t num_outputs,
2922 struct ac_vs_exp_inst *exp)
2923 {
2924 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2925 bool is_zero[4] = {}, is_one[4] = {};
2926
2927 for (i = 0; i < 4; i++) {
2928 /* It's a constant expression. Undef outputs are eliminated too. */
2929 if (exp->chan[i].type == AC_IR_UNDEF) {
2930 is_zero[i] = true;
2931 is_one[i] = true;
2932 } else if (exp->chan[i].type == AC_IR_CONST) {
2933 if (exp->chan[i].const_float == 0)
2934 is_zero[i] = true;
2935 else if (exp->chan[i].const_float == 1)
2936 is_one[i] = true;
2937 else
2938 return false; /* other constant */
2939 } else
2940 return false;
2941 }
2942
2943 /* Only certain combinations of 0 and 1 can be eliminated. */
2944 if (is_zero[0] && is_zero[1] && is_zero[2])
2945 default_val = is_zero[3] ? 0 : 1;
2946 else if (is_one[0] && is_one[1] && is_one[2])
2947 default_val = is_zero[3] ? 2 : 3;
2948 else
2949 return false;
2950
2951 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2952 LLVMInstructionEraseFromParent(exp->inst);
2953
2954 /* Change OFFSET to DEFAULT_VAL. */
2955 for (i = 0; i < num_outputs; i++) {
2956 if (vs_output_param_offset[i] == exp->offset) {
2957 vs_output_param_offset[i] =
2958 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
2959 break;
2960 }
2961 }
2962 return true;
2963 }
2964
2965 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
2966 uint8_t *vs_output_param_offset,
2967 uint32_t num_outputs,
2968 struct ac_vs_exports *processed,
2969 struct ac_vs_exp_inst *exp)
2970 {
2971 unsigned p, copy_back_channels = 0;
2972
2973 /* See if the output is already in the list of processed outputs.
2974 * The LLVMValueRef comparison relies on SSA.
2975 */
2976 for (p = 0; p < processed->num; p++) {
2977 bool different = false;
2978
2979 for (unsigned j = 0; j < 4; j++) {
2980 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
2981 struct ac_vs_exp_chan *c2 = &exp->chan[j];
2982
2983 /* Treat undef as a match. */
2984 if (c2->type == AC_IR_UNDEF)
2985 continue;
2986
2987 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2988 * and consider the instruction duplicated.
2989 */
2990 if (c1->type == AC_IR_UNDEF) {
2991 copy_back_channels |= 1 << j;
2992 continue;
2993 }
2994
2995 /* Test whether the channels are not equal. */
2996 if (c1->type != c2->type ||
2997 (c1->type == AC_IR_CONST &&
2998 c1->const_float != c2->const_float) ||
2999 (c1->type == AC_IR_VALUE &&
3000 c1->value != c2->value)) {
3001 different = true;
3002 break;
3003 }
3004 }
3005 if (!different)
3006 break;
3007
3008 copy_back_channels = 0;
3009 }
3010 if (p == processed->num)
3011 return false;
3012
3013 /* If a match was found, but the matching export has undef where the new
3014 * one has a normal value, copy the normal value to the undef channel.
3015 */
3016 struct ac_vs_exp_inst *match = &processed->exp[p];
3017
3018 /* Get current enabled channels mask. */
3019 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
3020 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
3021
3022 while (copy_back_channels) {
3023 unsigned chan = u_bit_scan(&copy_back_channels);
3024
3025 assert(match->chan[chan].type == AC_IR_UNDEF);
3026 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
3027 exp->chan[chan].value);
3028 match->chan[chan] = exp->chan[chan];
3029
3030 /* Update number of enabled channels because the original mask
3031 * is not always 0xf.
3032 */
3033 enabled_channels |= (1 << chan);
3034 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
3035 LLVMConstInt(ctx->i32, enabled_channels, 0));
3036 }
3037
3038 /* The PARAM export is duplicated. Kill it. */
3039 LLVMInstructionEraseFromParent(exp->inst);
3040
3041 /* Change OFFSET to the matching export. */
3042 for (unsigned i = 0; i < num_outputs; i++) {
3043 if (vs_output_param_offset[i] == exp->offset) {
3044 vs_output_param_offset[i] = match->offset;
3045 break;
3046 }
3047 }
3048 return true;
3049 }
3050
3051 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
3052 LLVMValueRef main_fn,
3053 uint8_t *vs_output_param_offset,
3054 uint32_t num_outputs,
3055 uint8_t *num_param_exports)
3056 {
3057 LLVMBasicBlockRef bb;
3058 bool removed_any = false;
3059 struct ac_vs_exports exports;
3060
3061 exports.num = 0;
3062
3063 /* Process all LLVM instructions. */
3064 bb = LLVMGetFirstBasicBlock(main_fn);
3065 while (bb) {
3066 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
3067
3068 while (inst) {
3069 LLVMValueRef cur = inst;
3070 inst = LLVMGetNextInstruction(inst);
3071 struct ac_vs_exp_inst exp;
3072
3073 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
3074 continue;
3075
3076 LLVMValueRef callee = ac_llvm_get_called_value(cur);
3077
3078 if (!ac_llvm_is_function(callee))
3079 continue;
3080
3081 const char *name = LLVMGetValueName(callee);
3082 unsigned num_args = LLVMCountParams(callee);
3083
3084 /* Check if this is an export instruction. */
3085 if ((num_args != 9 && num_args != 8) ||
3086 (strcmp(name, "llvm.SI.export") &&
3087 strcmp(name, "llvm.amdgcn.exp.f32")))
3088 continue;
3089
3090 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
3091 unsigned target = LLVMConstIntGetZExtValue(arg);
3092
3093 if (target < V_008DFC_SQ_EXP_PARAM)
3094 continue;
3095
3096 target -= V_008DFC_SQ_EXP_PARAM;
3097
3098 /* Parse the instruction. */
3099 memset(&exp, 0, sizeof(exp));
3100 exp.offset = target;
3101 exp.inst = cur;
3102
3103 for (unsigned i = 0; i < 4; i++) {
3104 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
3105
3106 exp.chan[i].value = v;
3107
3108 if (LLVMIsUndef(v)) {
3109 exp.chan[i].type = AC_IR_UNDEF;
3110 } else if (LLVMIsAConstantFP(v)) {
3111 LLVMBool loses_info;
3112 exp.chan[i].type = AC_IR_CONST;
3113 exp.chan[i].const_float =
3114 LLVMConstRealGetDouble(v, &loses_info);
3115 } else {
3116 exp.chan[i].type = AC_IR_VALUE;
3117 }
3118 }
3119
3120 /* Eliminate constant and duplicated PARAM exports. */
3121 if (ac_eliminate_const_output(vs_output_param_offset,
3122 num_outputs, &exp) ||
3123 ac_eliminate_duplicated_output(ctx,
3124 vs_output_param_offset,
3125 num_outputs, &exports,
3126 &exp)) {
3127 removed_any = true;
3128 } else {
3129 exports.exp[exports.num++] = exp;
3130 }
3131 }
3132 bb = LLVMGetNextBasicBlock(bb);
3133 }
3134
3135 /* Remove holes in export memory due to removed PARAM exports.
3136 * This is done by renumbering all PARAM exports.
3137 */
3138 if (removed_any) {
3139 uint8_t old_offset[VARYING_SLOT_MAX];
3140 unsigned out, i;
3141
3142 /* Make a copy of the offsets. We need the old version while
3143 * we are modifying some of them. */
3144 memcpy(old_offset, vs_output_param_offset,
3145 sizeof(old_offset));
3146
3147 for (i = 0; i < exports.num; i++) {
3148 unsigned offset = exports.exp[i].offset;
3149
3150 /* Update vs_output_param_offset. Multiple outputs can
3151 * have the same offset.
3152 */
3153 for (out = 0; out < num_outputs; out++) {
3154 if (old_offset[out] == offset)
3155 vs_output_param_offset[out] = i;
3156 }
3157
3158 /* Change the PARAM offset in the instruction. */
3159 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
3160 LLVMConstInt(ctx->i32,
3161 V_008DFC_SQ_EXP_PARAM + i, 0));
3162 }
3163 *num_param_exports = exports.num;
3164 }
3165 }
3166
3167 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
3168 {
3169 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
3170 ac_build_intrinsic(ctx,
3171 "llvm.amdgcn.init.exec", ctx->voidt,
3172 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
3173 }
3174
3175 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
3176 {
3177 unsigned lds_size = ctx->chip_class >= GFX7 ? 65536 : 32768;
3178 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
3179 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
3180 "lds");
3181 }
3182
3183 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
3184 LLVMValueRef dw_addr)
3185 {
3186 return LLVMBuildLoad(ctx->builder, ac_build_gep0(ctx, ctx->lds, dw_addr), "");
3187 }
3188
3189 void ac_lds_store(struct ac_llvm_context *ctx,
3190 LLVMValueRef dw_addr,
3191 LLVMValueRef value)
3192 {
3193 value = ac_to_integer(ctx, value);
3194 ac_build_indexed_store(ctx, ctx->lds,
3195 dw_addr, value);
3196 }
3197
3198 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
3199 LLVMTypeRef dst_type,
3200 LLVMValueRef src0)
3201 {
3202 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3203 const char *intrin_name;
3204 LLVMTypeRef type;
3205 LLVMValueRef zero;
3206
3207 switch (src0_bitsize) {
3208 case 64:
3209 intrin_name = "llvm.cttz.i64";
3210 type = ctx->i64;
3211 zero = ctx->i64_0;
3212 break;
3213 case 32:
3214 intrin_name = "llvm.cttz.i32";
3215 type = ctx->i32;
3216 zero = ctx->i32_0;
3217 break;
3218 case 16:
3219 intrin_name = "llvm.cttz.i16";
3220 type = ctx->i16;
3221 zero = ctx->i16_0;
3222 break;
3223 case 8:
3224 intrin_name = "llvm.cttz.i8";
3225 type = ctx->i8;
3226 zero = ctx->i8_0;
3227 break;
3228 default:
3229 unreachable(!"invalid bitsize");
3230 }
3231
3232 LLVMValueRef params[2] = {
3233 src0,
3234
3235 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3236 * add special code to check for x=0. The reason is that
3237 * the LLVM behavior for x=0 is different from what we
3238 * need here. However, LLVM also assumes that ffs(x) is
3239 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3240 * a conditional assignment to handle 0 is still required.
3241 *
3242 * The hardware already implements the correct behavior.
3243 */
3244 ctx->i1true,
3245 };
3246
3247 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
3248 params, 2,
3249 AC_FUNC_ATTR_READNONE);
3250
3251 if (src0_bitsize == 64) {
3252 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
3253 } else if (src0_bitsize < 32) {
3254 lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
3255 }
3256
3257 /* TODO: We need an intrinsic to skip this conditional. */
3258 /* Check for zero: */
3259 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
3260 LLVMIntEQ, src0,
3261 zero, ""),
3262 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
3263 }
3264
3265 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
3266 {
3267 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST);
3268 }
3269
3270 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
3271 {
3272 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST_32BIT);
3273 }
3274
3275 static struct ac_llvm_flow *
3276 get_current_flow(struct ac_llvm_context *ctx)
3277 {
3278 if (ctx->flow->depth > 0)
3279 return &ctx->flow->stack[ctx->flow->depth - 1];
3280 return NULL;
3281 }
3282
3283 static struct ac_llvm_flow *
3284 get_innermost_loop(struct ac_llvm_context *ctx)
3285 {
3286 for (unsigned i = ctx->flow->depth; i > 0; --i) {
3287 if (ctx->flow->stack[i - 1].loop_entry_block)
3288 return &ctx->flow->stack[i - 1];
3289 }
3290 return NULL;
3291 }
3292
3293 static struct ac_llvm_flow *
3294 push_flow(struct ac_llvm_context *ctx)
3295 {
3296 struct ac_llvm_flow *flow;
3297
3298 if (ctx->flow->depth >= ctx->flow->depth_max) {
3299 unsigned new_max = MAX2(ctx->flow->depth << 1,
3300 AC_LLVM_INITIAL_CF_DEPTH);
3301
3302 ctx->flow->stack = realloc(ctx->flow->stack, new_max * sizeof(*ctx->flow->stack));
3303 ctx->flow->depth_max = new_max;
3304 }
3305
3306 flow = &ctx->flow->stack[ctx->flow->depth];
3307 ctx->flow->depth++;
3308
3309 flow->next_block = NULL;
3310 flow->loop_entry_block = NULL;
3311 return flow;
3312 }
3313
3314 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
3315 int label_id)
3316 {
3317 char buf[32];
3318 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
3319 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
3320 }
3321
3322 /* Append a basic block at the level of the parent flow.
3323 */
3324 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
3325 const char *name)
3326 {
3327 assert(ctx->flow->depth >= 1);
3328
3329 if (ctx->flow->depth >= 2) {
3330 struct ac_llvm_flow *flow = &ctx->flow->stack[ctx->flow->depth - 2];
3331
3332 return LLVMInsertBasicBlockInContext(ctx->context,
3333 flow->next_block, name);
3334 }
3335
3336 LLVMValueRef main_fn =
3337 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
3338 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
3339 }
3340
3341 /* Emit a branch to the given default target for the current block if
3342 * applicable -- that is, if the current block does not already contain a
3343 * branch from a break or continue.
3344 */
3345 static void emit_default_branch(LLVMBuilderRef builder,
3346 LLVMBasicBlockRef target)
3347 {
3348 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
3349 LLVMBuildBr(builder, target);
3350 }
3351
3352 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
3353 {
3354 struct ac_llvm_flow *flow = push_flow(ctx);
3355 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
3356 flow->next_block = append_basic_block(ctx, "ENDLOOP");
3357 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
3358 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3359 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
3360 }
3361
3362 void ac_build_break(struct ac_llvm_context *ctx)
3363 {
3364 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3365 LLVMBuildBr(ctx->builder, flow->next_block);
3366 }
3367
3368 void ac_build_continue(struct ac_llvm_context *ctx)
3369 {
3370 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3371 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3372 }
3373
3374 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
3375 {
3376 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3377 LLVMBasicBlockRef endif_block;
3378
3379 assert(!current_branch->loop_entry_block);
3380
3381 endif_block = append_basic_block(ctx, "ENDIF");
3382 emit_default_branch(ctx->builder, endif_block);
3383
3384 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3385 set_basicblock_name(current_branch->next_block, "else", label_id);
3386
3387 current_branch->next_block = endif_block;
3388 }
3389
3390 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
3391 {
3392 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3393
3394 assert(!current_branch->loop_entry_block);
3395
3396 emit_default_branch(ctx->builder, current_branch->next_block);
3397 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3398 set_basicblock_name(current_branch->next_block, "endif", label_id);
3399
3400 ctx->flow->depth--;
3401 }
3402
3403 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
3404 {
3405 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
3406
3407 assert(current_loop->loop_entry_block);
3408
3409 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
3410
3411 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
3412 set_basicblock_name(current_loop->next_block, "endloop", label_id);
3413 ctx->flow->depth--;
3414 }
3415
3416 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
3417 {
3418 struct ac_llvm_flow *flow = push_flow(ctx);
3419 LLVMBasicBlockRef if_block;
3420
3421 if_block = append_basic_block(ctx, "IF");
3422 flow->next_block = append_basic_block(ctx, "ELSE");
3423 set_basicblock_name(if_block, "if", label_id);
3424 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
3425 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
3426 }
3427
3428 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
3429 int label_id)
3430 {
3431 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
3432 value, ctx->f32_0, "");
3433 ac_build_ifcc(ctx, cond, label_id);
3434 }
3435
3436 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
3437 int label_id)
3438 {
3439 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
3440 ac_to_integer(ctx, value),
3441 ctx->i32_0, "");
3442 ac_build_ifcc(ctx, cond, label_id);
3443 }
3444
3445 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
3446 const char *name)
3447 {
3448 LLVMBuilderRef builder = ac->builder;
3449 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
3450 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
3451 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
3452 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
3453 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
3454 LLVMValueRef res;
3455
3456 if (first_instr) {
3457 LLVMPositionBuilderBefore(first_builder, first_instr);
3458 } else {
3459 LLVMPositionBuilderAtEnd(first_builder, first_block);
3460 }
3461
3462 res = LLVMBuildAlloca(first_builder, type, name);
3463 LLVMDisposeBuilder(first_builder);
3464 return res;
3465 }
3466
3467 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
3468 LLVMTypeRef type, const char *name)
3469 {
3470 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
3471 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
3472 return ptr;
3473 }
3474
3475 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
3476 LLVMTypeRef type)
3477 {
3478 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3479 return LLVMBuildBitCast(ctx->builder, ptr,
3480 LLVMPointerType(type, addr_space), "");
3481 }
3482
3483 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
3484 unsigned count)
3485 {
3486 unsigned num_components = ac_get_llvm_num_components(value);
3487 if (count == num_components)
3488 return value;
3489
3490 LLVMValueRef masks[MAX2(count, 2)];
3491 masks[0] = ctx->i32_0;
3492 masks[1] = ctx->i32_1;
3493 for (unsigned i = 2; i < count; i++)
3494 masks[i] = LLVMConstInt(ctx->i32, i, false);
3495
3496 if (count == 1)
3497 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
3498 "");
3499
3500 LLVMValueRef swizzle = LLVMConstVector(masks, count);
3501 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
3502 }
3503
3504 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
3505 unsigned rshift, unsigned bitwidth)
3506 {
3507 LLVMValueRef value = param;
3508 if (rshift)
3509 value = LLVMBuildLShr(ctx->builder, value,
3510 LLVMConstInt(ctx->i32, rshift, false), "");
3511
3512 if (rshift + bitwidth < 32) {
3513 unsigned mask = (1 << bitwidth) - 1;
3514 value = LLVMBuildAnd(ctx->builder, value,
3515 LLVMConstInt(ctx->i32, mask, false), "");
3516 }
3517 return value;
3518 }
3519
3520 /* Adjust the sample index according to FMASK.
3521 *
3522 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3523 * which is the identity mapping. Each nibble says which physical sample
3524 * should be fetched to get that sample.
3525 *
3526 * For example, 0x11111100 means there are only 2 samples stored and
3527 * the second sample covers 3/4 of the pixel. When reading samples 0
3528 * and 1, return physical sample 0 (determined by the first two 0s
3529 * in FMASK), otherwise return physical sample 1.
3530 *
3531 * The sample index should be adjusted as follows:
3532 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3533 */
3534 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
3535 LLVMValueRef *addr, bool is_array_tex)
3536 {
3537 struct ac_image_args fmask_load = {};
3538 fmask_load.opcode = ac_image_load;
3539 fmask_load.resource = fmask;
3540 fmask_load.dmask = 0xf;
3541 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
3542 fmask_load.attributes = AC_FUNC_ATTR_READNONE;
3543
3544 fmask_load.coords[0] = addr[0];
3545 fmask_load.coords[1] = addr[1];
3546 if (is_array_tex)
3547 fmask_load.coords[2] = addr[2];
3548
3549 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
3550 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
3551 ac->i32_0, "");
3552
3553 /* Apply the formula. */
3554 unsigned sample_chan = is_array_tex ? 3 : 2;
3555 LLVMValueRef final_sample;
3556 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
3557 LLVMConstInt(ac->i32, 4, 0), "");
3558 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
3559 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3560 * with EQAA, so those will map to 0. */
3561 final_sample = LLVMBuildAnd(ac->builder, final_sample,
3562 LLVMConstInt(ac->i32, 0x7, 0), "");
3563
3564 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3565 * resource descriptor is 0 (invalid).
3566 */
3567 LLVMValueRef tmp;
3568 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
3569 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
3570 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
3571
3572 /* Replace the MSAA sample index. */
3573 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
3574 addr[sample_chan], "");
3575 }
3576
3577 static LLVMValueRef
3578 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3579 {
3580 LLVMTypeRef type = LLVMTypeOf(src);
3581 LLVMValueRef result;
3582
3583 ac_build_optimization_barrier(ctx, &src);
3584
3585 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3586 if (lane)
3587 lane = LLVMBuildZExt(ctx->builder, lane, ctx->i32, "");
3588
3589 result = ac_build_intrinsic(ctx,
3590 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3591 ctx->i32, (LLVMValueRef []) { src, lane },
3592 lane == NULL ? 1 : 2,
3593 AC_FUNC_ATTR_READNONE |
3594 AC_FUNC_ATTR_CONVERGENT);
3595
3596 return LLVMBuildTrunc(ctx->builder, result, type, "");
3597 }
3598
3599 /**
3600 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3601 * @param ctx
3602 * @param src
3603 * @param lane - id of the lane or NULL for the first active lane
3604 * @return value of the lane
3605 */
3606 LLVMValueRef
3607 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3608 {
3609 LLVMTypeRef src_type = LLVMTypeOf(src);
3610 src = ac_to_integer(ctx, src);
3611 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3612 LLVMValueRef ret;
3613
3614 if (bits > 32) {
3615 assert(bits % 32 == 0);
3616 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3617 LLVMValueRef src_vector =
3618 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3619 ret = LLVMGetUndef(vec_type);
3620 for (unsigned i = 0; i < bits / 32; i++) {
3621 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3622 LLVMConstInt(ctx->i32, i, 0), "");
3623 LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
3624 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
3625 LLVMConstInt(ctx->i32, i, 0), "");
3626 }
3627 } else {
3628 ret = _ac_build_readlane(ctx, src, lane);
3629 }
3630
3631 if (LLVMGetTypeKind(src_type) == LLVMPointerTypeKind)
3632 return LLVMBuildIntToPtr(ctx->builder, ret, src_type, "");
3633 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3634 }
3635
3636 LLVMValueRef
3637 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
3638 {
3639 return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
3640 (LLVMValueRef []) {value, lane, src}, 3,
3641 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3642 }
3643
3644 LLVMValueRef
3645 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
3646 {
3647 if (ctx->wave_size == 32) {
3648 return ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3649 (LLVMValueRef []) { mask, ctx->i32_0 },
3650 2, AC_FUNC_ATTR_READNONE);
3651 }
3652 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
3653 LLVMVectorType(ctx->i32, 2),
3654 "");
3655 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
3656 ctx->i32_0, "");
3657 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
3658 ctx->i32_1, "");
3659 LLVMValueRef val =
3660 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3661 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
3662 2, AC_FUNC_ATTR_READNONE);
3663 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
3664 (LLVMValueRef []) { mask_hi, val },
3665 2, AC_FUNC_ATTR_READNONE);
3666 return val;
3667 }
3668
3669 enum dpp_ctrl {
3670 _dpp_quad_perm = 0x000,
3671 _dpp_row_sl = 0x100,
3672 _dpp_row_sr = 0x110,
3673 _dpp_row_rr = 0x120,
3674 dpp_wf_sl1 = 0x130,
3675 dpp_wf_rl1 = 0x134,
3676 dpp_wf_sr1 = 0x138,
3677 dpp_wf_rr1 = 0x13C,
3678 dpp_row_mirror = 0x140,
3679 dpp_row_half_mirror = 0x141,
3680 dpp_row_bcast15 = 0x142,
3681 dpp_row_bcast31 = 0x143
3682 };
3683
3684 static inline enum dpp_ctrl
3685 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3686 {
3687 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
3688 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
3689 }
3690
3691 static inline enum dpp_ctrl
3692 dpp_row_sl(unsigned amount)
3693 {
3694 assert(amount > 0 && amount < 16);
3695 return _dpp_row_sl | amount;
3696 }
3697
3698 static inline enum dpp_ctrl
3699 dpp_row_sr(unsigned amount)
3700 {
3701 assert(amount > 0 && amount < 16);
3702 return _dpp_row_sr | amount;
3703 }
3704
3705 static LLVMValueRef
3706 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3707 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3708 bool bound_ctrl)
3709 {
3710 LLVMTypeRef type = LLVMTypeOf(src);
3711 LLVMValueRef res;
3712
3713 old = LLVMBuildZExt(ctx->builder, old, ctx->i32, "");
3714 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3715
3716 res = ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32", ctx->i32,
3717 (LLVMValueRef[]) {
3718 old, src,
3719 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3720 LLVMConstInt(ctx->i32, row_mask, 0),
3721 LLVMConstInt(ctx->i32, bank_mask, 0),
3722 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3723 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3724
3725 return LLVMBuildTrunc(ctx->builder, res, type, "");
3726 }
3727
3728 static LLVMValueRef
3729 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3730 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3731 bool bound_ctrl)
3732 {
3733 LLVMTypeRef src_type = LLVMTypeOf(src);
3734 src = ac_to_integer(ctx, src);
3735 old = ac_to_integer(ctx, old);
3736 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3737 LLVMValueRef ret;
3738 if (bits > 32) {
3739 assert(bits % 32 == 0);
3740 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3741 LLVMValueRef src_vector =
3742 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3743 LLVMValueRef old_vector =
3744 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3745 ret = LLVMGetUndef(vec_type);
3746 for (unsigned i = 0; i < bits / 32; i++) {
3747 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3748 LLVMConstInt(ctx->i32, i,
3749 0), "");
3750 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3751 LLVMConstInt(ctx->i32, i,
3752 0), "");
3753 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3754 dpp_ctrl,
3755 row_mask,
3756 bank_mask,
3757 bound_ctrl);
3758 ret = LLVMBuildInsertElement(ctx->builder, ret,
3759 ret_comp,
3760 LLVMConstInt(ctx->i32, i,
3761 0), "");
3762 }
3763 } else {
3764 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3765 bank_mask, bound_ctrl);
3766 }
3767 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3768 }
3769
3770 static LLVMValueRef
3771 _ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3772 bool exchange_rows, bool bound_ctrl)
3773 {
3774 LLVMValueRef args[6] = {
3775 src,
3776 src,
3777 LLVMConstInt(ctx->i32, sel, false),
3778 LLVMConstInt(ctx->i32, sel >> 32, false),
3779 ctx->i1true, /* fi */
3780 bound_ctrl ? ctx->i1true : ctx->i1false,
3781 };
3782 return ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
3783 : "llvm.amdgcn.permlane16",
3784 ctx->i32, args, 6,
3785 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3786 }
3787
3788 static LLVMValueRef
3789 ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3790 bool exchange_rows, bool bound_ctrl)
3791 {
3792 LLVMTypeRef src_type = LLVMTypeOf(src);
3793 src = ac_to_integer(ctx, src);
3794 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3795 LLVMValueRef ret;
3796 if (bits == 32) {
3797 ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
3798 bound_ctrl);
3799 } else {
3800 assert(bits % 32 == 0);
3801 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3802 LLVMValueRef src_vector =
3803 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3804 ret = LLVMGetUndef(vec_type);
3805 for (unsigned i = 0; i < bits / 32; i++) {
3806 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3807 LLVMConstInt(ctx->i32, i,
3808 0), "");
3809 LLVMValueRef ret_comp =
3810 _ac_build_permlane16(ctx, src, sel,
3811 exchange_rows,
3812 bound_ctrl);
3813 ret = LLVMBuildInsertElement(ctx->builder, ret,
3814 ret_comp,
3815 LLVMConstInt(ctx->i32, i,
3816 0), "");
3817 }
3818 }
3819 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3820 }
3821
3822 static inline unsigned
3823 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
3824 {
3825 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
3826 return and_mask | (or_mask << 5) | (xor_mask << 10);
3827 }
3828
3829 static LLVMValueRef
3830 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3831 {
3832 LLVMTypeRef src_type = LLVMTypeOf(src);
3833 LLVMValueRef ret;
3834
3835 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3836
3837 ret = ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle", ctx->i32,
3838 (LLVMValueRef []) {
3839 src, LLVMConstInt(ctx->i32, mask, 0) },
3840 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3841
3842 return LLVMBuildTrunc(ctx->builder, ret, src_type, "");
3843 }
3844
3845 LLVMValueRef
3846 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3847 {
3848 LLVMTypeRef src_type = LLVMTypeOf(src);
3849 src = ac_to_integer(ctx, src);
3850 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3851 LLVMValueRef ret;
3852 if (bits > 32) {
3853 assert(bits % 32 == 0);
3854 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3855 LLVMValueRef src_vector =
3856 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3857 ret = LLVMGetUndef(vec_type);
3858 for (unsigned i = 0; i < bits / 32; i++) {
3859 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3860 LLVMConstInt(ctx->i32, i,
3861 0), "");
3862 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
3863 mask);
3864 ret = LLVMBuildInsertElement(ctx->builder, ret,
3865 ret_comp,
3866 LLVMConstInt(ctx->i32, i,
3867 0), "");
3868 }
3869 } else {
3870 ret = _ac_build_ds_swizzle(ctx, src, mask);
3871 }
3872 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3873 }
3874
3875 static LLVMValueRef
3876 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
3877 {
3878 LLVMTypeRef src_type = LLVMTypeOf(src);
3879 unsigned bitsize = ac_get_elem_bits(ctx, src_type);
3880 char name[32], type[8];
3881 LLVMValueRef ret;
3882
3883 src = ac_to_integer(ctx, src);
3884
3885 if (bitsize < 32)
3886 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3887
3888 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3889 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
3890 ret = ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
3891 (LLVMValueRef []) { src }, 1,
3892 AC_FUNC_ATTR_READNONE);
3893
3894 if (bitsize < 32)
3895 ret = LLVMBuildTrunc(ctx->builder, ret,
3896 ac_to_integer_type(ctx, src_type), "");
3897
3898 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3899 }
3900
3901 static LLVMValueRef
3902 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
3903 LLVMValueRef inactive)
3904 {
3905 char name[33], type[8];
3906 LLVMTypeRef src_type = LLVMTypeOf(src);
3907 unsigned bitsize = ac_get_elem_bits(ctx, src_type);
3908 src = ac_to_integer(ctx, src);
3909 inactive = ac_to_integer(ctx, inactive);
3910
3911 if (bitsize < 32) {
3912 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3913 inactive = LLVMBuildZExt(ctx->builder, inactive, ctx->i32, "");
3914 }
3915
3916 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3917 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
3918 LLVMValueRef ret =
3919 ac_build_intrinsic(ctx, name,
3920 LLVMTypeOf(src), (LLVMValueRef []) {
3921 src, inactive }, 2,
3922 AC_FUNC_ATTR_READNONE |
3923 AC_FUNC_ATTR_CONVERGENT);
3924 if (bitsize < 32)
3925 ret = LLVMBuildTrunc(ctx->builder, ret, src_type, "");
3926
3927 return ret;
3928 }
3929
3930 static LLVMValueRef
3931 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
3932 {
3933 if (type_size == 1) {
3934 switch (op) {
3935 case nir_op_iadd: return ctx->i8_0;
3936 case nir_op_imul: return ctx->i8_1;
3937 case nir_op_imin: return LLVMConstInt(ctx->i8, INT8_MAX, 0);
3938 case nir_op_umin: return LLVMConstInt(ctx->i8, UINT8_MAX, 0);
3939 case nir_op_imax: return LLVMConstInt(ctx->i8, INT8_MIN, 0);
3940 case nir_op_umax: return ctx->i8_0;
3941 case nir_op_iand: return LLVMConstInt(ctx->i8, -1, 0);
3942 case nir_op_ior: return ctx->i8_0;
3943 case nir_op_ixor: return ctx->i8_0;
3944 default:
3945 unreachable("bad reduction intrinsic");
3946 }
3947 } else if (type_size == 2) {
3948 switch (op) {
3949 case nir_op_iadd: return ctx->i16_0;
3950 case nir_op_fadd: return ctx->f16_0;
3951 case nir_op_imul: return ctx->i16_1;
3952 case nir_op_fmul: return ctx->f16_1;
3953 case nir_op_imin: return LLVMConstInt(ctx->i16, INT16_MAX, 0);
3954 case nir_op_umin: return LLVMConstInt(ctx->i16, UINT16_MAX, 0);
3955 case nir_op_fmin: return LLVMConstReal(ctx->f16, INFINITY);
3956 case nir_op_imax: return LLVMConstInt(ctx->i16, INT16_MIN, 0);
3957 case nir_op_umax: return ctx->i16_0;
3958 case nir_op_fmax: return LLVMConstReal(ctx->f16, -INFINITY);
3959 case nir_op_iand: return LLVMConstInt(ctx->i16, -1, 0);
3960 case nir_op_ior: return ctx->i16_0;
3961 case nir_op_ixor: return ctx->i16_0;
3962 default:
3963 unreachable("bad reduction intrinsic");
3964 }
3965 } else if (type_size == 4) {
3966 switch (op) {
3967 case nir_op_iadd: return ctx->i32_0;
3968 case nir_op_fadd: return ctx->f32_0;
3969 case nir_op_imul: return ctx->i32_1;
3970 case nir_op_fmul: return ctx->f32_1;
3971 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
3972 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
3973 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
3974 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
3975 case nir_op_umax: return ctx->i32_0;
3976 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
3977 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
3978 case nir_op_ior: return ctx->i32_0;
3979 case nir_op_ixor: return ctx->i32_0;
3980 default:
3981 unreachable("bad reduction intrinsic");
3982 }
3983 } else { /* type_size == 64bit */
3984 switch (op) {
3985 case nir_op_iadd: return ctx->i64_0;
3986 case nir_op_fadd: return ctx->f64_0;
3987 case nir_op_imul: return ctx->i64_1;
3988 case nir_op_fmul: return ctx->f64_1;
3989 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
3990 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
3991 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
3992 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
3993 case nir_op_umax: return ctx->i64_0;
3994 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
3995 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
3996 case nir_op_ior: return ctx->i64_0;
3997 case nir_op_ixor: return ctx->i64_0;
3998 default:
3999 unreachable("bad reduction intrinsic");
4000 }
4001 }
4002 }
4003
4004 static LLVMValueRef
4005 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
4006 {
4007 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
4008 bool _32bit = ac_get_type_size(LLVMTypeOf(lhs)) == 4;
4009 switch (op) {
4010 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
4011 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
4012 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
4013 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
4014 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
4015 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
4016 lhs, rhs, "");
4017 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
4018 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
4019 lhs, rhs, "");
4020 case nir_op_fmin: return ac_build_intrinsic(ctx,
4021 _64bit ? "llvm.minnum.f64" : _32bit ? "llvm.minnum.f32" : "llvm.minnum.f16",
4022 _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
4023 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4024 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
4025 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
4026 lhs, rhs, "");
4027 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
4028 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
4029 lhs, rhs, "");
4030 case nir_op_fmax: return ac_build_intrinsic(ctx,
4031 _64bit ? "llvm.maxnum.f64" : _32bit ? "llvm.maxnum.f32" : "llvm.maxnum.f16",
4032 _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
4033 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4034 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
4035 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
4036 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
4037 default:
4038 unreachable("bad reduction intrinsic");
4039 }
4040 }
4041
4042 /**
4043 * \param maxprefix specifies that the result only needs to be correct for a
4044 * prefix of this many threads
4045 *
4046 * TODO: add inclusive and excluse scan functions for GFX6.
4047 */
4048 static LLVMValueRef
4049 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
4050 unsigned maxprefix, bool inclusive)
4051 {
4052 LLVMValueRef result, tmp;
4053
4054 if (ctx->chip_class >= GFX10) {
4055 result = inclusive ? src : identity;
4056 } else {
4057 if (!inclusive)
4058 src = ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
4059 result = src;
4060 }
4061 if (maxprefix <= 1)
4062 return result;
4063 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
4064 result = ac_build_alu_op(ctx, result, tmp, op);
4065 if (maxprefix <= 2)
4066 return result;
4067 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
4068 result = ac_build_alu_op(ctx, result, tmp, op);
4069 if (maxprefix <= 3)
4070 return result;
4071 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
4072 result = ac_build_alu_op(ctx, result, tmp, op);
4073 if (maxprefix <= 4)
4074 return result;
4075 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
4076 result = ac_build_alu_op(ctx, result, tmp, op);
4077 if (maxprefix <= 8)
4078 return result;
4079 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
4080 result = ac_build_alu_op(ctx, result, tmp, op);
4081 if (maxprefix <= 16)
4082 return result;
4083
4084 if (ctx->chip_class >= GFX10) {
4085 /* dpp_row_bcast{15,31} are not supported on gfx10. */
4086 LLVMBuilderRef builder = ctx->builder;
4087 LLVMValueRef tid = ac_get_thread_id(ctx);
4088 LLVMValueRef cc;
4089 /* TODO-GFX10: Can we get better code-gen by putting this into
4090 * a branch so that LLVM generates EXEC mask manipulations? */
4091 if (inclusive)
4092 tmp = result;
4093 else
4094 tmp = ac_build_alu_op(ctx, result, src, op);
4095 tmp = ac_build_permlane16(ctx, tmp, ~(uint64_t)0, true, false);
4096 tmp = ac_build_alu_op(ctx, result, tmp, op);
4097 cc = LLVMBuildAnd(builder, tid, LLVMConstInt(ctx->i32, 16, false), "");
4098 cc = LLVMBuildICmp(builder, LLVMIntNE, cc, ctx->i32_0, "");
4099 result = LLVMBuildSelect(builder, cc, tmp, result, "");
4100 if (maxprefix <= 32)
4101 return result;
4102
4103 if (inclusive)
4104 tmp = result;
4105 else
4106 tmp = ac_build_alu_op(ctx, result, src, op);
4107 tmp = ac_build_readlane(ctx, tmp, LLVMConstInt(ctx->i32, 31, false));
4108 tmp = ac_build_alu_op(ctx, result, tmp, op);
4109 cc = LLVMBuildICmp(builder, LLVMIntUGE, tid,
4110 LLVMConstInt(ctx->i32, 32, false), "");
4111 result = LLVMBuildSelect(builder, cc, tmp, result, "");
4112 return result;
4113 }
4114
4115 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4116 result = ac_build_alu_op(ctx, result, tmp, op);
4117 if (maxprefix <= 32)
4118 return result;
4119 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4120 result = ac_build_alu_op(ctx, result, tmp, op);
4121 return result;
4122 }
4123
4124 LLVMValueRef
4125 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4126 {
4127 LLVMValueRef result;
4128
4129 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4130 LLVMBuilderRef builder = ctx->builder;
4131 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4132 result = ac_build_ballot(ctx, src);
4133 result = ac_build_mbcnt(ctx, result);
4134 result = LLVMBuildAdd(builder, result, src, "");
4135 return result;
4136 }
4137
4138 ac_build_optimization_barrier(ctx, &src);
4139
4140 LLVMValueRef identity =
4141 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4142 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4143 LLVMTypeOf(identity), "");
4144 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
4145
4146 return ac_build_wwm(ctx, result);
4147 }
4148
4149 LLVMValueRef
4150 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4151 {
4152 LLVMValueRef result;
4153
4154 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4155 LLVMBuilderRef builder = ctx->builder;
4156 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4157 result = ac_build_ballot(ctx, src);
4158 result = ac_build_mbcnt(ctx, result);
4159 return result;
4160 }
4161
4162 ac_build_optimization_barrier(ctx, &src);
4163
4164 LLVMValueRef identity =
4165 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4166 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4167 LLVMTypeOf(identity), "");
4168 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
4169
4170 return ac_build_wwm(ctx, result);
4171 }
4172
4173 LLVMValueRef
4174 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
4175 {
4176 if (cluster_size == 1) return src;
4177 ac_build_optimization_barrier(ctx, &src);
4178 LLVMValueRef result, swap;
4179 LLVMValueRef identity = get_reduction_identity(ctx, op,
4180 ac_get_type_size(LLVMTypeOf(src)));
4181 result = LLVMBuildBitCast(ctx->builder,
4182 ac_build_set_inactive(ctx, src, identity),
4183 LLVMTypeOf(identity), "");
4184 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
4185 result = ac_build_alu_op(ctx, result, swap, op);
4186 if (cluster_size == 2) return ac_build_wwm(ctx, result);
4187
4188 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
4189 result = ac_build_alu_op(ctx, result, swap, op);
4190 if (cluster_size == 4) return ac_build_wwm(ctx, result);
4191
4192 if (ctx->chip_class >= GFX8)
4193 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
4194 else
4195 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
4196 result = ac_build_alu_op(ctx, result, swap, op);
4197 if (cluster_size == 8) return ac_build_wwm(ctx, result);
4198
4199 if (ctx->chip_class >= GFX8)
4200 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
4201 else
4202 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
4203 result = ac_build_alu_op(ctx, result, swap, op);
4204 if (cluster_size == 16) return ac_build_wwm(ctx, result);
4205
4206 if (ctx->chip_class >= GFX10)
4207 swap = ac_build_permlane16(ctx, result, 0, true, false);
4208 else if (ctx->chip_class >= GFX8 && cluster_size != 32)
4209 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4210 else
4211 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
4212 result = ac_build_alu_op(ctx, result, swap, op);
4213 if (cluster_size == 32) return ac_build_wwm(ctx, result);
4214
4215 if (ctx->chip_class >= GFX8) {
4216 if (ctx->chip_class >= GFX10)
4217 swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
4218 else
4219 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4220 result = ac_build_alu_op(ctx, result, swap, op);
4221 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
4222 return ac_build_wwm(ctx, result);
4223 } else {
4224 swap = ac_build_readlane(ctx, result, ctx->i32_0);
4225 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
4226 result = ac_build_alu_op(ctx, result, swap, op);
4227 return ac_build_wwm(ctx, result);
4228 }
4229 }
4230
4231 /**
4232 * "Top half" of a scan that reduces per-wave values across an entire
4233 * workgroup.
4234 *
4235 * The source value must be present in the highest lane of the wave, and the
4236 * highest lane must be live.
4237 */
4238 void
4239 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4240 {
4241 if (ws->maxwaves <= 1)
4242 return;
4243
4244 const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
4245 LLVMBuilderRef builder = ctx->builder;
4246 LLVMValueRef tid = ac_get_thread_id(ctx);
4247 LLVMValueRef tmp;
4248
4249 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
4250 ac_build_ifcc(ctx, tmp, 1000);
4251 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
4252 ac_build_endif(ctx, 1000);
4253 }
4254
4255 /**
4256 * "Bottom half" of a scan that reduces per-wave values across an entire
4257 * workgroup.
4258 *
4259 * The caller must place a barrier between the top and bottom halves.
4260 */
4261 void
4262 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4263 {
4264 const LLVMTypeRef type = LLVMTypeOf(ws->src);
4265 const LLVMValueRef identity =
4266 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
4267
4268 if (ws->maxwaves <= 1) {
4269 ws->result_reduce = ws->src;
4270 ws->result_inclusive = ws->src;
4271 ws->result_exclusive = identity;
4272 return;
4273 }
4274 assert(ws->maxwaves <= 32);
4275
4276 LLVMBuilderRef builder = ctx->builder;
4277 LLVMValueRef tid = ac_get_thread_id(ctx);
4278 LLVMBasicBlockRef bbs[2];
4279 LLVMValueRef phivalues_scan[2];
4280 LLVMValueRef tmp, tmp2;
4281
4282 bbs[0] = LLVMGetInsertBlock(builder);
4283 phivalues_scan[0] = LLVMGetUndef(type);
4284
4285 if (ws->enable_reduce)
4286 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
4287 else if (ws->enable_inclusive)
4288 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
4289 else
4290 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
4291 ac_build_ifcc(ctx, tmp, 1001);
4292 {
4293 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
4294
4295 ac_build_optimization_barrier(ctx, &tmp);
4296
4297 bbs[1] = LLVMGetInsertBlock(builder);
4298 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves, true);
4299 }
4300 ac_build_endif(ctx, 1001);
4301
4302 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
4303
4304 if (ws->enable_reduce) {
4305 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
4306 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
4307 }
4308 if (ws->enable_inclusive)
4309 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
4310 if (ws->enable_exclusive) {
4311 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
4312 tmp = ac_build_readlane(ctx, scan, tmp);
4313 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
4314 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
4315 }
4316 }
4317
4318 /**
4319 * Inclusive scan of a per-wave value across an entire workgroup.
4320 *
4321 * This implies an s_barrier instruction.
4322 *
4323 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4324 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4325 * useful manner because of the barrier in the algorithm.)
4326 */
4327 void
4328 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4329 {
4330 ac_build_wg_wavescan_top(ctx, ws);
4331 ac_build_s_barrier(ctx);
4332 ac_build_wg_wavescan_bottom(ctx, ws);
4333 }
4334
4335 /**
4336 * "Top half" of a scan that reduces per-thread values across an entire
4337 * workgroup.
4338 *
4339 * All lanes must be active when this code runs.
4340 */
4341 void
4342 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4343 {
4344 if (ws->enable_exclusive) {
4345 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
4346 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
4347 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
4348 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
4349 } else {
4350 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
4351 }
4352
4353 bool enable_inclusive = ws->enable_inclusive;
4354 bool enable_exclusive = ws->enable_exclusive;
4355 ws->enable_inclusive = false;
4356 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4357 ac_build_wg_wavescan_top(ctx, ws);
4358 ws->enable_inclusive = enable_inclusive;
4359 ws->enable_exclusive = enable_exclusive;
4360 }
4361
4362 /**
4363 * "Bottom half" of a scan that reduces per-thread values across an entire
4364 * workgroup.
4365 *
4366 * The caller must place a barrier between the top and bottom halves.
4367 */
4368 void
4369 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4370 {
4371 bool enable_inclusive = ws->enable_inclusive;
4372 bool enable_exclusive = ws->enable_exclusive;
4373 ws->enable_inclusive = false;
4374 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4375 ac_build_wg_wavescan_bottom(ctx, ws);
4376 ws->enable_inclusive = enable_inclusive;
4377 ws->enable_exclusive = enable_exclusive;
4378
4379 /* ws->result_reduce is already the correct value */
4380 if (ws->enable_inclusive)
4381 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_inclusive, ws->src, ws->op);
4382 if (ws->enable_exclusive)
4383 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
4384 }
4385
4386 /**
4387 * A scan that reduces per-thread values across an entire workgroup.
4388 *
4389 * The caller must ensure that all lanes are active when this code runs
4390 * (WWM is insufficient!), because there is an implied barrier.
4391 */
4392 void
4393 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4394 {
4395 ac_build_wg_scan_top(ctx, ws);
4396 ac_build_s_barrier(ctx);
4397 ac_build_wg_scan_bottom(ctx, ws);
4398 }
4399
4400 LLVMValueRef
4401 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
4402 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
4403 {
4404 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
4405 if (ctx->chip_class >= GFX8) {
4406 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
4407 } else {
4408 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
4409 }
4410 }
4411
4412 LLVMValueRef
4413 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
4414 {
4415 LLVMTypeRef type = LLVMTypeOf(src);
4416 LLVMValueRef result;
4417
4418 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4419 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
4420
4421 result = ac_build_intrinsic(ctx, "llvm.amdgcn.ds.bpermute", ctx->i32,
4422 (LLVMValueRef []) {index, src}, 2,
4423 AC_FUNC_ATTR_READNONE |
4424 AC_FUNC_ATTR_CONVERGENT);
4425 return LLVMBuildTrunc(ctx->builder, result, type, "");
4426 }
4427
4428 LLVMValueRef
4429 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
4430 unsigned bitsize)
4431 {
4432 LLVMTypeRef type;
4433 char *intr;
4434
4435 if (bitsize == 16) {
4436 intr = "llvm.amdgcn.frexp.exp.i16.f16";
4437 type = ctx->i16;
4438 } else if (bitsize == 32) {
4439 intr = "llvm.amdgcn.frexp.exp.i32.f32";
4440 type = ctx->i32;
4441 } else {
4442 intr = "llvm.amdgcn.frexp.exp.i32.f64";
4443 type = ctx->i32;
4444 }
4445
4446 LLVMValueRef params[] = {
4447 src0,
4448 };
4449 return ac_build_intrinsic(ctx, intr, type, params, 1,
4450 AC_FUNC_ATTR_READNONE);
4451 }
4452 LLVMValueRef
4453 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
4454 unsigned bitsize)
4455 {
4456 LLVMTypeRef type;
4457 char *intr;
4458
4459 if (bitsize == 16) {
4460 intr = "llvm.amdgcn.frexp.mant.f16";
4461 type = ctx->f16;
4462 } else if (bitsize == 32) {
4463 intr = "llvm.amdgcn.frexp.mant.f32";
4464 type = ctx->f32;
4465 } else {
4466 intr = "llvm.amdgcn.frexp.mant.f64";
4467 type = ctx->f64;
4468 }
4469
4470 LLVMValueRef params[] = {
4471 src0,
4472 };
4473 return ac_build_intrinsic(ctx, intr, type, params, 1,
4474 AC_FUNC_ATTR_READNONE);
4475 }
4476
4477 LLVMValueRef
4478 ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
4479 unsigned bitsize)
4480 {
4481 LLVMTypeRef type;
4482 char *intr;
4483
4484 if (bitsize == 16) {
4485 intr = "llvm.canonicalize.f16";
4486 type = ctx->f16;
4487 } else if (bitsize == 32) {
4488 intr = "llvm.canonicalize.f32";
4489 type = ctx->f32;
4490 } else if (bitsize == 64) {
4491 intr = "llvm.canonicalize.f64";
4492 type = ctx->f64;
4493 }
4494
4495 LLVMValueRef params[] = {
4496 src0,
4497 };
4498 return ac_build_intrinsic(ctx, intr, type, params, 1,
4499 AC_FUNC_ATTR_READNONE);
4500 }
4501
4502 /*
4503 * this takes an I,J coordinate pair,
4504 * and works out the X and Y derivatives.
4505 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4506 */
4507 LLVMValueRef
4508 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
4509 {
4510 LLVMValueRef result[4], a;
4511 unsigned i;
4512
4513 for (i = 0; i < 2; i++) {
4514 a = LLVMBuildExtractElement(ctx->builder, interp_ij,
4515 LLVMConstInt(ctx->i32, i, false), "");
4516 result[i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 1, a);
4517 result[2+i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 2, a);
4518 }
4519 return ac_build_gather_values(ctx, result, 4);
4520 }
4521
4522 LLVMValueRef
4523 ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
4524 {
4525 LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4526 ctx->i1, NULL, 0,
4527 AC_FUNC_ATTR_READNONE);
4528 result = LLVMBuildNot(ctx->builder, result, "");
4529 return LLVMBuildSExt(ctx->builder, result, ctx->i32, "");
4530 }
4531
4532 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
4533 LLVMValueRef *args, unsigned num_args)
4534 {
4535 LLVMValueRef ret = LLVMBuildCall(ctx->builder, func, args, num_args, "");
4536 LLVMSetInstructionCallConv(ret, LLVMGetFunctionCallConv(func));
4537 return ret;
4538 }
4539
4540 void
4541 ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
4542 LLVMValueRef stencil, LLVMValueRef samplemask,
4543 struct ac_export_args *args)
4544 {
4545 unsigned mask = 0;
4546 unsigned format = ac_get_spi_shader_z_format(depth != NULL,
4547 stencil != NULL,
4548 samplemask != NULL);
4549
4550 assert(depth || stencil || samplemask);
4551
4552 memset(args, 0, sizeof(*args));
4553
4554 args->valid_mask = 1; /* whether the EXEC mask is valid */
4555 args->done = 1; /* DONE bit */
4556
4557 /* Specify the target we are exporting */
4558 args->target = V_008DFC_SQ_EXP_MRTZ;
4559
4560 args->compr = 0; /* COMP flag */
4561 args->out[0] = LLVMGetUndef(ctx->f32); /* R, depth */
4562 args->out[1] = LLVMGetUndef(ctx->f32); /* G, stencil test val[0:7], stencil op val[8:15] */
4563 args->out[2] = LLVMGetUndef(ctx->f32); /* B, sample mask */
4564 args->out[3] = LLVMGetUndef(ctx->f32); /* A, alpha to mask */
4565
4566 if (format == V_028710_SPI_SHADER_UINT16_ABGR) {
4567 assert(!depth);
4568 args->compr = 1; /* COMPR flag */
4569
4570 if (stencil) {
4571 /* Stencil should be in X[23:16]. */
4572 stencil = ac_to_integer(ctx, stencil);
4573 stencil = LLVMBuildShl(ctx->builder, stencil,
4574 LLVMConstInt(ctx->i32, 16, 0), "");
4575 args->out[0] = ac_to_float(ctx, stencil);
4576 mask |= 0x3;
4577 }
4578 if (samplemask) {
4579 /* SampleMask should be in Y[15:0]. */
4580 args->out[1] = samplemask;
4581 mask |= 0xc;
4582 }
4583 } else {
4584 if (depth) {
4585 args->out[0] = depth;
4586 mask |= 0x1;
4587 }
4588 if (stencil) {
4589 args->out[1] = stencil;
4590 mask |= 0x2;
4591 }
4592 if (samplemask) {
4593 args->out[2] = samplemask;
4594 mask |= 0x4;
4595 }
4596 }
4597
4598 /* GFX6 (except OLAND and HAINAN) has a bug that it only looks
4599 * at the X writemask component. */
4600 if (ctx->chip_class == GFX6 &&
4601 ctx->family != CHIP_OLAND &&
4602 ctx->family != CHIP_HAINAN)
4603 mask |= 0x1;
4604
4605 /* Specify which components to enable */
4606 args->enabled_channels = mask;
4607 }
4608