projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
amd/common: move ac_build_phi from radeonsi
[mesa.git] / src / amd / common / ac_llvm_build.c
1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
11 *
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
19 *
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
22 * of the Software.
23 *
24 */
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
27
28 #include <llvm-c/Core.h>
29
30 #include "c11/threads.h"
31
32 #include <assert.h>
33 #include <stdio.h>
34
35 #include "ac_llvm_util.h"
36 #include "ac_exp_param.h"
37 #include "util/bitscan.h"
38 #include "util/macros.h"
39 #include "util/u_atomic.h"
40 #include "sid.h"
41
42 #include "shader_enums.h"
43
44 /* Initialize module-independent parts of the context.
45 *
46 * The caller is responsible for initializing ctx::module and ctx::builder.
47 */
48 void
49 ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context,
50 enum chip_class chip_class)
51 {
52 LLVMValueRef args[1];
53
54 ctx->chip_class = chip_class;
55
56 ctx->context = context;
57 ctx->module = NULL;
58 ctx->builder = NULL;
59
60 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
61 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
62 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
63 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
64 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
65 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
66 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
67 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
68 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
69 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
70 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
71 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
72
73 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
74 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
75 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
76 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
77
78 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
79 "range", 5);
80
81 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
82 "invariant.load", 14);
83
84 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
85
86 args[0] = LLVMConstReal(ctx->f32, 2.5);
87 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
88
89 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
90 "amdgpu.uniform", 14);
91
92 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
93 }
94
95 unsigned
96 ac_get_type_size(LLVMTypeRef type)
97 {
98 LLVMTypeKind kind = LLVMGetTypeKind(type);
99
100 switch (kind) {
101 case LLVMIntegerTypeKind:
102 return LLVMGetIntTypeWidth(type) / 8;
103 case LLVMFloatTypeKind:
104 return 4;
105 case LLVMDoubleTypeKind:
106 case LLVMPointerTypeKind:
107 return 8;
108 case LLVMVectorTypeKind:
109 return LLVMGetVectorSize(type) *
110 ac_get_type_size(LLVMGetElementType(type));
111 case LLVMArrayTypeKind:
112 return LLVMGetArrayLength(type) *
113 ac_get_type_size(LLVMGetElementType(type));
114 default:
115 assert(0);
116 return 0;
117 }
118 }
119
120 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
121 {
122 if (t == ctx->f16 || t == ctx->i16)
123 return ctx->i16;
124 else if (t == ctx->f32 || t == ctx->i32)
125 return ctx->i32;
126 else if (t == ctx->f64 || t == ctx->i64)
127 return ctx->i64;
128 else
129 unreachable("Unhandled integer size");
130 }
131
132 LLVMTypeRef
133 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
134 {
135 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
136 LLVMTypeRef elem_type = LLVMGetElementType(t);
137 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
138 LLVMGetVectorSize(t));
139 }
140 return to_integer_type_scalar(ctx, t);
141 }
142
143 LLVMValueRef
144 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
145 {
146 LLVMTypeRef type = LLVMTypeOf(v);
147 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
148 }
149
150 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
151 {
152 if (t == ctx->i16 || t == ctx->f16)
153 return ctx->f16;
154 else if (t == ctx->i32 || t == ctx->f32)
155 return ctx->f32;
156 else if (t == ctx->i64 || t == ctx->f64)
157 return ctx->f64;
158 else
159 unreachable("Unhandled float size");
160 }
161
162 LLVMTypeRef
163 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
164 {
165 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
166 LLVMTypeRef elem_type = LLVMGetElementType(t);
167 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
168 LLVMGetVectorSize(t));
169 }
170 return to_float_type_scalar(ctx, t);
171 }
172
173 LLVMValueRef
174 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
175 {
176 LLVMTypeRef type = LLVMTypeOf(v);
177 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
178 }
179
180
181 LLVMValueRef
182 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
183 LLVMTypeRef return_type, LLVMValueRef *params,
184 unsigned param_count, unsigned attrib_mask)
185 {
186 LLVMValueRef function, call;
187 bool set_callsite_attrs = HAVE_LLVM >= 0x0400 &&
188 !(attrib_mask & AC_FUNC_ATTR_LEGACY);
189
190 function = LLVMGetNamedFunction(ctx->module, name);
191 if (!function) {
192 LLVMTypeRef param_types[32], function_type;
193 unsigned i;
194
195 assert(param_count <= 32);
196
197 for (i = 0; i < param_count; ++i) {
198 assert(params[i]);
199 param_types[i] = LLVMTypeOf(params[i]);
200 }
201 function_type =
202 LLVMFunctionType(return_type, param_types, param_count, 0);
203 function = LLVMAddFunction(ctx->module, name, function_type);
204
205 LLVMSetFunctionCallConv(function, LLVMCCallConv);
206 LLVMSetLinkage(function, LLVMExternalLinkage);
207
208 if (!set_callsite_attrs)
209 ac_add_func_attributes(ctx->context, function, attrib_mask);
210 }
211
212 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
213 if (set_callsite_attrs)
214 ac_add_func_attributes(ctx->context, call, attrib_mask);
215 return call;
216 }
217
218 /**
219 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
220 * intrinsic names).
221 */
222 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
223 {
224 LLVMTypeRef elem_type = type;
225
226 assert(bufsize >= 8);
227
228 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
229 int ret = snprintf(buf, bufsize, "v%u",
230 LLVMGetVectorSize(type));
231 if (ret < 0) {
232 char *type_name = LLVMPrintTypeToString(type);
233 fprintf(stderr, "Error building type name for: %s\n",
234 type_name);
235 return;
236 }
237 elem_type = LLVMGetElementType(type);
238 buf += ret;
239 bufsize -= ret;
240 }
241 switch (LLVMGetTypeKind(elem_type)) {
242 default: break;
243 case LLVMIntegerTypeKind:
244 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
245 break;
246 case LLVMFloatTypeKind:
247 snprintf(buf, bufsize, "f32");
248 break;
249 case LLVMDoubleTypeKind:
250 snprintf(buf, bufsize, "f64");
251 break;
252 }
253 }
254
255 /**
256 * Helper function that builds an LLVM IR PHI node and immediately adds
257 * incoming edges.
258 */
259 LLVMValueRef
260 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
261 unsigned count_incoming, LLVMValueRef *values,
262 LLVMBasicBlockRef *blocks)
263 {
264 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
265 LLVMAddIncoming(phi, values, blocks, count_incoming);
266 return phi;
267 }
268
269 /* Prevent optimizations (at least of memory accesses) across the current
270 * point in the program by emitting empty inline assembly that is marked as
271 * having side effects.
272 *
273 * Optionally, a value can be passed through the inline assembly to prevent
274 * LLVM from hoisting calls to ReadNone functions.
275 */
276 void
277 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
278 LLVMValueRef *pvgpr)
279 {
280 static int counter = 0;
281
282 LLVMBuilderRef builder = ctx->builder;
283 char code[16];
284
285 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
286
287 if (!pvgpr) {
288 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
289 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
290 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
291 } else {
292 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
293 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
294 LLVMValueRef vgpr = *pvgpr;
295 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
296 unsigned vgpr_size = ac_get_type_size(vgpr_type);
297 LLVMValueRef vgpr0;
298
299 assert(vgpr_size % 4 == 0);
300
301 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
302 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
303 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
304 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
305 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
306
307 *pvgpr = vgpr;
308 }
309 }
310
311 LLVMValueRef
312 ac_build_ballot(struct ac_llvm_context *ctx,
313 LLVMValueRef value)
314 {
315 LLVMValueRef args[3] = {
316 value,
317 ctx->i32_0,
318 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
319 };
320
321 /* We currently have no other way to prevent LLVM from lifting the icmp
322 * calls to a dominating basic block.
323 */
324 ac_build_optimization_barrier(ctx, &args[0]);
325
326 if (LLVMTypeOf(args[0]) != ctx->i32)
327 args[0] = LLVMBuildBitCast(ctx->builder, args[0], ctx->i32, "");
328
329 return ac_build_intrinsic(ctx,
330 "llvm.amdgcn.icmp.i32",
331 ctx->i64, args, 3,
332 AC_FUNC_ATTR_NOUNWIND |
333 AC_FUNC_ATTR_READNONE |
334 AC_FUNC_ATTR_CONVERGENT);
335 }
336
337 LLVMValueRef
338 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
339 {
340 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
341 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
342 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
343 }
344
345 LLVMValueRef
346 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
347 {
348 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
349 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
350 LLVMConstInt(ctx->i64, 0, 0), "");
351 }
352
353 LLVMValueRef
354 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
355 {
356 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
357 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
358
359 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
360 vote_set, active_set, "");
361 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
362 vote_set,
363 LLVMConstInt(ctx->i64, 0, 0), "");
364 return LLVMBuildOr(ctx->builder, all, none, "");
365 }
366
367 LLVMValueRef
368 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
369 LLVMValueRef *values,
370 unsigned value_count,
371 unsigned value_stride,
372 bool load,
373 bool always_vector)
374 {
375 LLVMBuilderRef builder = ctx->builder;
376 LLVMValueRef vec = NULL;
377 unsigned i;
378
379 if (value_count == 1 && !always_vector) {
380 if (load)
381 return LLVMBuildLoad(builder, values[0], "");
382 return values[0];
383 } else if (!value_count)
384 unreachable("value_count is 0");
385
386 for (i = 0; i < value_count; i++) {
387 LLVMValueRef value = values[i * value_stride];
388 if (load)
389 value = LLVMBuildLoad(builder, value, "");
390
391 if (!i)
392 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
393 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
394 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
395 }
396 return vec;
397 }
398
399 LLVMValueRef
400 ac_build_gather_values(struct ac_llvm_context *ctx,
401 LLVMValueRef *values,
402 unsigned value_count)
403 {
404 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
405 }
406
407 LLVMValueRef
408 ac_build_fdiv(struct ac_llvm_context *ctx,
409 LLVMValueRef num,
410 LLVMValueRef den)
411 {
412 LLVMValueRef ret = LLVMBuildFDiv(ctx->builder, num, den, "");
413
414 if (!LLVMIsConstant(ret))
415 LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
416 return ret;
417 }
418
419 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
420 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
421 * already multiplied by two. id is the cube face number.
422 */
423 struct cube_selection_coords {
424 LLVMValueRef stc[2];
425 LLVMValueRef ma;
426 LLVMValueRef id;
427 };
428
429 static void
430 build_cube_intrinsic(struct ac_llvm_context *ctx,
431 LLVMValueRef in[3],
432 struct cube_selection_coords *out)
433 {
434 LLVMTypeRef f32 = ctx->f32;
435
436 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
437 f32, in, 3, AC_FUNC_ATTR_READNONE);
438 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
439 f32, in, 3, AC_FUNC_ATTR_READNONE);
440 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
441 f32, in, 3, AC_FUNC_ATTR_READNONE);
442 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
443 f32, in, 3, AC_FUNC_ATTR_READNONE);
444 }
445
446 /**
447 * Build a manual selection sequence for cube face sc/tc coordinates and
448 * major axis vector (multiplied by 2 for consistency) for the given
449 * vec3 \p coords, for the face implied by \p selcoords.
450 *
451 * For the major axis, we always adjust the sign to be in the direction of
452 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
453 * the selcoords major axis.
454 */
455 static void build_cube_select(struct ac_llvm_context *ctx,
456 const struct cube_selection_coords *selcoords,
457 const LLVMValueRef *coords,
458 LLVMValueRef *out_st,
459 LLVMValueRef *out_ma)
460 {
461 LLVMBuilderRef builder = ctx->builder;
462 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
463 LLVMValueRef is_ma_positive;
464 LLVMValueRef sgn_ma;
465 LLVMValueRef is_ma_z, is_not_ma_z;
466 LLVMValueRef is_ma_y;
467 LLVMValueRef is_ma_x;
468 LLVMValueRef sgn;
469 LLVMValueRef tmp;
470
471 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
472 selcoords->ma, LLVMConstReal(f32, 0.0), "");
473 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
474 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
475
476 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
477 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
478 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
479 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
480 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
481
482 /* Select sc */
483 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
484 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
485 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
486 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
487 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
488
489 /* Select tc */
490 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
491 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
492 LLVMConstReal(f32, -1.0), "");
493 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
494
495 /* Select ma */
496 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
497 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
498 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
499 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
500 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
501 }
502
503 void
504 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
505 bool is_deriv, bool is_array, bool is_lod,
506 LLVMValueRef *coords_arg,
507 LLVMValueRef *derivs_arg)
508 {
509
510 LLVMBuilderRef builder = ctx->builder;
511 struct cube_selection_coords selcoords;
512 LLVMValueRef coords[3];
513 LLVMValueRef invma;
514
515 if (is_array && !is_lod) {
516 LLVMValueRef tmp = coords_arg[3];
517 tmp = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &tmp, 1, 0);
518
519 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
520 *
521 * "For Array forms, the array layer used will be
522 *
523 * max(0, min(d−1, floor(layer+0.5)))
524 *
525 * where d is the depth of the texture array and layer
526 * comes from the component indicated in the tables below.
527 * Workaroudn for an issue where the layer is taken from a
528 * helper invocation which happens to fall on a different
529 * layer due to extrapolation."
530 *
531 * VI and earlier attempt to implement this in hardware by
532 * clamping the value of coords[2] = (8 * layer) + face.
533 * Unfortunately, this means that the we end up with the wrong
534 * face when clamping occurs.
535 *
536 * Clamp the layer earlier to work around the issue.
537 */
538 if (ctx->chip_class <= VI) {
539 LLVMValueRef ge0;
540 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
541 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
542 }
543
544 coords_arg[3] = tmp;
545 }
546
547 build_cube_intrinsic(ctx, coords_arg, &selcoords);
548
549 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
550 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
551 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
552
553 for (int i = 0; i < 2; ++i)
554 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
555
556 coords[2] = selcoords.id;
557
558 if (is_deriv && derivs_arg) {
559 LLVMValueRef derivs[4];
560 int axis;
561
562 /* Convert cube derivatives to 2D derivatives. */
563 for (axis = 0; axis < 2; axis++) {
564 LLVMValueRef deriv_st[2];
565 LLVMValueRef deriv_ma;
566
567 /* Transform the derivative alongside the texture
568 * coordinate. Mathematically, the correct formula is
569 * as follows. Assume we're projecting onto the +Z face
570 * and denote by dx/dh the derivative of the (original)
571 * X texture coordinate with respect to horizontal
572 * window coordinates. The projection onto the +Z face
573 * plane is:
574 *
575 * f(x,z) = x/z
576 *
577 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
578 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
579 *
580 * This motivatives the implementation below.
581 *
582 * Whether this actually gives the expected results for
583 * apps that might feed in derivatives obtained via
584 * finite differences is anyone's guess. The OpenGL spec
585 * seems awfully quiet about how textureGrad for cube
586 * maps should be handled.
587 */
588 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
589 deriv_st, &deriv_ma);
590
591 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
592
593 for (int i = 0; i < 2; ++i)
594 derivs[axis * 2 + i] =
595 LLVMBuildFSub(builder,
596 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
597 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
598 }
599
600 memcpy(derivs_arg, derivs, sizeof(derivs));
601 }
602
603 /* Shift the texture coordinate. This must be applied after the
604 * derivative calculation.
605 */
606 for (int i = 0; i < 2; ++i)
607 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
608
609 if (is_array) {
610 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
611 /* coords_arg.w component - array_index for cube arrays */
612 LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
613 coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
614 }
615
616 memcpy(coords_arg, coords, sizeof(coords));
617 }
618
619
620 LLVMValueRef
621 ac_build_fs_interp(struct ac_llvm_context *ctx,
622 LLVMValueRef llvm_chan,
623 LLVMValueRef attr_number,
624 LLVMValueRef params,
625 LLVMValueRef i,
626 LLVMValueRef j)
627 {
628 LLVMValueRef args[5];
629 LLVMValueRef p1;
630
631 if (HAVE_LLVM < 0x0400) {
632 LLVMValueRef ij[2];
633 ij[0] = LLVMBuildBitCast(ctx->builder, i, ctx->i32, "");
634 ij[1] = LLVMBuildBitCast(ctx->builder, j, ctx->i32, "");
635
636 args[0] = llvm_chan;
637 args[1] = attr_number;
638 args[2] = params;
639 args[3] = ac_build_gather_values(ctx, ij, 2);
640 return ac_build_intrinsic(ctx, "llvm.SI.fs.interp",
641 ctx->f32, args, 4,
642 AC_FUNC_ATTR_READNONE);
643 }
644
645 args[0] = i;
646 args[1] = llvm_chan;
647 args[2] = attr_number;
648 args[3] = params;
649
650 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
651 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
652
653 args[0] = p1;
654 args[1] = j;
655 args[2] = llvm_chan;
656 args[3] = attr_number;
657 args[4] = params;
658
659 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
660 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
661 }
662
663 LLVMValueRef
664 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
665 LLVMValueRef parameter,
666 LLVMValueRef llvm_chan,
667 LLVMValueRef attr_number,
668 LLVMValueRef params)
669 {
670 LLVMValueRef args[4];
671 if (HAVE_LLVM < 0x0400) {
672 args[0] = llvm_chan;
673 args[1] = attr_number;
674 args[2] = params;
675
676 return ac_build_intrinsic(ctx,
677 "llvm.SI.fs.constant",
678 ctx->f32, args, 3,
679 AC_FUNC_ATTR_READNONE);
680 }
681
682 args[0] = parameter;
683 args[1] = llvm_chan;
684 args[2] = attr_number;
685 args[3] = params;
686
687 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
688 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
689 }
690
691 LLVMValueRef
692 ac_build_gep0(struct ac_llvm_context *ctx,
693 LLVMValueRef base_ptr,
694 LLVMValueRef index)
695 {
696 LLVMValueRef indices[2] = {
697 LLVMConstInt(ctx->i32, 0, 0),
698 index,
699 };
700 return LLVMBuildGEP(ctx->builder, base_ptr,
701 indices, 2, "");
702 }
703
704 void
705 ac_build_indexed_store(struct ac_llvm_context *ctx,
706 LLVMValueRef base_ptr, LLVMValueRef index,
707 LLVMValueRef value)
708 {
709 LLVMBuildStore(ctx->builder, value,
710 ac_build_gep0(ctx, base_ptr, index));
711 }
712
713 /**
714 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
715 * It's equivalent to doing a load from &base_ptr[index].
716 *
717 * \param base_ptr Where the array starts.
718 * \param index The element index into the array.
719 * \param uniform Whether the base_ptr and index can be assumed to be
720 * dynamically uniform
721 */
722 LLVMValueRef
723 ac_build_indexed_load(struct ac_llvm_context *ctx,
724 LLVMValueRef base_ptr, LLVMValueRef index,
725 bool uniform)
726 {
727 LLVMValueRef pointer;
728
729 pointer = ac_build_gep0(ctx, base_ptr, index);
730 if (uniform)
731 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
732 return LLVMBuildLoad(ctx->builder, pointer, "");
733 }
734
735 /**
736 * Do a load from &base_ptr[index], but also add a flag that it's loading
737 * a constant from a dynamically uniform index.
738 */
739 LLVMValueRef
740 ac_build_indexed_load_const(struct ac_llvm_context *ctx,
741 LLVMValueRef base_ptr, LLVMValueRef index)
742 {
743 LLVMValueRef result = ac_build_indexed_load(ctx, base_ptr, index, true);
744 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
745 return result;
746 }
747
748 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
749 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
750 * or v4i32 (num_channels=3,4).
751 */
752 void
753 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
754 LLVMValueRef rsrc,
755 LLVMValueRef vdata,
756 unsigned num_channels,
757 LLVMValueRef voffset,
758 LLVMValueRef soffset,
759 unsigned inst_offset,
760 bool glc,
761 bool slc,
762 bool writeonly_memory,
763 bool has_add_tid)
764 {
765 /* TODO: Fix stores with ADD_TID and remove the "has_add_tid" flag. */
766 if (!has_add_tid) {
767 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
768 * intrinsics. */
769 if (num_channels == 3) {
770 LLVMValueRef v[3], v01;
771
772 for (int i = 0; i < 3; i++) {
773 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
774 LLVMConstInt(ctx->i32, i, 0), "");
775 }
776 v01 = ac_build_gather_values(ctx, v, 2);
777
778 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
779 soffset, inst_offset, glc, slc,
780 writeonly_memory, has_add_tid);
781 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
782 soffset, inst_offset + 8,
783 glc, slc,
784 writeonly_memory, has_add_tid);
785 return;
786 }
787
788 unsigned func = CLAMP(num_channels, 1, 3) - 1;
789 static const char *types[] = {"f32", "v2f32", "v4f32"};
790 char name[256];
791 LLVMValueRef offset = soffset;
792
793 if (inst_offset)
794 offset = LLVMBuildAdd(ctx->builder, offset,
795 LLVMConstInt(ctx->i32, inst_offset, 0), "");
796 if (voffset)
797 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
798
799 LLVMValueRef args[] = {
800 ac_to_float(ctx, vdata),
801 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
802 LLVMConstInt(ctx->i32, 0, 0),
803 offset,
804 LLVMConstInt(ctx->i1, glc, 0),
805 LLVMConstInt(ctx->i1, slc, 0),
806 };
807
808 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
809 types[func]);
810
811 ac_build_intrinsic(ctx, name, ctx->voidt,
812 args, ARRAY_SIZE(args),
813 writeonly_memory ?
814 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
815 AC_FUNC_ATTR_WRITEONLY);
816 return;
817 }
818
819 static unsigned dfmt[] = {
820 V_008F0C_BUF_DATA_FORMAT_32,
821 V_008F0C_BUF_DATA_FORMAT_32_32,
822 V_008F0C_BUF_DATA_FORMAT_32_32_32,
823 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
824 };
825 assert(num_channels >= 1 && num_channels <= 4);
826
827 LLVMValueRef args[] = {
828 rsrc,
829 vdata,
830 LLVMConstInt(ctx->i32, num_channels, 0),
831 voffset ? voffset : LLVMGetUndef(ctx->i32),
832 soffset,
833 LLVMConstInt(ctx->i32, inst_offset, 0),
834 LLVMConstInt(ctx->i32, dfmt[num_channels - 1], 0),
835 LLVMConstInt(ctx->i32, V_008F0C_BUF_NUM_FORMAT_UINT, 0),
836 LLVMConstInt(ctx->i32, voffset != NULL, 0),
837 LLVMConstInt(ctx->i32, 0, 0), /* idxen */
838 LLVMConstInt(ctx->i32, glc, 0),
839 LLVMConstInt(ctx->i32, slc, 0),
840 LLVMConstInt(ctx->i32, 0, 0), /* tfe*/
841 };
842
843 /* The instruction offset field has 12 bits */
844 assert(voffset || inst_offset < (1 << 12));
845
846 /* The intrinsic is overloaded, we need to add a type suffix for overloading to work. */
847 unsigned func = CLAMP(num_channels, 1, 3) - 1;
848 const char *types[] = {"i32", "v2i32", "v4i32"};
849 char name[256];
850 snprintf(name, sizeof(name), "llvm.SI.tbuffer.store.%s", types[func]);
851
852 ac_build_intrinsic(ctx, name, ctx->voidt,
853 args, ARRAY_SIZE(args),
854 AC_FUNC_ATTR_LEGACY);
855 }
856
857 LLVMValueRef
858 ac_build_buffer_load(struct ac_llvm_context *ctx,
859 LLVMValueRef rsrc,
860 int num_channels,
861 LLVMValueRef vindex,
862 LLVMValueRef voffset,
863 LLVMValueRef soffset,
864 unsigned inst_offset,
865 unsigned glc,
866 unsigned slc,
867 bool can_speculate,
868 bool allow_smem)
869 {
870 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
871 if (voffset)
872 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
873 if (soffset)
874 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
875
876 /* TODO: VI and later generations can use SMEM with GLC=1.*/
877 if (allow_smem && !glc && !slc) {
878 assert(vindex == NULL);
879
880 LLVMValueRef result[4];
881
882 for (int i = 0; i < num_channels; i++) {
883 if (i) {
884 offset = LLVMBuildAdd(ctx->builder, offset,
885 LLVMConstInt(ctx->i32, 4, 0), "");
886 }
887 LLVMValueRef args[2] = {rsrc, offset};
888 result[i] = ac_build_intrinsic(ctx, "llvm.SI.load.const.v4i32",
889 ctx->f32, args, 2,
890 AC_FUNC_ATTR_READNONE |
891 AC_FUNC_ATTR_LEGACY);
892 }
893 if (num_channels == 1)
894 return result[0];
895
896 if (num_channels == 3)
897 result[num_channels++] = LLVMGetUndef(ctx->f32);
898 return ac_build_gather_values(ctx, result, num_channels);
899 }
900
901 unsigned func = CLAMP(num_channels, 1, 3) - 1;
902
903 LLVMValueRef args[] = {
904 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
905 vindex ? vindex : LLVMConstInt(ctx->i32, 0, 0),
906 offset,
907 LLVMConstInt(ctx->i1, glc, 0),
908 LLVMConstInt(ctx->i1, slc, 0)
909 };
910
911 LLVMTypeRef types[] = {ctx->f32, LLVMVectorType(ctx->f32, 2),
912 ctx->v4f32};
913 const char *type_names[] = {"f32", "v2f32", "v4f32"};
914 char name[256];
915
916 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
917 type_names[func]);
918
919 return ac_build_intrinsic(ctx, name, types[func], args,
920 ARRAY_SIZE(args),
921 /* READNONE means writes can't affect it, while
922 * READONLY means that writes can affect it. */
923 can_speculate && HAVE_LLVM >= 0x0400 ?
924 AC_FUNC_ATTR_READNONE :
925 AC_FUNC_ATTR_READONLY);
926 }
927
928 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
929 LLVMValueRef rsrc,
930 LLVMValueRef vindex,
931 LLVMValueRef voffset,
932 bool can_speculate)
933 {
934 LLVMValueRef args [] = {
935 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
936 vindex,
937 voffset,
938 LLVMConstInt(ctx->i1, 0, 0), /* glc */
939 LLVMConstInt(ctx->i1, 0, 0), /* slc */
940 };
941
942 return ac_build_intrinsic(ctx,
943 "llvm.amdgcn.buffer.load.format.v4f32",
944 ctx->v4f32, args, ARRAY_SIZE(args),
945 /* READNONE means writes can't affect it, while
946 * READONLY means that writes can affect it. */
947 can_speculate && HAVE_LLVM >= 0x0400 ?
948 AC_FUNC_ATTR_READNONE :
949 AC_FUNC_ATTR_READONLY);
950 }
951
952 /**
953 * Set range metadata on an instruction. This can only be used on load and
954 * call instructions. If you know an instruction can only produce the values
955 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
956 * \p lo is the minimum value inclusive.
957 * \p hi is the maximum value exclusive.
958 */
959 static void set_range_metadata(struct ac_llvm_context *ctx,
960 LLVMValueRef value, unsigned lo, unsigned hi)
961 {
962 LLVMValueRef range_md, md_args[2];
963 LLVMTypeRef type = LLVMTypeOf(value);
964 LLVMContextRef context = LLVMGetTypeContext(type);
965
966 md_args[0] = LLVMConstInt(type, lo, false);
967 md_args[1] = LLVMConstInt(type, hi, false);
968 range_md = LLVMMDNodeInContext(context, md_args, 2);
969 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
970 }
971
972 LLVMValueRef
973 ac_get_thread_id(struct ac_llvm_context *ctx)
974 {
975 LLVMValueRef tid;
976
977 LLVMValueRef tid_args[2];
978 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
979 tid_args[1] = LLVMConstInt(ctx->i32, 0, false);
980 tid_args[1] = ac_build_intrinsic(ctx,
981 "llvm.amdgcn.mbcnt.lo", ctx->i32,
982 tid_args, 2, AC_FUNC_ATTR_READNONE);
983
984 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
985 ctx->i32, tid_args,
986 2, AC_FUNC_ATTR_READNONE);
987 set_range_metadata(ctx, tid, 0, 64);
988 return tid;
989 }
990
991 /*
992 * SI implements derivatives using the local data store (LDS)
993 * All writes to the LDS happen in all executing threads at
994 * the same time. TID is the Thread ID for the current
995 * thread and is a value between 0 and 63, representing
996 * the thread's position in the wavefront.
997 *
998 * For the pixel shader threads are grouped into quads of four pixels.
999 * The TIDs of the pixels of a quad are:
1000 *
1001 * +------+------+
1002 * |4n + 0|4n + 1|
1003 * +------+------+
1004 * |4n + 2|4n + 3|
1005 * +------+------+
1006 *
1007 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1008 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1009 * the current pixel's column, and masking with 0xfffffffe yields the TID
1010 * of the left pixel of the current pixel's row.
1011 *
1012 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1013 * adding 2 yields the TID of the pixel below the top pixel.
1014 */
1015 LLVMValueRef
1016 ac_build_ddxy(struct ac_llvm_context *ctx,
1017 uint32_t mask,
1018 int idx,
1019 LLVMValueRef val)
1020 {
1021 LLVMValueRef tl, trbl, args[2];
1022 LLVMValueRef result;
1023
1024 if (ctx->chip_class >= VI) {
1025 LLVMValueRef thread_id, tl_tid, trbl_tid;
1026 thread_id = ac_get_thread_id(ctx);
1027
1028 tl_tid = LLVMBuildAnd(ctx->builder, thread_id,
1029 LLVMConstInt(ctx->i32, mask, false), "");
1030
1031 trbl_tid = LLVMBuildAdd(ctx->builder, tl_tid,
1032 LLVMConstInt(ctx->i32, idx, false), "");
1033
1034 args[0] = LLVMBuildMul(ctx->builder, tl_tid,
1035 LLVMConstInt(ctx->i32, 4, false), "");
1036 args[1] = val;
1037 tl = ac_build_intrinsic(ctx,
1038 "llvm.amdgcn.ds.bpermute", ctx->i32,
1039 args, 2,
1040 AC_FUNC_ATTR_READNONE |
1041 AC_FUNC_ATTR_CONVERGENT);
1042
1043 args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
1044 LLVMConstInt(ctx->i32, 4, false), "");
1045 trbl = ac_build_intrinsic(ctx,
1046 "llvm.amdgcn.ds.bpermute", ctx->i32,
1047 args, 2,
1048 AC_FUNC_ATTR_READNONE |
1049 AC_FUNC_ATTR_CONVERGENT);
1050 } else {
1051 uint32_t masks[2];
1052
1053 switch (mask) {
1054 case AC_TID_MASK_TOP_LEFT:
1055 masks[0] = 0x8000;
1056 if (idx == 1)
1057 masks[1] = 0x8055;
1058 else
1059 masks[1] = 0x80aa;
1060
1061 break;
1062 case AC_TID_MASK_TOP:
1063 masks[0] = 0x8044;
1064 masks[1] = 0x80ee;
1065 break;
1066 case AC_TID_MASK_LEFT:
1067 masks[0] = 0x80a0;
1068 masks[1] = 0x80f5;
1069 break;
1070 }
1071
1072 args[0] = val;
1073 args[1] = LLVMConstInt(ctx->i32, masks[0], false);
1074
1075 tl = ac_build_intrinsic(ctx,
1076 "llvm.amdgcn.ds.swizzle", ctx->i32,
1077 args, 2,
1078 AC_FUNC_ATTR_READNONE |
1079 AC_FUNC_ATTR_CONVERGENT);
1080
1081 args[1] = LLVMConstInt(ctx->i32, masks[1], false);
1082 trbl = ac_build_intrinsic(ctx,
1083 "llvm.amdgcn.ds.swizzle", ctx->i32,
1084 args, 2,
1085 AC_FUNC_ATTR_READNONE |
1086 AC_FUNC_ATTR_CONVERGENT);
1087 }
1088
1089 tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
1090 trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
1091 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
1092 return result;
1093 }
1094
1095 void
1096 ac_build_sendmsg(struct ac_llvm_context *ctx,
1097 uint32_t msg,
1098 LLVMValueRef wave_id)
1099 {
1100 LLVMValueRef args[2];
1101 const char *intr_name = (HAVE_LLVM < 0x0400) ? "llvm.SI.sendmsg" : "llvm.amdgcn.s.sendmsg";
1102 args[0] = LLVMConstInt(ctx->i32, msg, false);
1103 args[1] = wave_id;
1104 ac_build_intrinsic(ctx, intr_name, ctx->voidt, args, 2, 0);
1105 }
1106
1107 LLVMValueRef
1108 ac_build_imsb(struct ac_llvm_context *ctx,
1109 LLVMValueRef arg,
1110 LLVMTypeRef dst_type)
1111 {
1112 const char *intr_name = (HAVE_LLVM < 0x0400) ? "llvm.AMDGPU.flbit.i32" :
1113 "llvm.amdgcn.sffbh.i32";
1114 LLVMValueRef msb = ac_build_intrinsic(ctx, intr_name,
1115 dst_type, &arg, 1,
1116 AC_FUNC_ATTR_READNONE);
1117
1118 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1119 * the index from LSB. Invert it by doing "31 - msb". */
1120 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1121 msb, "");
1122
1123 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
1124 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
1125 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1126 arg, LLVMConstInt(ctx->i32, 0, 0), ""),
1127 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1128 arg, all_ones, ""), "");
1129
1130 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
1131 }
1132
1133 LLVMValueRef
1134 ac_build_umsb(struct ac_llvm_context *ctx,
1135 LLVMValueRef arg,
1136 LLVMTypeRef dst_type)
1137 {
1138 LLVMValueRef args[2] = {
1139 arg,
1140 LLVMConstInt(ctx->i1, 1, 0),
1141 };
1142 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.ctlz.i32",
1143 dst_type, args, ARRAY_SIZE(args),
1144 AC_FUNC_ATTR_READNONE);
1145
1146 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1147 * the index from LSB. Invert it by doing "31 - msb". */
1148 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1149 msb, "");
1150
1151 /* check for zero */
1152 return LLVMBuildSelect(ctx->builder,
1153 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg,
1154 LLVMConstInt(ctx->i32, 0, 0), ""),
1155 LLVMConstInt(ctx->i32, -1, true), msb, "");
1156 }
1157
1158 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
1159 LLVMValueRef b)
1160 {
1161 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
1162 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1163 }
1164
1165 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
1166 {
1167 if (HAVE_LLVM >= 0x0500) {
1168 LLVMValueRef max[2] = {
1169 value,
1170 LLVMConstReal(ctx->f32, 0),
1171 };
1172 LLVMValueRef min[2] = {
1173 LLVMConstReal(ctx->f32, 1),
1174 };
1175
1176 min[1] = ac_build_intrinsic(ctx, "llvm.maxnum.f32",
1177 ctx->f32, max, 2,
1178 AC_FUNC_ATTR_READNONE);
1179 return ac_build_intrinsic(ctx, "llvm.minnum.f32",
1180 ctx->f32, min, 2,
1181 AC_FUNC_ATTR_READNONE);
1182 }
1183
1184 LLVMValueRef args[3] = {
1185 value,
1186 LLVMConstReal(ctx->f32, 0),
1187 LLVMConstReal(ctx->f32, 1),
1188 };
1189
1190 return ac_build_intrinsic(ctx, "llvm.AMDGPU.clamp.", ctx->f32, args, 3,
1191 AC_FUNC_ATTR_READNONE |
1192 AC_FUNC_ATTR_LEGACY);
1193 }
1194
1195 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
1196 {
1197 LLVMValueRef args[9];
1198
1199 if (HAVE_LLVM >= 0x0500) {
1200 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
1201 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1202
1203 if (a->compr) {
1204 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
1205 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
1206
1207 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
1208 v2i16, "");
1209 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
1210 v2i16, "");
1211 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
1212 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1213
1214 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
1215 ctx->voidt, args, 6, 0);
1216 } else {
1217 args[2] = a->out[0];
1218 args[3] = a->out[1];
1219 args[4] = a->out[2];
1220 args[5] = a->out[3];
1221 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
1222 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1223
1224 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
1225 ctx->voidt, args, 8, 0);
1226 }
1227 return;
1228 }
1229
1230 args[0] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1231 args[1] = LLVMConstInt(ctx->i32, a->valid_mask, 0);
1232 args[2] = LLVMConstInt(ctx->i32, a->done, 0);
1233 args[3] = LLVMConstInt(ctx->i32, a->target, 0);
1234 args[4] = LLVMConstInt(ctx->i32, a->compr, 0);
1235 memcpy(args + 5, a->out, sizeof(a->out[0]) * 4);
1236
1237 ac_build_intrinsic(ctx, "llvm.SI.export", ctx->voidt, args, 9,
1238 AC_FUNC_ATTR_LEGACY);
1239 }
1240
1241 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
1242 struct ac_image_args *a)
1243 {
1244 LLVMTypeRef dst_type;
1245 LLVMValueRef args[11];
1246 unsigned num_args = 0;
1247 const char *name;
1248 char intr_name[128], type[64];
1249
1250 if (HAVE_LLVM >= 0x0400) {
1251 bool sample = a->opcode == ac_image_sample ||
1252 a->opcode == ac_image_gather4 ||
1253 a->opcode == ac_image_get_lod;
1254
1255 if (sample)
1256 args[num_args++] = ac_to_float(ctx, a->addr);
1257 else
1258 args[num_args++] = a->addr;
1259
1260 args[num_args++] = a->resource;
1261 if (sample)
1262 args[num_args++] = a->sampler;
1263 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, 0);
1264 if (sample)
1265 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, 0);
1266 args[num_args++] = LLVMConstInt(ctx->i1, 0, 0); /* glc */
1267 args[num_args++] = LLVMConstInt(ctx->i1, 0, 0); /* slc */
1268 args[num_args++] = LLVMConstInt(ctx->i1, 0, 0); /* lwe */
1269 args[num_args++] = LLVMConstInt(ctx->i1, a->da, 0);
1270
1271 switch (a->opcode) {
1272 case ac_image_sample:
1273 name = "llvm.amdgcn.image.sample";
1274 break;
1275 case ac_image_gather4:
1276 name = "llvm.amdgcn.image.gather4";
1277 break;
1278 case ac_image_load:
1279 name = "llvm.amdgcn.image.load";
1280 break;
1281 case ac_image_load_mip:
1282 name = "llvm.amdgcn.image.load.mip";
1283 break;
1284 case ac_image_get_lod:
1285 name = "llvm.amdgcn.image.getlod";
1286 break;
1287 case ac_image_get_resinfo:
1288 name = "llvm.amdgcn.image.getresinfo";
1289 break;
1290 default:
1291 unreachable("invalid image opcode");
1292 }
1293
1294 ac_build_type_name_for_intr(LLVMTypeOf(args[0]), type,
1295 sizeof(type));
1296
1297 snprintf(intr_name, sizeof(intr_name), "%s%s%s%s.v4f32.%s.v8i32",
1298 name,
1299 a->compare ? ".c" : "",
1300 a->bias ? ".b" :
1301 a->lod ? ".l" :
1302 a->deriv ? ".d" :
1303 a->level_zero ? ".lz" : "",
1304 a->offset ? ".o" : "",
1305 type);
1306
1307 LLVMValueRef result =
1308 ac_build_intrinsic(ctx, intr_name,
1309 ctx->v4f32, args, num_args,
1310 AC_FUNC_ATTR_READNONE);
1311 if (!sample) {
1312 result = LLVMBuildBitCast(ctx->builder, result,
1313 ctx->v4i32, "");
1314 }
1315 return result;
1316 }
1317
1318 args[num_args++] = a->addr;
1319 args[num_args++] = a->resource;
1320
1321 if (a->opcode == ac_image_load ||
1322 a->opcode == ac_image_load_mip ||
1323 a->opcode == ac_image_get_resinfo) {
1324 dst_type = ctx->v4i32;
1325 } else {
1326 dst_type = ctx->v4f32;
1327 args[num_args++] = a->sampler;
1328 }
1329
1330 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, 0);
1331 args[num_args++] = LLVMConstInt(ctx->i32, a->unorm, 0);
1332 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* r128 */
1333 args[num_args++] = LLVMConstInt(ctx->i32, a->da, 0);
1334 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* glc */
1335 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* slc */
1336 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* tfe */
1337 args[num_args++] = LLVMConstInt(ctx->i32, 0, 0); /* lwe */
1338
1339 switch (a->opcode) {
1340 case ac_image_sample:
1341 name = "llvm.SI.image.sample";
1342 break;
1343 case ac_image_gather4:
1344 name = "llvm.SI.gather4";
1345 break;
1346 case ac_image_load:
1347 name = "llvm.SI.image.load";
1348 break;
1349 case ac_image_load_mip:
1350 name = "llvm.SI.image.load.mip";
1351 break;
1352 case ac_image_get_lod:
1353 name = "llvm.SI.getlod";
1354 break;
1355 case ac_image_get_resinfo:
1356 name = "llvm.SI.getresinfo";
1357 break;
1358 }
1359
1360 ac_build_type_name_for_intr(LLVMTypeOf(a->addr), type, sizeof(type));
1361 snprintf(intr_name, sizeof(intr_name), "%s%s%s%s.%s",
1362 name,
1363 a->compare ? ".c" : "",
1364 a->bias ? ".b" :
1365 a->lod ? ".l" :
1366 a->deriv ? ".d" :
1367 a->level_zero ? ".lz" : "",
1368 a->offset ? ".o" : "",
1369 type);
1370
1371 return ac_build_intrinsic(ctx, intr_name,
1372 dst_type, args, num_args,
1373 AC_FUNC_ATTR_READNONE |
1374 AC_FUNC_ATTR_LEGACY);
1375 }
1376
1377 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
1378 LLVMValueRef args[2])
1379 {
1380 if (HAVE_LLVM >= 0x0500) {
1381 LLVMTypeRef v2f16 =
1382 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
1383 LLVMValueRef res =
1384 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz",
1385 v2f16, args, 2,
1386 AC_FUNC_ATTR_READNONE);
1387 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1388 }
1389
1390 return ac_build_intrinsic(ctx, "llvm.SI.packf16", ctx->i32, args, 2,
1391 AC_FUNC_ATTR_READNONE |
1392 AC_FUNC_ATTR_LEGACY);
1393 }
1394
1395 /**
1396 * KILL, AKA discard in GLSL.
1397 *
1398 * \param value kill if value < 0.0 or value == NULL.
1399 */
1400 void ac_build_kill(struct ac_llvm_context *ctx, LLVMValueRef value)
1401 {
1402 if (value) {
1403 ac_build_intrinsic(ctx, "llvm.AMDGPU.kill", ctx->voidt,
1404 &value, 1, AC_FUNC_ATTR_LEGACY);
1405 } else {
1406 ac_build_intrinsic(ctx, "llvm.AMDGPU.kilp", ctx->voidt,
1407 NULL, 0, AC_FUNC_ATTR_LEGACY);
1408 }
1409 }
1410
1411 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
1412 LLVMValueRef offset, LLVMValueRef width,
1413 bool is_signed)
1414 {
1415 LLVMValueRef args[] = {
1416 input,
1417 offset,
1418 width,
1419 };
1420
1421 if (HAVE_LLVM >= 0x0500) {
1422 return ac_build_intrinsic(ctx,
1423 is_signed ? "llvm.amdgcn.sbfe.i32" :
1424 "llvm.amdgcn.ubfe.i32",
1425 ctx->i32, args, 3,
1426 AC_FUNC_ATTR_READNONE);
1427 }
1428
1429 return ac_build_intrinsic(ctx,
1430 is_signed ? "llvm.AMDGPU.bfe.i32" :
1431 "llvm.AMDGPU.bfe.u32",
1432 ctx->i32, args, 3,
1433 AC_FUNC_ATTR_READNONE |
1434 AC_FUNC_ATTR_LEGACY);
1435 }
1436
1437 void ac_get_image_intr_name(const char *base_name,
1438 LLVMTypeRef data_type,
1439 LLVMTypeRef coords_type,
1440 LLVMTypeRef rsrc_type,
1441 char *out_name, unsigned out_len)
1442 {
1443 char coords_type_name[8];
1444
1445 ac_build_type_name_for_intr(coords_type, coords_type_name,
1446 sizeof(coords_type_name));
1447
1448 if (HAVE_LLVM <= 0x0309) {
1449 snprintf(out_name, out_len, "%s.%s", base_name, coords_type_name);
1450 } else {
1451 char data_type_name[8];
1452 char rsrc_type_name[8];
1453
1454 ac_build_type_name_for_intr(data_type, data_type_name,
1455 sizeof(data_type_name));
1456 ac_build_type_name_for_intr(rsrc_type, rsrc_type_name,
1457 sizeof(rsrc_type_name));
1458 snprintf(out_name, out_len, "%s.%s.%s.%s", base_name,
1459 data_type_name, coords_type_name, rsrc_type_name);
1460 }
1461 }
1462
1463 #define AC_EXP_TARGET (HAVE_LLVM >= 0x0500 ? 0 : 3)
1464 #define AC_EXP_OUT0 (HAVE_LLVM >= 0x0500 ? 2 : 5)
1465
1466 enum ac_ir_type {
1467 AC_IR_UNDEF,
1468 AC_IR_CONST,
1469 AC_IR_VALUE,
1470 };
1471
1472 struct ac_vs_exp_chan
1473 {
1474 LLVMValueRef value;
1475 float const_float;
1476 enum ac_ir_type type;
1477 };
1478
1479 struct ac_vs_exp_inst {
1480 unsigned offset;
1481 LLVMValueRef inst;
1482 struct ac_vs_exp_chan chan[4];
1483 };
1484
1485 struct ac_vs_exports {
1486 unsigned num;
1487 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
1488 };
1489
1490 /* Return true if the PARAM export has been eliminated. */
1491 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
1492 uint32_t num_outputs,
1493 struct ac_vs_exp_inst *exp)
1494 {
1495 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
1496 bool is_zero[4] = {}, is_one[4] = {};
1497
1498 for (i = 0; i < 4; i++) {
1499 /* It's a constant expression. Undef outputs are eliminated too. */
1500 if (exp->chan[i].type == AC_IR_UNDEF) {
1501 is_zero[i] = true;
1502 is_one[i] = true;
1503 } else if (exp->chan[i].type == AC_IR_CONST) {
1504 if (exp->chan[i].const_float == 0)
1505 is_zero[i] = true;
1506 else if (exp->chan[i].const_float == 1)
1507 is_one[i] = true;
1508 else
1509 return false; /* other constant */
1510 } else
1511 return false;
1512 }
1513
1514 /* Only certain combinations of 0 and 1 can be eliminated. */
1515 if (is_zero[0] && is_zero[1] && is_zero[2])
1516 default_val = is_zero[3] ? 0 : 1;
1517 else if (is_one[0] && is_one[1] && is_one[2])
1518 default_val = is_zero[3] ? 2 : 3;
1519 else
1520 return false;
1521
1522 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
1523 LLVMInstructionEraseFromParent(exp->inst);
1524
1525 /* Change OFFSET to DEFAULT_VAL. */
1526 for (i = 0; i < num_outputs; i++) {
1527 if (vs_output_param_offset[i] == exp->offset) {
1528 vs_output_param_offset[i] =
1529 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
1530 break;
1531 }
1532 }
1533 return true;
1534 }
1535
1536 static bool ac_eliminate_duplicated_output(uint8_t *vs_output_param_offset,
1537 uint32_t num_outputs,
1538 struct ac_vs_exports *processed,
1539 struct ac_vs_exp_inst *exp)
1540 {
1541 unsigned p, copy_back_channels = 0;
1542
1543 /* See if the output is already in the list of processed outputs.
1544 * The LLVMValueRef comparison relies on SSA.
1545 */
1546 for (p = 0; p < processed->num; p++) {
1547 bool different = false;
1548
1549 for (unsigned j = 0; j < 4; j++) {
1550 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
1551 struct ac_vs_exp_chan *c2 = &exp->chan[j];
1552
1553 /* Treat undef as a match. */
1554 if (c2->type == AC_IR_UNDEF)
1555 continue;
1556
1557 /* If c1 is undef but c2 isn't, we can copy c2 to c1
1558 * and consider the instruction duplicated.
1559 */
1560 if (c1->type == AC_IR_UNDEF) {
1561 copy_back_channels |= 1 << j;
1562 continue;
1563 }
1564
1565 /* Test whether the channels are not equal. */
1566 if (c1->type != c2->type ||
1567 (c1->type == AC_IR_CONST &&
1568 c1->const_float != c2->const_float) ||
1569 (c1->type == AC_IR_VALUE &&
1570 c1->value != c2->value)) {
1571 different = true;
1572 break;
1573 }
1574 }
1575 if (!different)
1576 break;
1577
1578 copy_back_channels = 0;
1579 }
1580 if (p == processed->num)
1581 return false;
1582
1583 /* If a match was found, but the matching export has undef where the new
1584 * one has a normal value, copy the normal value to the undef channel.
1585 */
1586 struct ac_vs_exp_inst *match = &processed->exp[p];
1587
1588 while (copy_back_channels) {
1589 unsigned chan = u_bit_scan(&copy_back_channels);
1590
1591 assert(match->chan[chan].type == AC_IR_UNDEF);
1592 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
1593 exp->chan[chan].value);
1594 match->chan[chan] = exp->chan[chan];
1595 }
1596
1597 /* The PARAM export is duplicated. Kill it. */
1598 LLVMInstructionEraseFromParent(exp->inst);
1599
1600 /* Change OFFSET to the matching export. */
1601 for (unsigned i = 0; i < num_outputs; i++) {
1602 if (vs_output_param_offset[i] == exp->offset) {
1603 vs_output_param_offset[i] = match->offset;
1604 break;
1605 }
1606 }
1607 return true;
1608 }
1609
1610 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
1611 LLVMValueRef main_fn,
1612 uint8_t *vs_output_param_offset,
1613 uint32_t num_outputs,
1614 uint8_t *num_param_exports)
1615 {
1616 LLVMBasicBlockRef bb;
1617 bool removed_any = false;
1618 struct ac_vs_exports exports;
1619
1620 exports.num = 0;
1621
1622 /* Process all LLVM instructions. */
1623 bb = LLVMGetFirstBasicBlock(main_fn);
1624 while (bb) {
1625 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
1626
1627 while (inst) {
1628 LLVMValueRef cur = inst;
1629 inst = LLVMGetNextInstruction(inst);
1630 struct ac_vs_exp_inst exp;
1631
1632 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
1633 continue;
1634
1635 LLVMValueRef callee = ac_llvm_get_called_value(cur);
1636
1637 if (!ac_llvm_is_function(callee))
1638 continue;
1639
1640 const char *name = LLVMGetValueName(callee);
1641 unsigned num_args = LLVMCountParams(callee);
1642
1643 /* Check if this is an export instruction. */
1644 if ((num_args != 9 && num_args != 8) ||
1645 (strcmp(name, "llvm.SI.export") &&
1646 strcmp(name, "llvm.amdgcn.exp.f32")))
1647 continue;
1648
1649 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
1650 unsigned target = LLVMConstIntGetZExtValue(arg);
1651
1652 if (target < V_008DFC_SQ_EXP_PARAM)
1653 continue;
1654
1655 target -= V_008DFC_SQ_EXP_PARAM;
1656
1657 /* Parse the instruction. */
1658 memset(&exp, 0, sizeof(exp));
1659 exp.offset = target;
1660 exp.inst = cur;
1661
1662 for (unsigned i = 0; i < 4; i++) {
1663 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
1664
1665 exp.chan[i].value = v;
1666
1667 if (LLVMIsUndef(v)) {
1668 exp.chan[i].type = AC_IR_UNDEF;
1669 } else if (LLVMIsAConstantFP(v)) {
1670 LLVMBool loses_info;
1671 exp.chan[i].type = AC_IR_CONST;
1672 exp.chan[i].const_float =
1673 LLVMConstRealGetDouble(v, &loses_info);
1674 } else {
1675 exp.chan[i].type = AC_IR_VALUE;
1676 }
1677 }
1678
1679 /* Eliminate constant and duplicated PARAM exports. */
1680 if (ac_eliminate_const_output(vs_output_param_offset,
1681 num_outputs, &exp) ||
1682 ac_eliminate_duplicated_output(vs_output_param_offset,
1683 num_outputs, &exports,
1684 &exp)) {
1685 removed_any = true;
1686 } else {
1687 exports.exp[exports.num++] = exp;
1688 }
1689 }
1690 bb = LLVMGetNextBasicBlock(bb);
1691 }
1692
1693 /* Remove holes in export memory due to removed PARAM exports.
1694 * This is done by renumbering all PARAM exports.
1695 */
1696 if (removed_any) {
1697 uint8_t old_offset[VARYING_SLOT_MAX];
1698 unsigned out, i;
1699
1700 /* Make a copy of the offsets. We need the old version while
1701 * we are modifying some of them. */
1702 memcpy(old_offset, vs_output_param_offset,
1703 sizeof(old_offset));
1704
1705 for (i = 0; i < exports.num; i++) {
1706 unsigned offset = exports.exp[i].offset;
1707
1708 /* Update vs_output_param_offset. Multiple outputs can
1709 * have the same offset.
1710 */
1711 for (out = 0; out < num_outputs; out++) {
1712 if (old_offset[out] == offset)
1713 vs_output_param_offset[out] = i;
1714 }
1715
1716 /* Change the PARAM offset in the instruction. */
1717 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
1718 LLVMConstInt(ctx->i32,
1719 V_008DFC_SQ_EXP_PARAM + i, 0));
1720 }
1721 *num_param_exports = exports.num;
1722 }
1723 }