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