amd: Move all amd/common code that depends on LLVM to amd/llvm.
[mesa.git] / src / amd / llvm / ac_llvm_util.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_util.h"
27 #include "ac_llvm_build.h"
28 #include "util/bitscan.h"
29 #include <llvm-c/Core.h>
30 #include <llvm-c/Support.h>
31 #include <llvm-c/Transforms/IPO.h>
32 #include <llvm-c/Transforms/Scalar.h>
33 #include <llvm-c/Transforms/Utils.h>
34 #include "c11/threads.h"
35 #include "gallivm/lp_bld_misc.h"
36 #include "util/u_math.h"
37
38 #include <assert.h>
39 #include <stdio.h>
40 #include <string.h>
41
42 static void ac_init_llvm_target()
43 {
44 LLVMInitializeAMDGPUTargetInfo();
45 LLVMInitializeAMDGPUTarget();
46 LLVMInitializeAMDGPUTargetMC();
47 LLVMInitializeAMDGPUAsmPrinter();
48
49 /* For inline assembly. */
50 LLVMInitializeAMDGPUAsmParser();
51
52 /* For ACO disassembly. */
53 LLVMInitializeAMDGPUDisassembler();
54
55 /* Workaround for bug in llvm 4.0 that causes image intrinsics
56 * to disappear.
57 * https://reviews.llvm.org/D26348
58 *
59 * "mesa" is the prefix for error messages.
60 *
61 * -global-isel-abort=2 is a no-op unless global isel has been enabled.
62 * This option tells the backend to fall-back to SelectionDAG and print
63 * a diagnostic message if global isel fails.
64 */
65 const char *argv[] = {
66 "mesa",
67 "-simplifycfg-sink-common=false",
68 "-global-isel-abort=2",
69 #if LLVM_VERSION_MAJOR >= 10
70 /* Atomic optimizations require LLVM 10.0 for gfx10 support. */
71 "-amdgpu-atomic-optimizations=true",
72 #endif
73 };
74 LLVMParseCommandLineOptions(ARRAY_SIZE(argv), argv, NULL);
75 }
76
77 static once_flag ac_init_llvm_target_once_flag = ONCE_FLAG_INIT;
78
79 void ac_init_llvm_once(void)
80 {
81 call_once(&ac_init_llvm_target_once_flag, ac_init_llvm_target);
82 }
83
84 static LLVMTargetRef ac_get_llvm_target(const char *triple)
85 {
86 LLVMTargetRef target = NULL;
87 char *err_message = NULL;
88
89 if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
90 fprintf(stderr, "Cannot find target for triple %s ", triple);
91 if (err_message) {
92 fprintf(stderr, "%s\n", err_message);
93 }
94 LLVMDisposeMessage(err_message);
95 return NULL;
96 }
97 return target;
98 }
99
100 const char *ac_get_llvm_processor_name(enum radeon_family family)
101 {
102 switch (family) {
103 case CHIP_TAHITI:
104 return "tahiti";
105 case CHIP_PITCAIRN:
106 return "pitcairn";
107 case CHIP_VERDE:
108 return "verde";
109 case CHIP_OLAND:
110 return "oland";
111 case CHIP_HAINAN:
112 return "hainan";
113 case CHIP_BONAIRE:
114 return "bonaire";
115 case CHIP_KABINI:
116 return "kabini";
117 case CHIP_KAVERI:
118 return "kaveri";
119 case CHIP_HAWAII:
120 return "hawaii";
121 case CHIP_TONGA:
122 return "tonga";
123 case CHIP_ICELAND:
124 return "iceland";
125 case CHIP_CARRIZO:
126 return "carrizo";
127 case CHIP_FIJI:
128 return "fiji";
129 case CHIP_STONEY:
130 return "stoney";
131 case CHIP_POLARIS10:
132 return "polaris10";
133 case CHIP_POLARIS11:
134 case CHIP_POLARIS12:
135 case CHIP_VEGAM:
136 return "polaris11";
137 case CHIP_VEGA10:
138 return "gfx900";
139 case CHIP_RAVEN:
140 return "gfx902";
141 case CHIP_VEGA12:
142 return "gfx904";
143 case CHIP_VEGA20:
144 return "gfx906";
145 case CHIP_RAVEN2:
146 case CHIP_RENOIR:
147 return "gfx909";
148 case CHIP_ARCTURUS:
149 return "gfx908";
150 case CHIP_NAVI10:
151 return "gfx1010";
152 case CHIP_NAVI12:
153 return "gfx1011";
154 case CHIP_NAVI14:
155 return "gfx1012";
156 default:
157 return "";
158 }
159 }
160
161 static LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family,
162 enum ac_target_machine_options tm_options,
163 LLVMCodeGenOptLevel level,
164 const char **out_triple)
165 {
166 assert(family >= CHIP_TAHITI);
167 char features[256];
168 const char *triple = (tm_options & AC_TM_SUPPORTS_SPILL) ? "amdgcn-mesa-mesa3d" : "amdgcn--";
169 LLVMTargetRef target = ac_get_llvm_target(triple);
170
171 snprintf(features, sizeof(features),
172 "+DumpCode,-fp32-denormals,+fp64-denormals%s%s%s%s%s%s",
173 family >= CHIP_NAVI10 && !(tm_options & AC_TM_WAVE32) ?
174 ",+wavefrontsize64,-wavefrontsize32" : "",
175 tm_options & AC_TM_SISCHED ? ",+si-scheduler" : "",
176 tm_options & AC_TM_FORCE_ENABLE_XNACK ? ",+xnack" : "",
177 tm_options & AC_TM_FORCE_DISABLE_XNACK ? ",-xnack" : "",
178 tm_options & AC_TM_PROMOTE_ALLOCA_TO_SCRATCH ? ",-promote-alloca" : "",
179 tm_options & AC_TM_NO_LOAD_STORE_OPT ? ",-load-store-opt" : "");
180
181 LLVMTargetMachineRef tm = LLVMCreateTargetMachine(
182 target,
183 triple,
184 ac_get_llvm_processor_name(family),
185 features,
186 level,
187 LLVMRelocDefault,
188 LLVMCodeModelDefault);
189
190 if (out_triple)
191 *out_triple = triple;
192 if (tm_options & AC_TM_ENABLE_GLOBAL_ISEL)
193 ac_enable_global_isel(tm);
194 return tm;
195 }
196
197 static LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_library_info,
198 bool check_ir)
199 {
200 LLVMPassManagerRef passmgr = LLVMCreatePassManager();
201 if (!passmgr)
202 return NULL;
203
204 if (target_library_info)
205 LLVMAddTargetLibraryInfo(target_library_info,
206 passmgr);
207
208 if (check_ir)
209 LLVMAddVerifierPass(passmgr);
210 LLVMAddAlwaysInlinerPass(passmgr);
211 /* Normally, the pass manager runs all passes on one function before
212 * moving onto another. Adding a barrier no-op pass forces the pass
213 * manager to run the inliner on all functions first, which makes sure
214 * that the following passes are only run on the remaining non-inline
215 * function, so it removes useless work done on dead inline functions.
216 */
217 ac_llvm_add_barrier_noop_pass(passmgr);
218 /* This pass should eliminate all the load and store instructions. */
219 LLVMAddPromoteMemoryToRegisterPass(passmgr);
220 LLVMAddScalarReplAggregatesPass(passmgr);
221 LLVMAddLICMPass(passmgr);
222 LLVMAddAggressiveDCEPass(passmgr);
223 LLVMAddCFGSimplificationPass(passmgr);
224 /* This is recommended by the instruction combining pass. */
225 LLVMAddEarlyCSEMemSSAPass(passmgr);
226 LLVMAddInstructionCombiningPass(passmgr);
227 return passmgr;
228 }
229
230 static const char *attr_to_str(enum ac_func_attr attr)
231 {
232 switch (attr) {
233 case AC_FUNC_ATTR_ALWAYSINLINE: return "alwaysinline";
234 case AC_FUNC_ATTR_INREG: return "inreg";
235 case AC_FUNC_ATTR_NOALIAS: return "noalias";
236 case AC_FUNC_ATTR_NOUNWIND: return "nounwind";
237 case AC_FUNC_ATTR_READNONE: return "readnone";
238 case AC_FUNC_ATTR_READONLY: return "readonly";
239 case AC_FUNC_ATTR_WRITEONLY: return "writeonly";
240 case AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY: return "inaccessiblememonly";
241 case AC_FUNC_ATTR_CONVERGENT: return "convergent";
242 default:
243 fprintf(stderr, "Unhandled function attribute: %x\n", attr);
244 return 0;
245 }
246 }
247
248 void
249 ac_add_function_attr(LLVMContextRef ctx, LLVMValueRef function,
250 int attr_idx, enum ac_func_attr attr)
251 {
252 const char *attr_name = attr_to_str(attr);
253 unsigned kind_id = LLVMGetEnumAttributeKindForName(attr_name,
254 strlen(attr_name));
255 LLVMAttributeRef llvm_attr = LLVMCreateEnumAttribute(ctx, kind_id, 0);
256
257 if (LLVMIsAFunction(function))
258 LLVMAddAttributeAtIndex(function, attr_idx, llvm_attr);
259 else
260 LLVMAddCallSiteAttribute(function, attr_idx, llvm_attr);
261 }
262
263 void ac_add_func_attributes(LLVMContextRef ctx, LLVMValueRef function,
264 unsigned attrib_mask)
265 {
266 attrib_mask |= AC_FUNC_ATTR_NOUNWIND;
267 attrib_mask &= ~AC_FUNC_ATTR_LEGACY;
268
269 while (attrib_mask) {
270 enum ac_func_attr attr = 1u << u_bit_scan(&attrib_mask);
271 ac_add_function_attr(ctx, function, -1, attr);
272 }
273 }
274
275 void
276 ac_dump_module(LLVMModuleRef module)
277 {
278 char *str = LLVMPrintModuleToString(module);
279 fprintf(stderr, "%s", str);
280 LLVMDisposeMessage(str);
281 }
282
283 void
284 ac_llvm_add_target_dep_function_attr(LLVMValueRef F,
285 const char *name, unsigned value)
286 {
287 char str[16];
288
289 snprintf(str, sizeof(str), "0x%x", value);
290 LLVMAddTargetDependentFunctionAttr(F, name, str);
291 }
292
293 void ac_llvm_set_workgroup_size(LLVMValueRef F, unsigned size)
294 {
295 if (!size)
296 return;
297
298 char str[32];
299 snprintf(str, sizeof(str), "%u,%u", size, size);
300 LLVMAddTargetDependentFunctionAttr(F, "amdgpu-flat-work-group-size", str);
301 }
302
303 unsigned
304 ac_count_scratch_private_memory(LLVMValueRef function)
305 {
306 unsigned private_mem_vgprs = 0;
307
308 /* Process all LLVM instructions. */
309 LLVMBasicBlockRef bb = LLVMGetFirstBasicBlock(function);
310 while (bb) {
311 LLVMValueRef next = LLVMGetFirstInstruction(bb);
312
313 while (next) {
314 LLVMValueRef inst = next;
315 next = LLVMGetNextInstruction(next);
316
317 if (LLVMGetInstructionOpcode(inst) != LLVMAlloca)
318 continue;
319
320 LLVMTypeRef type = LLVMGetElementType(LLVMTypeOf(inst));
321 /* No idea why LLVM aligns allocas to 4 elements. */
322 unsigned alignment = LLVMGetAlignment(inst);
323 unsigned dw_size = align(ac_get_type_size(type) / 4, alignment);
324 private_mem_vgprs += dw_size;
325 }
326 bb = LLVMGetNextBasicBlock(bb);
327 }
328
329 return private_mem_vgprs;
330 }
331
332 bool
333 ac_init_llvm_compiler(struct ac_llvm_compiler *compiler,
334 enum radeon_family family,
335 enum ac_target_machine_options tm_options)
336 {
337 const char *triple;
338 memset(compiler, 0, sizeof(*compiler));
339
340 compiler->tm = ac_create_target_machine(family, tm_options,
341 LLVMCodeGenLevelDefault,
342 &triple);
343 if (!compiler->tm)
344 return false;
345
346 if (tm_options & AC_TM_CREATE_LOW_OPT) {
347 compiler->low_opt_tm =
348 ac_create_target_machine(family, tm_options,
349 LLVMCodeGenLevelLess, NULL);
350 if (!compiler->low_opt_tm)
351 goto fail;
352 }
353
354 if (family >= CHIP_NAVI10) {
355 assert(!(tm_options & AC_TM_CREATE_LOW_OPT));
356 compiler->tm_wave32 = ac_create_target_machine(family,
357 tm_options | AC_TM_WAVE32,
358 LLVMCodeGenLevelDefault,
359 NULL);
360 if (!compiler->tm_wave32)
361 goto fail;
362 }
363
364 compiler->target_library_info =
365 ac_create_target_library_info(triple);
366 if (!compiler->target_library_info)
367 goto fail;
368
369 compiler->passmgr = ac_create_passmgr(compiler->target_library_info,
370 tm_options & AC_TM_CHECK_IR);
371 if (!compiler->passmgr)
372 goto fail;
373
374 return true;
375 fail:
376 ac_destroy_llvm_compiler(compiler);
377 return false;
378 }
379
380 void
381 ac_destroy_llvm_compiler(struct ac_llvm_compiler *compiler)
382 {
383 ac_destroy_llvm_passes(compiler->passes);
384 ac_destroy_llvm_passes(compiler->passes_wave32);
385 ac_destroy_llvm_passes(compiler->low_opt_passes);
386
387 if (compiler->passmgr)
388 LLVMDisposePassManager(compiler->passmgr);
389 if (compiler->target_library_info)
390 ac_dispose_target_library_info(compiler->target_library_info);
391 if (compiler->low_opt_tm)
392 LLVMDisposeTargetMachine(compiler->low_opt_tm);
393 if (compiler->tm)
394 LLVMDisposeTargetMachine(compiler->tm);
395 if (compiler->tm_wave32)
396 LLVMDisposeTargetMachine(compiler->tm_wave32);
397 }