amd: add support for Arcturus
[mesa.git] / src / amd / common / 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 /* Workaround for bug in llvm 4.0 that causes image intrinsics
53 * to disappear.
54 * https://reviews.llvm.org/D26348
55 *
56 * "mesa" is the prefix for error messages.
57 *
58 * -global-isel-abort=2 is a no-op unless global isel has been enabled.
59 * This option tells the backend to fall-back to SelectionDAG and print
60 * a diagnostic message if global isel fails.
61 */
62 const char *argv[] = { "mesa", "-simplifycfg-sink-common=false", "-global-isel-abort=2" };
63 LLVMParseCommandLineOptions(ARRAY_SIZE(argv), argv, NULL);
64 }
65
66 static once_flag ac_init_llvm_target_once_flag = ONCE_FLAG_INIT;
67
68 void ac_init_llvm_once(void)
69 {
70 call_once(&ac_init_llvm_target_once_flag, ac_init_llvm_target);
71 }
72
73 static LLVMTargetRef ac_get_llvm_target(const char *triple)
74 {
75 LLVMTargetRef target = NULL;
76 char *err_message = NULL;
77
78 if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
79 fprintf(stderr, "Cannot find target for triple %s ", triple);
80 if (err_message) {
81 fprintf(stderr, "%s\n", err_message);
82 }
83 LLVMDisposeMessage(err_message);
84 return NULL;
85 }
86 return target;
87 }
88
89 const char *ac_get_llvm_processor_name(enum radeon_family family)
90 {
91 switch (family) {
92 case CHIP_TAHITI:
93 return "tahiti";
94 case CHIP_PITCAIRN:
95 return "pitcairn";
96 case CHIP_VERDE:
97 return "verde";
98 case CHIP_OLAND:
99 return "oland";
100 case CHIP_HAINAN:
101 return "hainan";
102 case CHIP_BONAIRE:
103 return "bonaire";
104 case CHIP_KABINI:
105 return "kabini";
106 case CHIP_KAVERI:
107 return "kaveri";
108 case CHIP_HAWAII:
109 return "hawaii";
110 case CHIP_TONGA:
111 return "tonga";
112 case CHIP_ICELAND:
113 return "iceland";
114 case CHIP_CARRIZO:
115 return "carrizo";
116 case CHIP_FIJI:
117 return "fiji";
118 case CHIP_STONEY:
119 return "stoney";
120 case CHIP_POLARIS10:
121 return "polaris10";
122 case CHIP_POLARIS11:
123 case CHIP_POLARIS12:
124 case CHIP_VEGAM:
125 return "polaris11";
126 case CHIP_VEGA10:
127 return "gfx900";
128 case CHIP_RAVEN:
129 return "gfx902";
130 case CHIP_VEGA12:
131 return "gfx904";
132 case CHIP_VEGA20:
133 return "gfx906";
134 case CHIP_RAVEN2:
135 return HAVE_LLVM >= 0x0800 ? "gfx909" : "gfx902";
136 case CHIP_ARCTURUS:
137 return "gfx908";
138 case CHIP_NAVI10:
139 return "gfx1010";
140 case CHIP_NAVI12:
141 return "gfx1011";
142 case CHIP_NAVI14:
143 return "gfx1012";
144 default:
145 return "";
146 }
147 }
148
149 static LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family,
150 enum ac_target_machine_options tm_options,
151 LLVMCodeGenOptLevel level,
152 const char **out_triple)
153 {
154 assert(family >= CHIP_TAHITI);
155 char features[256];
156 const char *triple = (tm_options & AC_TM_SUPPORTS_SPILL) ? "amdgcn-mesa-mesa3d" : "amdgcn--";
157 LLVMTargetRef target = ac_get_llvm_target(triple);
158
159 snprintf(features, sizeof(features),
160 "+DumpCode,-fp32-denormals,+fp64-denormals%s%s%s%s%s%s%s",
161 HAVE_LLVM >= 0x0800 ? "" : ",+vgpr-spilling",
162 family >= CHIP_NAVI10 && !(tm_options & AC_TM_WAVE32) ?
163 ",+wavefrontsize64,-wavefrontsize32" : "",
164 tm_options & AC_TM_SISCHED ? ",+si-scheduler" : "",
165 tm_options & AC_TM_FORCE_ENABLE_XNACK ? ",+xnack" : "",
166 tm_options & AC_TM_FORCE_DISABLE_XNACK ? ",-xnack" : "",
167 tm_options & AC_TM_PROMOTE_ALLOCA_TO_SCRATCH ? ",-promote-alloca" : "",
168 tm_options & AC_TM_NO_LOAD_STORE_OPT ? ",-load-store-opt" : "");
169
170 LLVMTargetMachineRef tm = LLVMCreateTargetMachine(
171 target,
172 triple,
173 ac_get_llvm_processor_name(family),
174 features,
175 level,
176 LLVMRelocDefault,
177 LLVMCodeModelDefault);
178
179 if (out_triple)
180 *out_triple = triple;
181 if (tm_options & AC_TM_ENABLE_GLOBAL_ISEL)
182 ac_enable_global_isel(tm);
183 return tm;
184 }
185
186 static LLVMPassManagerRef ac_create_passmgr(LLVMTargetLibraryInfoRef target_library_info,
187 bool check_ir)
188 {
189 LLVMPassManagerRef passmgr = LLVMCreatePassManager();
190 if (!passmgr)
191 return NULL;
192
193 if (target_library_info)
194 LLVMAddTargetLibraryInfo(target_library_info,
195 passmgr);
196
197 if (check_ir)
198 LLVMAddVerifierPass(passmgr);
199 LLVMAddAlwaysInlinerPass(passmgr);
200 /* Normally, the pass manager runs all passes on one function before
201 * moving onto another. Adding a barrier no-op pass forces the pass
202 * manager to run the inliner on all functions first, which makes sure
203 * that the following passes are only run on the remaining non-inline
204 * function, so it removes useless work done on dead inline functions.
205 */
206 ac_llvm_add_barrier_noop_pass(passmgr);
207 /* This pass should eliminate all the load and store instructions. */
208 LLVMAddPromoteMemoryToRegisterPass(passmgr);
209 LLVMAddScalarReplAggregatesPass(passmgr);
210 LLVMAddLICMPass(passmgr);
211 LLVMAddAggressiveDCEPass(passmgr);
212 LLVMAddCFGSimplificationPass(passmgr);
213 /* This is recommended by the instruction combining pass. */
214 LLVMAddEarlyCSEMemSSAPass(passmgr);
215 LLVMAddInstructionCombiningPass(passmgr);
216 return passmgr;
217 }
218
219 static const char *attr_to_str(enum ac_func_attr attr)
220 {
221 switch (attr) {
222 case AC_FUNC_ATTR_ALWAYSINLINE: return "alwaysinline";
223 case AC_FUNC_ATTR_INREG: return "inreg";
224 case AC_FUNC_ATTR_NOALIAS: return "noalias";
225 case AC_FUNC_ATTR_NOUNWIND: return "nounwind";
226 case AC_FUNC_ATTR_READNONE: return "readnone";
227 case AC_FUNC_ATTR_READONLY: return "readonly";
228 case AC_FUNC_ATTR_WRITEONLY: return "writeonly";
229 case AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY: return "inaccessiblememonly";
230 case AC_FUNC_ATTR_CONVERGENT: return "convergent";
231 default:
232 fprintf(stderr, "Unhandled function attribute: %x\n", attr);
233 return 0;
234 }
235 }
236
237 void
238 ac_add_function_attr(LLVMContextRef ctx, LLVMValueRef function,
239 int attr_idx, enum ac_func_attr attr)
240 {
241 const char *attr_name = attr_to_str(attr);
242 unsigned kind_id = LLVMGetEnumAttributeKindForName(attr_name,
243 strlen(attr_name));
244 LLVMAttributeRef llvm_attr = LLVMCreateEnumAttribute(ctx, kind_id, 0);
245
246 if (LLVMIsAFunction(function))
247 LLVMAddAttributeAtIndex(function, attr_idx, llvm_attr);
248 else
249 LLVMAddCallSiteAttribute(function, attr_idx, llvm_attr);
250 }
251
252 void ac_add_func_attributes(LLVMContextRef ctx, LLVMValueRef function,
253 unsigned attrib_mask)
254 {
255 attrib_mask |= AC_FUNC_ATTR_NOUNWIND;
256 attrib_mask &= ~AC_FUNC_ATTR_LEGACY;
257
258 while (attrib_mask) {
259 enum ac_func_attr attr = 1u << u_bit_scan(&attrib_mask);
260 ac_add_function_attr(ctx, function, -1, attr);
261 }
262 }
263
264 void
265 ac_dump_module(LLVMModuleRef module)
266 {
267 char *str = LLVMPrintModuleToString(module);
268 fprintf(stderr, "%s", str);
269 LLVMDisposeMessage(str);
270 }
271
272 void
273 ac_llvm_add_target_dep_function_attr(LLVMValueRef F,
274 const char *name, unsigned value)
275 {
276 char str[16];
277
278 snprintf(str, sizeof(str), "0x%x", value);
279 LLVMAddTargetDependentFunctionAttr(F, name, str);
280 }
281
282 void ac_llvm_set_workgroup_size(LLVMValueRef F, unsigned size)
283 {
284 if (!size)
285 return;
286
287 char str[32];
288 snprintf(str, sizeof(str), "%u,%u", size, size);
289 LLVMAddTargetDependentFunctionAttr(F, "amdgpu-flat-work-group-size", str);
290 }
291
292 unsigned
293 ac_count_scratch_private_memory(LLVMValueRef function)
294 {
295 unsigned private_mem_vgprs = 0;
296
297 /* Process all LLVM instructions. */
298 LLVMBasicBlockRef bb = LLVMGetFirstBasicBlock(function);
299 while (bb) {
300 LLVMValueRef next = LLVMGetFirstInstruction(bb);
301
302 while (next) {
303 LLVMValueRef inst = next;
304 next = LLVMGetNextInstruction(next);
305
306 if (LLVMGetInstructionOpcode(inst) != LLVMAlloca)
307 continue;
308
309 LLVMTypeRef type = LLVMGetElementType(LLVMTypeOf(inst));
310 /* No idea why LLVM aligns allocas to 4 elements. */
311 unsigned alignment = LLVMGetAlignment(inst);
312 unsigned dw_size = align(ac_get_type_size(type) / 4, alignment);
313 private_mem_vgprs += dw_size;
314 }
315 bb = LLVMGetNextBasicBlock(bb);
316 }
317
318 return private_mem_vgprs;
319 }
320
321 bool
322 ac_init_llvm_compiler(struct ac_llvm_compiler *compiler,
323 enum radeon_family family,
324 enum ac_target_machine_options tm_options)
325 {
326 const char *triple;
327 memset(compiler, 0, sizeof(*compiler));
328
329 compiler->tm = ac_create_target_machine(family, tm_options,
330 LLVMCodeGenLevelDefault,
331 &triple);
332 if (!compiler->tm)
333 return false;
334
335 if (tm_options & AC_TM_CREATE_LOW_OPT) {
336 compiler->low_opt_tm =
337 ac_create_target_machine(family, tm_options,
338 LLVMCodeGenLevelLess, NULL);
339 if (!compiler->low_opt_tm)
340 goto fail;
341 }
342
343 if (family >= CHIP_NAVI10) {
344 assert(!(tm_options & AC_TM_CREATE_LOW_OPT));
345 compiler->tm_wave32 = ac_create_target_machine(family,
346 tm_options | AC_TM_WAVE32,
347 LLVMCodeGenLevelDefault,
348 NULL);
349 if (!compiler->tm_wave32)
350 goto fail;
351 }
352
353 compiler->target_library_info =
354 ac_create_target_library_info(triple);
355 if (!compiler->target_library_info)
356 goto fail;
357
358 compiler->passmgr = ac_create_passmgr(compiler->target_library_info,
359 tm_options & AC_TM_CHECK_IR);
360 if (!compiler->passmgr)
361 goto fail;
362
363 return true;
364 fail:
365 ac_destroy_llvm_compiler(compiler);
366 return false;
367 }
368
369 void
370 ac_destroy_llvm_compiler(struct ac_llvm_compiler *compiler)
371 {
372 ac_destroy_llvm_passes(compiler->passes);
373 ac_destroy_llvm_passes(compiler->passes_wave32);
374 ac_destroy_llvm_passes(compiler->low_opt_passes);
375
376 if (compiler->passmgr)
377 LLVMDisposePassManager(compiler->passmgr);
378 if (compiler->target_library_info)
379 ac_dispose_target_library_info(compiler->target_library_info);
380 if (compiler->low_opt_tm)
381 LLVMDisposeTargetMachine(compiler->low_opt_tm);
382 if (compiler->tm)
383 LLVMDisposeTargetMachine(compiler->tm);
384 if (compiler->tm_wave32)
385 LLVMDisposeTargetMachine(compiler->tm_wave32);
386 }