1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28
29 #include "pipe/p_config.h"
30 #include "pipe/p_compiler.h"
31 #include "util/u_cpu_detect.h"
32 #include "util/u_debug.h"
33 #include "util/u_memory.h"
34 #include "util/os_time.h"
35 #include "lp_bld.h"
36 #include "lp_bld_debug.h"
37 #include "lp_bld_misc.h"
38 #include "lp_bld_init.h"
39 #include "lp_bld_coro.h"
40 #include "lp_bld_printf.h"
41
42 #include <llvm/Config/llvm-config.h>
43 #include <llvm-c/Analysis.h>
44 #include <llvm-c/Transforms/Scalar.h>
45 #if LLVM_VERSION_MAJOR >= 7
46 #include <llvm-c/Transforms/Utils.h>
47 #endif
48 #include <llvm-c/BitWriter.h>
49 #if GALLIVM_USE_NEW_PASS == 1
50 #include <llvm-c/Transforms/PassBuilder.h>
51 #elif GALLIVM_HAVE_CORO == 1
52 #if LLVM_VERSION_MAJOR <= 8 && (defined(PIPE_ARCH_AARCH64) || defined (PIPE_ARCH_ARM) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_MIPS64))
53 #include <llvm-c/Transforms/IPO.h>
54 #endif
55 #include <llvm-c/Transforms/Coroutines.h>
56 #endif
57
58 unsigned gallivm_perf = 0;
59
60 static const struct debug_named_value lp_bld_perf_flags[] = {
61 { "brilinear", GALLIVM_PERF_BRILINEAR, "enable brilinear optimization" },
62 { "rho_approx", GALLIVM_PERF_RHO_APPROX, "enable rho_approx optimization" },
63 { "no_quad_lod", GALLIVM_PERF_NO_QUAD_LOD, "disable quad_lod optimization" },
64 { "no_aos_sampling", GALLIVM_PERF_NO_AOS_SAMPLING, "disable aos sampling optimization" },
65 { "nopt", GALLIVM_PERF_NO_OPT, "disable optimization passes to speed up shader compilation" },
66 DEBUG_NAMED_VALUE_END
67 };
68
69 #ifdef DEBUG
70 unsigned gallivm_debug = 0;
71
72 static const struct debug_named_value lp_bld_debug_flags[] = {
73 { "tgsi", GALLIVM_DEBUG_TGSI, NULL },
74 { "ir", GALLIVM_DEBUG_IR, NULL },
75 { "asm", GALLIVM_DEBUG_ASM, NULL },
76 { "perf", GALLIVM_DEBUG_PERF, NULL },
77 { "gc", GALLIVM_DEBUG_GC, NULL },
78 { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL },
79 DEBUG_NAMED_VALUE_END
80 };
81
82 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
83 #endif
84
85
86 static boolean gallivm_initialized = FALSE;
87
88 unsigned lp_native_vector_width;
89
90
91 /*
92 * Optimization values are:
93 * - 0: None (-O0)
94 * - 1: Less (-O1)
95 * - 2: Default (-O2, -Os)
96 * - 3: Aggressive (-O3)
97 *
98 * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
99 */
100 enum LLVM_CodeGenOpt_Level {
101 None, // -O0
102 Less, // -O1
103 Default, // -O2, -Os
104 Aggressive // -O3
105 };
106
107
108 /**
109 * Create the LLVM (optimization) pass manager and install
110 * relevant optimization passes.
111 * \return TRUE for success, FALSE for failure
112 */
113 static boolean
create_pass_manager(struct gallivm_state * gallivm)114 create_pass_manager(struct gallivm_state *gallivm)
115 {
116 #if GALLIVM_USE_NEW_PASS == 0
117 assert(!gallivm->passmgr);
118 assert(gallivm->target);
119
120 gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
121 if (!gallivm->passmgr)
122 return FALSE;
123
124 #if GALLIVM_HAVE_CORO == 1
125 gallivm->cgpassmgr = LLVMCreatePassManager();
126 #endif
127 /*
128 * TODO: some per module pass manager with IPO passes might be helpful -
129 * the generated texture functions may benefit from inlining if they are
130 * simple, or constant propagation into them, etc.
131 */
132
133 {
134 char *td_str;
135 // New ones from the Module.
136 td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
137 LLVMSetDataLayout(gallivm->module, td_str);
138 free(td_str);
139 }
140
141 #if GALLIVM_HAVE_CORO == 1
142 #if LLVM_VERSION_MAJOR <= 8 && (defined(PIPE_ARCH_AARCH64) || defined (PIPE_ARCH_ARM) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_MIPS64))
143 LLVMAddArgumentPromotionPass(gallivm->cgpassmgr);
144 LLVMAddFunctionAttrsPass(gallivm->cgpassmgr);
145 #endif
146 LLVMAddCoroEarlyPass(gallivm->cgpassmgr);
147 LLVMAddCoroSplitPass(gallivm->cgpassmgr);
148 LLVMAddCoroElidePass(gallivm->cgpassmgr);
149 #endif
150
151 if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
152 /*
153 * TODO: Evaluate passes some more - keeping in mind
154 * both quality of generated code and compile times.
155 */
156 /*
157 * NOTE: if you change this, don't forget to change the output
158 * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
159 */
160 LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
161 LLVMAddEarlyCSEPass(gallivm->passmgr);
162 LLVMAddCFGSimplificationPass(gallivm->passmgr);
163 /*
164 * FIXME: LICM is potentially quite useful. However, for some
165 * rather crazy shaders the compile time can reach _hours_ per shader,
166 * due to licm implying lcssa (since llvm 3.5), which can take forever.
167 * Even for sane shaders, the cost of licm is rather high (and not just
168 * due to lcssa, licm itself too), though mostly only in cases when it
169 * can actually move things, so having to disable it is a pity.
170 * LLVMAddLICMPass(gallivm->passmgr);
171 */
172 LLVMAddReassociatePass(gallivm->passmgr);
173 LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
174 #if LLVM_VERSION_MAJOR <= 11
175 LLVMAddConstantPropagationPass(gallivm->passmgr);
176 #else
177 LLVMAddInstructionSimplifyPass(gallivm->passmgr);
178 #endif
179 LLVMAddInstructionCombiningPass(gallivm->passmgr);
180 LLVMAddGVNPass(gallivm->passmgr);
181 }
182 else {
183 /* We need at least this pass to prevent the backends to fail in
184 * unexpected ways.
185 */
186 LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
187 }
188 #if GALLIVM_HAVE_CORO == 1
189 LLVMAddCoroCleanupPass(gallivm->passmgr);
190 #endif
191 #endif
192 return TRUE;
193 }
194
195 /**
196 * Free gallivm object's LLVM allocations, but not any generated code
197 * nor the gallivm object itself.
198 */
199 void
gallivm_free_ir(struct gallivm_state * gallivm)200 gallivm_free_ir(struct gallivm_state *gallivm)
201 {
202 #if GALLIVM_USE_NEW_PASS == 0
203 if (gallivm->passmgr) {
204 LLVMDisposePassManager(gallivm->passmgr);
205 }
206
207 #if GALLIVM_HAVE_CORO == 1
208 if (gallivm->cgpassmgr) {
209 LLVMDisposePassManager(gallivm->cgpassmgr);
210 }
211 #endif
212 #endif
213
214 if (gallivm->engine) {
215 /* This will already destroy any associated module */
216 LLVMDisposeExecutionEngine(gallivm->engine);
217 } else if (gallivm->module) {
218 LLVMDisposeModule(gallivm->module);
219 }
220
221 if (gallivm->cache) {
222 lp_free_objcache(gallivm->cache->jit_obj_cache);
223 free(gallivm->cache->data);
224 }
225 FREE(gallivm->module_name);
226
227 if (gallivm->target) {
228 LLVMDisposeTargetData(gallivm->target);
229 }
230
231 if (gallivm->builder)
232 LLVMDisposeBuilder(gallivm->builder);
233
234 /* The LLVMContext should be owned by the parent of gallivm. */
235
236 gallivm->engine = NULL;
237 gallivm->target = NULL;
238 gallivm->module = NULL;
239 gallivm->module_name = NULL;
240 #if GALLIVM_USE_NEW_PASS == 0
241 #if GALLIVM_HAVE_CORO == 1
242 gallivm->cgpassmgr = NULL;
243 #endif
244 gallivm->passmgr = NULL;
245 #endif
246 gallivm->context = NULL;
247 gallivm->builder = NULL;
248 gallivm->cache = NULL;
249 }
250
251
252 /**
253 * Free LLVM-generated code. Should be done AFTER gallivm_free_ir().
254 */
255 static void
gallivm_free_code(struct gallivm_state * gallivm)256 gallivm_free_code(struct gallivm_state *gallivm)
257 {
258 assert(!gallivm->module);
259 assert(!gallivm->engine);
260 lp_free_generated_code(gallivm->code);
261 gallivm->code = NULL;
262 lp_free_memory_manager(gallivm->memorymgr);
263 gallivm->memorymgr = NULL;
264 }
265
266
267 static boolean
init_gallivm_engine(struct gallivm_state * gallivm)268 init_gallivm_engine(struct gallivm_state *gallivm)
269 {
270 if (1) {
271 enum LLVM_CodeGenOpt_Level optlevel;
272 char *error = NULL;
273 int ret;
274
275 if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
276 optlevel = None;
277 }
278 else {
279 optlevel = Default;
280 }
281
282 ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
283 &gallivm->code,
284 gallivm->cache,
285 gallivm->module,
286 gallivm->memorymgr,
287 (unsigned) optlevel,
288 &error);
289 if (ret) {
290 _debug_printf("%s\n", error);
291 LLVMDisposeMessage(error);
292 goto fail;
293 }
294 }
295
296 if (0) {
297 /*
298 * Dump the data layout strings.
299 */
300
301 LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
302 char *data_layout;
303 char *engine_data_layout;
304
305 data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
306 engine_data_layout = LLVMCopyStringRepOfTargetData(target);
307
308 if (1) {
309 debug_printf("module target data = %s\n", data_layout);
310 debug_printf("engine target data = %s\n", engine_data_layout);
311 }
312
313 free(data_layout);
314 free(engine_data_layout);
315 }
316
317 return TRUE;
318
319 fail:
320 return FALSE;
321 }
322
323
324 /**
325 * Allocate gallivm LLVM objects.
326 * \return TRUE for success, FALSE for failure
327 */
328 static boolean
init_gallivm_state(struct gallivm_state * gallivm,const char * name,LLVMContextRef context,struct lp_cached_code * cache)329 init_gallivm_state(struct gallivm_state *gallivm, const char *name,
330 LLVMContextRef context, struct lp_cached_code *cache)
331 {
332 assert(!gallivm->context);
333 assert(!gallivm->module);
334
335 if (!lp_build_init())
336 return FALSE;
337
338 gallivm->context = context;
339 gallivm->cache = cache;
340 if (!gallivm->context)
341 goto fail;
342
343 gallivm->module_name = NULL;
344 if (name) {
345 size_t size = strlen(name) + 1;
346 gallivm->module_name = MALLOC(size);
347 if (gallivm->module_name) {
348 memcpy(gallivm->module_name, name, size);
349 }
350 }
351
352 gallivm->module = LLVMModuleCreateWithNameInContext(name,
353 gallivm->context);
354 if (!gallivm->module)
355 goto fail;
356
357 #if defined(PIPE_ARCH_X86)
358 lp_set_module_stack_alignment_override(gallivm->module, 4);
359 #endif
360
361 gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
362 if (!gallivm->builder)
363 goto fail;
364
365 gallivm->memorymgr = lp_get_default_memory_manager();
366 if (!gallivm->memorymgr)
367 goto fail;
368
369 /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
370 * complete when MC-JIT is created. So defer the MC-JIT engine creation for
371 * now.
372 */
373
374 /*
375 * MC-JIT engine compiles the module immediately on creation, so we can't
376 * obtain the target data from it. Instead we create a target data layout
377 * from a string.
378 *
379 * The produced layout strings are not precisely the same, but should make
380 * no difference for the kind of optimization passes we run.
381 *
382 * For reference this is the layout string on x64:
383 *
384 * e-p:64:64:64-S128-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f16:16:16-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-f128:128:128-n8:16:32:64
385 *
386 * See also:
387 * - http://llvm.org/docs/LangRef.html#datalayout
388 */
389
390 {
391 const unsigned pointer_size = 8 * sizeof(void *);
392 char layout[512];
393 snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
394 #if UTIL_ARCH_LITTLE_ENDIAN
395 'e', // little endian
396 #else
397 'E', // big endian
398 #endif
399 pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
400 pointer_size, // aggregate preferred alignment
401 pointer_size, pointer_size); // stack objects abi alignment, preferred alignment
402
403 gallivm->target = LLVMCreateTargetData(layout);
404 if (!gallivm->target) {
405 return FALSE;
406 }
407 }
408
409 if (!create_pass_manager(gallivm))
410 goto fail;
411
412 lp_build_coro_declare_malloc_hooks(gallivm);
413 return TRUE;
414
415 fail:
416 gallivm_free_ir(gallivm);
417 gallivm_free_code(gallivm);
418 return FALSE;
419 }
420
421
422 boolean
lp_build_init(void)423 lp_build_init(void)
424 {
425 if (gallivm_initialized)
426 return TRUE;
427
428
429 /* LLVMLinkIn* are no-ops at runtime. They just ensure the respective
430 * component is linked at buildtime, which is sufficient for its static
431 * constructors to be called at load time.
432 */
433 LLVMLinkInMCJIT();
434
435 #ifdef DEBUG
436 gallivm_debug = debug_get_option_gallivm_debug();
437 #endif
438
439 gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );
440
441 lp_set_target_options();
442
443 /* For simulating less capable machines */
444 #ifdef DEBUG
445 if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
446 extern struct util_cpu_caps_t util_cpu_caps;
447 assert(util_cpu_caps.has_sse2);
448 util_cpu_caps.has_sse3 = 0;
449 util_cpu_caps.has_ssse3 = 0;
450 util_cpu_caps.has_sse4_1 = 0;
451 util_cpu_caps.has_sse4_2 = 0;
452 util_cpu_caps.has_avx = 0;
453 util_cpu_caps.has_avx2 = 0;
454 util_cpu_caps.has_f16c = 0;
455 util_cpu_caps.has_fma = 0;
456 }
457 #endif
458
459 if (util_get_cpu_caps()->has_avx2 || util_get_cpu_caps()->has_avx) {
460 lp_native_vector_width = 256;
461 } else {
462 /* Leave it at 128, even when no SIMD extensions are available.
463 * Really needs to be a multiple of 128 so can fit 4 floats.
464 */
465 lp_native_vector_width = 128;
466 }
467
468 lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
469 lp_native_vector_width);
470
471 #if LLVM_VERSION_MAJOR < 4
472 if (lp_native_vector_width <= 128) {
473 /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
474 * "util_get_cpu_caps()->has_avx" predicate, and lack the
475 * "lp_native_vector_width > 128" predicate. And also to ensure a more
476 * consistent behavior, allowing one to test SSE2 on AVX machines.
477 * XXX: should not play games with util_cpu_caps directly as it might
478 * get used for other things outside llvm too.
479 */
480 util_get_cpu_caps()->has_avx = 0;
481 util_get_cpu_caps()->has_avx2 = 0;
482 util_get_cpu_caps()->has_f16c = 0;
483 util_get_cpu_caps()->has_fma = 0;
484 }
485 #endif
486
487 #ifdef PIPE_ARCH_PPC_64
488 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
489 * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
490 * that some rounding and half-float to float handling does not round
491 * incorrectly to 0.
492 * XXX: should eventually follow same logic on all platforms.
493 * Right now denorms get explicitly disabled (but elsewhere) for x86,
494 * whereas ppc64 explicitly enables them...
495 */
496 if (util_get_cpu_caps()->has_altivec) {
497 unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
498 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
499 __asm (
500 "mfvscr %%v1\n"
501 "vand %0,%%v1,%0\n"
502 "mtvscr %0"
503 :
504 : "r" (*mask)
505 );
506 }
507 #endif
508
509 gallivm_initialized = TRUE;
510
511 return TRUE;
512 }
513
514
515
516 /**
517 * Create a new gallivm_state object.
518 */
519 struct gallivm_state *
gallivm_create(const char * name,LLVMContextRef context,struct lp_cached_code * cache)520 gallivm_create(const char *name, LLVMContextRef context,
521 struct lp_cached_code *cache)
522 {
523 struct gallivm_state *gallivm;
524
525 gallivm = CALLOC_STRUCT(gallivm_state);
526 if (gallivm) {
527 if (!init_gallivm_state(gallivm, name, context, cache)) {
528 FREE(gallivm);
529 gallivm = NULL;
530 }
531 }
532
533 assert(gallivm != NULL);
534 return gallivm;
535 }
536
537
538 /**
539 * Destroy a gallivm_state object.
540 */
541 void
gallivm_destroy(struct gallivm_state * gallivm)542 gallivm_destroy(struct gallivm_state *gallivm)
543 {
544 gallivm_free_ir(gallivm);
545 gallivm_free_code(gallivm);
546 FREE(gallivm);
547 }
548
549
550 /**
551 * Validate a function.
552 * Verification is only done with debug builds.
553 */
554 void
gallivm_verify_function(struct gallivm_state * gallivm,LLVMValueRef func)555 gallivm_verify_function(struct gallivm_state *gallivm,
556 LLVMValueRef func)
557 {
558 /* Verify the LLVM IR. If invalid, dump and abort */
559 #ifdef DEBUG
560 if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
561 lp_debug_dump_value(func);
562 assert(0);
563 return;
564 }
565 #endif
566
567 if (gallivm_debug & GALLIVM_DEBUG_IR) {
568 /* Print the LLVM IR to stderr */
569 lp_debug_dump_value(func);
570 debug_printf("\n");
571 }
572 }
573
574
575 /**
576 * Compile a module.
577 * This does IR optimization on all functions in the module.
578 */
579 void
gallivm_compile_module(struct gallivm_state * gallivm)580 gallivm_compile_module(struct gallivm_state *gallivm)
581 {
582 int64_t time_begin = 0;
583
584 assert(!gallivm->compiled);
585
586 if (gallivm->builder) {
587 LLVMDisposeBuilder(gallivm->builder);
588 gallivm->builder = NULL;
589 }
590
591 LLVMSetDataLayout(gallivm->module, "");
592 assert(!gallivm->engine);
593 if (!init_gallivm_engine(gallivm)) {
594 assert(0);
595 }
596 assert(gallivm->engine);
597
598 if (gallivm->cache && gallivm->cache->data_size) {
599 goto skip_cached;
600 }
601
602 /* Dump bitcode to a file */
603 if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
604 char filename[256];
605 assert(gallivm->module_name);
606 snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
607 LLVMWriteBitcodeToFile(gallivm->module, filename);
608 debug_printf("%s written\n", filename);
609 debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
610 gallivm_perf & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
611 "-sroa -early-cse -simplifycfg -reassociate "
612 "-mem2reg -constprop -instcombine -gvn",
613 filename, gallivm_perf & GALLIVM_PERF_NO_OPT ? 0 : 2,
614 "[-mcpu=<-mcpu option>] ",
615 "[-mattr=<-mattr option(s)>]");
616 }
617
618 if (gallivm_debug & GALLIVM_DEBUG_PERF)
619 time_begin = os_time_get();
620
621 #if GALLIVM_USE_NEW_PASS == 1
622 char passes[1024];
623 passes[0] = 0;
624
625 /*
626 * there should be some way to combine these two pass runs but I'm not seeing it,
627 * at the time of writing.
628 */
629 strcpy(passes, "default<O0>");
630
631 LLVMPassBuilderOptionsRef opts = LLVMCreatePassBuilderOptions();
632 LLVMRunPasses(gallivm->module, passes, LLVMGetExecutionEngineTargetMachine(gallivm->engine), opts);
633
634 if (!(gallivm_perf & GALLIVM_PERF_NO_OPT))
635 strcpy(passes, "sroa,early-cse,simplifycfg,reassociate,mem2reg,instsimplify,instcombine");
636 else
637 strcpy(passes, "mem2reg");
638
639 LLVMRunPasses(gallivm->module, passes, LLVMGetExecutionEngineTargetMachine(gallivm->engine), opts);
640 LLVMDisposePassBuilderOptions(opts);
641 #else
642 #if GALLIVM_HAVE_CORO == 1
643 LLVMRunPassManager(gallivm->cgpassmgr, gallivm->module);
644 #endif
645 /* Run optimization passes */
646 LLVMInitializeFunctionPassManager(gallivm->passmgr);
647 LLVMValueRef func;
648 func = LLVMGetFirstFunction(gallivm->module);
649 while (func) {
650 if (0) {
651 debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
652 }
653
654 /* Disable frame pointer omission on debug/profile builds */
655 /* XXX: And workaround http://llvm.org/PR21435 */
656 #if defined(DEBUG) || defined(PROFILE) || defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
657 LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
658 LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
659 #endif
660
661 LLVMRunFunctionPassManager(gallivm->passmgr, func);
662 func = LLVMGetNextFunction(func);
663 }
664 LLVMFinalizeFunctionPassManager(gallivm->passmgr);
665 #endif
666 if (gallivm_debug & GALLIVM_DEBUG_PERF) {
667 int64_t time_end = os_time_get();
668 int time_msec = (int)((time_end - time_begin) / 1000);
669 assert(gallivm->module_name);
670 debug_printf("optimizing module %s took %d msec\n",
671 gallivm->module_name, time_msec);
672 }
673
674 /* Setting the module's DataLayout to an empty string will cause the
675 * ExecutionEngine to copy to the DataLayout string from its target machine
676 * to the module. As of LLVM 3.8 the module and the execution engine are
677 * required to have the same DataLayout.
678 *
679 * We must make sure we do this after running the optimization passes,
680 * because those passes need a correct datalayout string. For example, if
681 * those optimization passes see an empty datalayout, they will assume this
682 * is a little endian target and will do optimizations that break big endian
683 * machines.
684 *
685 * TODO: This is just a temporary work-around. The correct solution is for
686 * gallivm_init_state() to create a TargetMachine and pull the DataLayout
687 * from there. Currently, the TargetMachine used by llvmpipe is being
688 * implicitly created by the EngineBuilder in
689 * lp_build_create_jit_compiler_for_module()
690 */
691 skip_cached:
692
693 ++gallivm->compiled;
694
695 lp_init_printf_hook(gallivm);
696 LLVMAddGlobalMapping(gallivm->engine, gallivm->debug_printf_hook, debug_printf);
697
698
699 lp_build_coro_add_malloc_hooks(gallivm);
700
701 if (gallivm_debug & GALLIVM_DEBUG_ASM) {
702 LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
703
704 while (llvm_func) {
705 /*
706 * Need to filter out functions which don't have an implementation,
707 * such as the intrinsics. May not be sufficient in case of IPO?
708 * LLVMGetPointerToGlobal() will abort otherwise.
709 */
710 if (!LLVMIsDeclaration(llvm_func)) {
711 void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
712 lp_disassemble(llvm_func, func_code);
713 }
714 llvm_func = LLVMGetNextFunction(llvm_func);
715 }
716 }
717
718 #if defined(PROFILE)
719 {
720 LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
721
722 while (llvm_func) {
723 if (!LLVMIsDeclaration(llvm_func)) {
724 void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
725 lp_profile(llvm_func, func_code);
726 }
727 llvm_func = LLVMGetNextFunction(llvm_func);
728 }
729 }
730 #endif
731 }
732
733
734
735 func_pointer
gallivm_jit_function(struct gallivm_state * gallivm,LLVMValueRef func)736 gallivm_jit_function(struct gallivm_state *gallivm,
737 LLVMValueRef func)
738 {
739 void *code;
740 func_pointer jit_func;
741 int64_t time_begin = 0;
742
743 assert(gallivm->compiled);
744 assert(gallivm->engine);
745
746 if (gallivm_debug & GALLIVM_DEBUG_PERF)
747 time_begin = os_time_get();
748
749 code = LLVMGetPointerToGlobal(gallivm->engine, func);
750 assert(code);
751 jit_func = pointer_to_func(code);
752
753 if (gallivm_debug & GALLIVM_DEBUG_PERF) {
754 int64_t time_end = os_time_get();
755 int time_msec = (int)(time_end - time_begin) / 1000;
756 debug_printf(" jitting func %s took %d msec\n",
757 LLVMGetValueName(func), time_msec);
758 }
759
760 return jit_func;
761 }
762
gallivm_get_perf_flags(void)763 unsigned gallivm_get_perf_flags(void)
764 {
765 return gallivm_perf;
766 }
767