aboutsummaryrefslogtreecommitdiffstats
path: root/src/gallium/auxiliary/gallivm/lp_bld_init.c
blob: e8bca5bc3afce7210149cb5b100efd99ea1814c3 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
/**************************************************************************
 *
 * Copyright 2009 VMware, Inc.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the
 * "Software"), to deal in the Software without restriction, including
 * without limitation the rights to use, copy, modify, merge, publish,
 * distribute, sub license, and/or sell copies of the Software, and to
 * permit persons to whom the Software is furnished to do so, subject to
 * the following conditions:
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 **************************************************************************/


#include "pipe/p_config.h"
#include "pipe/p_compiler.h"
#include "util/u_cpu_detect.h"
#include "util/u_debug.h"
#include "util/u_memory.h"
#include "util/simple_list.h"
#include "util/os_time.h"
#include "lp_bld.h"
#include "lp_bld_debug.h"
#include "lp_bld_misc.h"
#include "lp_bld_init.h"

#include <llvm-c/Analysis.h>
#include <llvm-c/Transforms/Scalar.h>
#if HAVE_LLVM >= 0x0700
#include <llvm-c/Transforms/Utils.h>
#endif
#include <llvm-c/BitWriter.h>


/* Only MCJIT is available as of LLVM SVN r216982 */
#if HAVE_LLVM >= 0x0306
#  define USE_MCJIT 1
#elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
#  define USE_MCJIT 1
#endif

#if defined(USE_MCJIT)
static const bool use_mcjit = USE_MCJIT;
#else
static bool use_mcjit = FALSE;
#endif

unsigned gallivm_perf = 0;

static const struct debug_named_value lp_bld_perf_flags[] = {
   { "no_brilinear", GALLIVM_PERF_NO_BRILINEAR, "disable brilinear optimization" },
   { "no_rho_approx", GALLIVM_PERF_NO_RHO_APPROX, "disable rho_approx optimization" },
   { "no_quad_lod", GALLIVM_PERF_NO_QUAD_LOD, "disable quad_lod optimization" },
   { "nopt",   GALLIVM_PERF_NO_OPT, "disable optimization passes to speed up shader compilation" },
   { "no_filter_hacks", GALLIVM_PERF_NO_BRILINEAR | GALLIVM_PERF_NO_RHO_APPROX |
     GALLIVM_PERF_NO_QUAD_LOD, "disable filter optimization hacks" },
   DEBUG_NAMED_VALUE_END
};

#ifdef DEBUG
unsigned gallivm_debug = 0;

static const struct debug_named_value lp_bld_debug_flags[] = {
   { "tgsi",   GALLIVM_DEBUG_TGSI, NULL },
   { "ir",     GALLIVM_DEBUG_IR, NULL },
   { "asm",    GALLIVM_DEBUG_ASM, NULL },
   { "perf",   GALLIVM_DEBUG_PERF, NULL },
   { "gc",     GALLIVM_DEBUG_GC, NULL },
   { "dumpbc", GALLIVM_DEBUG_DUMP_BC, NULL },
   DEBUG_NAMED_VALUE_END
};

DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0)
#endif


static boolean gallivm_initialized = FALSE;

unsigned lp_native_vector_width;


/*
 * Optimization values are:
 * - 0: None (-O0)
 * - 1: Less (-O1)
 * - 2: Default (-O2, -Os)
 * - 3: Aggressive (-O3)
 *
 * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
 */
enum LLVM_CodeGenOpt_Level {
   None,        // -O0
   Less,        // -O1
   Default,     // -O2, -Os
   Aggressive   // -O3
};


/**
 * Create the LLVM (optimization) pass manager and install
 * relevant optimization passes.
 * \return  TRUE for success, FALSE for failure
 */
static boolean
create_pass_manager(struct gallivm_state *gallivm)
{
   assert(!gallivm->passmgr);
   assert(gallivm->target);

   gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
   if (!gallivm->passmgr)
      return FALSE;
   /*
    * TODO: some per module pass manager with IPO passes might be helpful -
    * the generated texture functions may benefit from inlining if they are
    * simple, or constant propagation into them, etc.
    */

#if HAVE_LLVM < 0x0309
   // Old versions of LLVM get the DataLayout from the pass manager.
   LLVMAddTargetData(gallivm->target, gallivm->passmgr);
#endif

   {
      char *td_str;
      // New ones from the Module.
      td_str = LLVMCopyStringRepOfTargetData(gallivm->target);
      LLVMSetDataLayout(gallivm->module, td_str);
      free(td_str);
   }

   if ((gallivm_perf & GALLIVM_PERF_NO_OPT) == 0) {
      /*
       * TODO: Evaluate passes some more - keeping in mind
       * both quality of generated code and compile times.
       */
      /*
       * NOTE: if you change this, don't forget to change the output
       * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
       */
      LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
      LLVMAddEarlyCSEPass(gallivm->passmgr);
      LLVMAddCFGSimplificationPass(gallivm->passmgr);
      /*
       * FIXME: LICM is potentially quite useful. However, for some
       * rather crazy shaders the compile time can reach _hours_ per shader,
       * due to licm implying lcssa (since llvm 3.5), which can take forever.
       * Even for sane shaders, the cost of licm is rather high (and not just
       * due to lcssa, licm itself too), though mostly only in cases when it
       * can actually move things, so having to disable it is a pity.
       * LLVMAddLICMPass(gallivm->passmgr);
       */
      LLVMAddReassociatePass(gallivm->passmgr);
      LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
      LLVMAddConstantPropagationPass(gallivm->passmgr);
      LLVMAddInstructionCombiningPass(gallivm->passmgr);
      LLVMAddGVNPass(gallivm->passmgr);
   }
   else {
      /* We need at least this pass to prevent the backends to fail in
       * unexpected ways.
       */
      LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
   }

   return TRUE;
}


/**
 * Free gallivm object's LLVM allocations, but not any generated code
 * nor the gallivm object itself.
 */
void
gallivm_free_ir(struct gallivm_state *gallivm)
{
   if (gallivm->passmgr) {
      LLVMDisposePassManager(gallivm->passmgr);
   }

   if (gallivm->engine) {
      /* This will already destroy any associated module */
      LLVMDisposeExecutionEngine(gallivm->engine);
   } else if (gallivm->module) {
      LLVMDisposeModule(gallivm->module);
   }

   FREE(gallivm->module_name);

   if (!use_mcjit) {
      /* Don't free the TargetData, it's owned by the exec engine */
   } else {
      if (gallivm->target) {
         LLVMDisposeTargetData(gallivm->target);
      }
   }

   if (gallivm->builder)
      LLVMDisposeBuilder(gallivm->builder);

   /* The LLVMContext should be owned by the parent of gallivm. */

   gallivm->engine = NULL;
   gallivm->target = NULL;
   gallivm->module = NULL;
   gallivm->module_name = NULL;
   gallivm->passmgr = NULL;
   gallivm->context = NULL;
   gallivm->builder = NULL;
}


/**
 * Free LLVM-generated code.  Should be done AFTER gallivm_free_ir().
 */
static void
gallivm_free_code(struct gallivm_state *gallivm)
{
   assert(!gallivm->module);
   assert(!gallivm->engine);
   lp_free_generated_code(gallivm->code);
   gallivm->code = NULL;
   lp_free_memory_manager(gallivm->memorymgr);
   gallivm->memorymgr = NULL;
}


static boolean
init_gallivm_engine(struct gallivm_state *gallivm)
{
   if (1) {
      enum LLVM_CodeGenOpt_Level optlevel;
      char *error = NULL;
      int ret;

      if (gallivm_perf & GALLIVM_PERF_NO_OPT) {
         optlevel = None;
      }
      else {
         optlevel = Default;
      }

      ret = lp_build_create_jit_compiler_for_module(&gallivm->engine,
                                                    &gallivm->code,
                                                    gallivm->module,
                                                    gallivm->memorymgr,
                                                    (unsigned) optlevel,
                                                    use_mcjit,
                                                    &error);
      if (ret) {
         _debug_printf("%s\n", error);
         LLVMDisposeMessage(error);
         goto fail;
      }
   }

   if (!use_mcjit) {
      gallivm->target = LLVMGetExecutionEngineTargetData(gallivm->engine);
      if (!gallivm->target)
         goto fail;
   } else {
      if (0) {
          /*
           * Dump the data layout strings.
           */

          LLVMTargetDataRef target = LLVMGetExecutionEngineTargetData(gallivm->engine);
          char *data_layout;
          char *engine_data_layout;

          data_layout = LLVMCopyStringRepOfTargetData(gallivm->target);
          engine_data_layout = LLVMCopyStringRepOfTargetData(target);

          if (1) {
             debug_printf("module target data = %s\n", data_layout);
             debug_printf("engine target data = %s\n", engine_data_layout);
          }

          free(data_layout);
          free(engine_data_layout);
      }
   }

   return TRUE;

fail:
   return FALSE;
}


/**
 * Allocate gallivm LLVM objects.
 * \return  TRUE for success, FALSE for failure
 */
static boolean
init_gallivm_state(struct gallivm_state *gallivm, const char *name,
                   LLVMContextRef context)
{
   assert(!gallivm->context);
   assert(!gallivm->module);

   if (!lp_build_init())
      return FALSE;

   gallivm->context = context;

   if (!gallivm->context)
      goto fail;

   gallivm->module_name = NULL;
   if (name) {
      size_t size = strlen(name) + 1;
      gallivm->module_name = MALLOC(size);
      if (gallivm->module_name) {
         memcpy(gallivm->module_name, name, size);
      }
   }

   gallivm->module = LLVMModuleCreateWithNameInContext(name,
                                                       gallivm->context);
   if (!gallivm->module)
      goto fail;

   gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
   if (!gallivm->builder)
      goto fail;

   gallivm->memorymgr = lp_get_default_memory_manager();
   if (!gallivm->memorymgr)
      goto fail;

   /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
    * complete when MC-JIT is created. So defer the MC-JIT engine creation for
    * now.
    */
   if (!use_mcjit) {
      if (!init_gallivm_engine(gallivm)) {
         goto fail;
      }
   } else {
      /*
       * MC-JIT engine compiles the module immediately on creation, so we can't
       * obtain the target data from it.  Instead we create a target data layout
       * from a string.
       *
       * The produced layout strings are not precisely the same, but should make
       * no difference for the kind of optimization passes we run.
       *
       * For reference this is the layout string on x64:
       *
       *   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
       *
       * See also:
       * - http://llvm.org/docs/LangRef.html#datalayout
       */

      {
         const unsigned pointer_size = 8 * sizeof(void *);
         char layout[512];
         util_snprintf(layout, sizeof layout, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
#ifdef PIPE_ARCH_LITTLE_ENDIAN
                       'e', // little endian
#else
                       'E', // big endian
#endif
                       pointer_size, pointer_size, pointer_size, // pointer size, abi alignment, preferred alignment
                       pointer_size, // aggregate preferred alignment
                       pointer_size, pointer_size); // stack objects abi alignment, preferred alignment

         gallivm->target = LLVMCreateTargetData(layout);
         if (!gallivm->target) {
            return FALSE;
         }
      }
   }

   if (!create_pass_manager(gallivm))
      goto fail;

   return TRUE;

fail:
   gallivm_free_ir(gallivm);
   gallivm_free_code(gallivm);
   return FALSE;
}


boolean
lp_build_init(void)
{
   if (gallivm_initialized)
      return TRUE;


   /* LLVMLinkIn* are no-ops at runtime.  They just ensure the respective
    * component is linked at buildtime, which is sufficient for its static
    * constructors to be called at load time.
    */
#if defined(USE_MCJIT)
#  if USE_MCJIT
      LLVMLinkInMCJIT();
#  else
      LLVMLinkInJIT();
#  endif
#else
   use_mcjit = debug_get_bool_option("GALLIVM_MCJIT", FALSE);
   LLVMLinkInJIT();
   LLVMLinkInMCJIT();
#endif

#ifdef DEBUG
   gallivm_debug = debug_get_option_gallivm_debug();
#endif

   gallivm_perf = debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags, 0 );

   lp_set_target_options();

   util_cpu_detect();

   /* For simulating less capable machines */
#ifdef DEBUG
   if (debug_get_bool_option("LP_FORCE_SSE2", FALSE)) {
      assert(util_cpu_caps.has_sse2);
      util_cpu_caps.has_sse3 = 0;
      util_cpu_caps.has_ssse3 = 0;
      util_cpu_caps.has_sse4_1 = 0;
      util_cpu_caps.has_sse4_2 = 0;
      util_cpu_caps.has_avx = 0;
      util_cpu_caps.has_avx2 = 0;
      util_cpu_caps.has_f16c = 0;
      util_cpu_caps.has_fma = 0;
   }
#endif

   /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
    * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
    * actually more efficient to use 4-wide vectors on this processor.
    *
    * See also:
    * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
    */
   if (util_cpu_caps.has_avx &&
       util_cpu_caps.has_intel) {
      lp_native_vector_width = 256;
   } else {
      /* Leave it at 128, even when no SIMD extensions are available.
       * Really needs to be a multiple of 128 so can fit 4 floats.
       */
      lp_native_vector_width = 128;
   }
 
   lp_native_vector_width = debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
                                                 lp_native_vector_width);

   if (lp_native_vector_width <= 128) {
      /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
       * "util_cpu_caps.has_avx" predicate, and lack the
       * "lp_native_vector_width > 128" predicate. And also to ensure a more
       * consistent behavior, allowing one to test SSE2 on AVX machines.
       * XXX: should not play games with util_cpu_caps directly as it might
       * get used for other things outside llvm too.
       */
      util_cpu_caps.has_avx = 0;
      util_cpu_caps.has_avx2 = 0;
      util_cpu_caps.has_f16c = 0;
      util_cpu_caps.has_fma = 0;
   }
   if (HAVE_LLVM < 0x0304 || !use_mcjit) {
      /* AVX2 support has only been tested with LLVM 3.4, and it requires
       * MCJIT. */
      util_cpu_caps.has_avx2 = 0;
   }

#ifdef PIPE_ARCH_PPC_64
   /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
    * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
    * that some rounding and half-float to float handling does not round
    * incorrectly to 0.
    * XXX: should eventually follow same logic on all platforms.
    * Right now denorms get explicitly disabled (but elsewhere) for x86,
    * whereas ppc64 explicitly enables them...
    */
   if (util_cpu_caps.has_altivec) {
      unsigned short mask[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
                                0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
      __asm (
        "mfvscr %%v1\n"
        "vand   %0,%%v1,%0\n"
        "mtvscr %0"
        :
        : "r" (*mask)
      );
   }
#endif

   gallivm_initialized = TRUE;

   return TRUE;
}



/**
 * Create a new gallivm_state object.
 */
struct gallivm_state *
gallivm_create(const char *name, LLVMContextRef context)
{
   struct gallivm_state *gallivm;

   gallivm = CALLOC_STRUCT(gallivm_state);
   if (gallivm) {
      if (!init_gallivm_state(gallivm, name, context)) {
         FREE(gallivm);
         gallivm = NULL;
      }
   }

   return gallivm;
}


/**
 * Destroy a gallivm_state object.
 */
void
gallivm_destroy(struct gallivm_state *gallivm)
{
   gallivm_free_ir(gallivm);
   gallivm_free_code(gallivm);
   FREE(gallivm);
}


/**
 * Validate a function.
 * Verification is only done with debug builds.
 */
void
gallivm_verify_function(struct gallivm_state *gallivm,
                        LLVMValueRef func)
{
   /* Verify the LLVM IR.  If invalid, dump and abort */
#ifdef DEBUG
   if (LLVMVerifyFunction(func, LLVMPrintMessageAction)) {
      lp_debug_dump_value(func);
      assert(0);
      return;
   }
#endif

   if (gallivm_debug & GALLIVM_DEBUG_IR) {
      /* Print the LLVM IR to stderr */
      lp_debug_dump_value(func);
      debug_printf("\n");
   }
}


/**
 * Compile a module.
 * This does IR optimization on all functions in the module.
 */
void
gallivm_compile_module(struct gallivm_state *gallivm)
{
   LLVMValueRef func;
   int64_t time_begin = 0;

   assert(!gallivm->compiled);

   if (gallivm->builder) {
      LLVMDisposeBuilder(gallivm->builder);
      gallivm->builder = NULL;
   }

   /* Dump bitcode to a file */
   if (gallivm_debug & GALLIVM_DEBUG_DUMP_BC) {
      char filename[256];
      assert(gallivm->module_name);
      util_snprintf(filename, sizeof(filename), "ir_%s.bc", gallivm->module_name);
      LLVMWriteBitcodeToFile(gallivm->module, filename);
      debug_printf("%s written\n", filename);
      debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
                   gallivm_debug & GALLIVM_PERF_NO_OPT ? "-mem2reg" :
                   "-sroa -early-cse -simplifycfg -reassociate "
                   "-mem2reg -constprop -instcombine -gvn",
                   filename, gallivm_debug & GALLIVM_PERF_NO_OPT ? 0 : 2,
                   (HAVE_LLVM >= 0x0305) ? "[-mcpu=<-mcpu option>] " : "",
                   "[-mattr=<-mattr option(s)>]");
   }

   if (gallivm_debug & GALLIVM_DEBUG_PERF)
      time_begin = os_time_get();

   /* Run optimization passes */
   LLVMInitializeFunctionPassManager(gallivm->passmgr);
   func = LLVMGetFirstFunction(gallivm->module);
   while (func) {
      if (0) {
         debug_printf("optimizing func %s...\n", LLVMGetValueName(func));
      }

   /* Disable frame pointer omission on debug/profile builds */
   /* XXX: And workaround http://llvm.org/PR21435 */
#if HAVE_LLVM >= 0x0307 && \
    (defined(DEBUG) || defined(PROFILE) || \
     defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
      LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim", "true");
      LLVMAddTargetDependentFunctionAttr(func, "no-frame-pointer-elim-non-leaf", "true");
#endif

      LLVMRunFunctionPassManager(gallivm->passmgr, func);
      func = LLVMGetNextFunction(func);
   }
   LLVMFinalizeFunctionPassManager(gallivm->passmgr);

   if (gallivm_debug & GALLIVM_DEBUG_PERF) {
      int64_t time_end = os_time_get();
      int time_msec = (int)((time_end - time_begin) / 1000);
      assert(gallivm->module_name);
      debug_printf("optimizing module %s took %d msec\n",
                   gallivm->module_name, time_msec);
   }

   if (use_mcjit) {
      /* Setting the module's DataLayout to an empty string will cause the
       * ExecutionEngine to copy to the DataLayout string from its target
       * machine to the module.  As of LLVM 3.8 the module and the execution
       * engine are required to have the same DataLayout.
       *
       * We must make sure we do this after running the optimization passes,
       * because those passes need a correct datalayout string.  For example,
       * if those optimization passes see an empty datalayout, they will assume
       * this is a little endian target and will do optimizations that break big
       * endian machines.
       *
       * TODO: This is just a temporary work-around.  The correct solution is
       * for gallivm_init_state() to create a TargetMachine and pull the
       * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
       * is being implicitly created by the EngineBuilder in
       * lp_build_create_jit_compiler_for_module()
       */
      LLVMSetDataLayout(gallivm->module, "");
      assert(!gallivm->engine);
      if (!init_gallivm_engine(gallivm)) {
         assert(0);
      }
   }
   assert(gallivm->engine);

   ++gallivm->compiled;

   if (gallivm_debug & GALLIVM_DEBUG_ASM) {
      LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);

      while (llvm_func) {
         /*
          * Need to filter out functions which don't have an implementation,
          * such as the intrinsics. May not be sufficient in case of IPO?
          * LLVMGetPointerToGlobal() will abort otherwise.
          */
         if (!LLVMIsDeclaration(llvm_func)) {
            void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
            lp_disassemble(llvm_func, func_code);
         }
         llvm_func = LLVMGetNextFunction(llvm_func);
      }
   }

#if defined(PROFILE)
   {
      LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);

      while (llvm_func) {
         if (!LLVMIsDeclaration(llvm_func)) {
            void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
            lp_profile(llvm_func, func_code);
         }
         llvm_func = LLVMGetNextFunction(llvm_func);
      }
   }
#endif
}



func_pointer
gallivm_jit_function(struct gallivm_state *gallivm,
                     LLVMValueRef func)
{
   void *code;
   func_pointer jit_func;
   int64_t time_begin = 0;

   assert(gallivm->compiled);
   assert(gallivm->engine);

   if (gallivm_debug & GALLIVM_DEBUG_PERF)
      time_begin = os_time_get();

   code = LLVMGetPointerToGlobal(gallivm->engine, func);
   assert(code);
   jit_func = pointer_to_func(code);

   if (gallivm_debug & GALLIVM_DEBUG_PERF) {
      int64_t time_end = os_time_get();
      int time_msec = (int)(time_end - time_begin) / 1000;
      debug_printf("   jitting func %s took %d msec\n",
                   LLVMGetValueName(func), time_msec);
   }

   return jit_func;
}