1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
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:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
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.
26 **************************************************************************/
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/simple_list.h"
35 #include "util/os_time.h"
37 #include "lp_bld_debug.h"
38 #include "lp_bld_misc.h"
39 #include "lp_bld_init.h"
41 #include <llvm/Config/llvm-config.h>
42 #include <llvm-c/Analysis.h>
43 #include <llvm-c/Transforms/Scalar.h>
44 #if LLVM_VERSION_MAJOR >= 7
45 #include <llvm-c/Transforms/Utils.h>
47 #include <llvm-c/BitWriter.h>
49 #include <llvm-c/Transforms/Coroutines.h>
52 /* Only MCJIT is available as of LLVM SVN r216982 */
53 #if LLVM_VERSION_MAJOR > 3 || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 6)
55 #elif defined(PIPE_ARCH_PPC_64) || defined(PIPE_ARCH_S390) || defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
59 #if defined(USE_MCJIT)
60 static const bool use_mcjit
= USE_MCJIT
;
62 static bool use_mcjit
= FALSE
;
65 unsigned gallivm_perf
= 0;
67 static const struct debug_named_value lp_bld_perf_flags
[] = {
68 { "no_brilinear", GALLIVM_PERF_NO_BRILINEAR
, "disable brilinear optimization" },
69 { "no_rho_approx", GALLIVM_PERF_NO_RHO_APPROX
, "disable rho_approx optimization" },
70 { "no_quad_lod", GALLIVM_PERF_NO_QUAD_LOD
, "disable quad_lod optimization" },
71 { "no_aos_sampling", GALLIVM_PERF_NO_AOS_SAMPLING
, "disable aos sampling optimization" },
72 { "nopt", GALLIVM_PERF_NO_OPT
, "disable optimization passes to speed up shader compilation" },
73 { "no_filter_hacks", GALLIVM_PERF_NO_BRILINEAR
| GALLIVM_PERF_NO_RHO_APPROX
|
74 GALLIVM_PERF_NO_QUAD_LOD
, "disable filter optimization hacks" },
79 unsigned gallivm_debug
= 0;
81 static const struct debug_named_value lp_bld_debug_flags
[] = {
82 { "tgsi", GALLIVM_DEBUG_TGSI
, NULL
},
83 { "ir", GALLIVM_DEBUG_IR
, NULL
},
84 { "asm", GALLIVM_DEBUG_ASM
, NULL
},
85 { "perf", GALLIVM_DEBUG_PERF
, NULL
},
86 { "gc", GALLIVM_DEBUG_GC
, NULL
},
87 { "dumpbc", GALLIVM_DEBUG_DUMP_BC
, NULL
},
91 DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug
, "GALLIVM_DEBUG", lp_bld_debug_flags
, 0)
95 static boolean gallivm_initialized
= FALSE
;
97 unsigned lp_native_vector_width
;
101 * Optimization values are:
104 * - 2: Default (-O2, -Os)
105 * - 3: Aggressive (-O3)
107 * See also CodeGenOpt::Level in llvm/Target/TargetMachine.h
109 enum LLVM_CodeGenOpt_Level
{
118 * Create the LLVM (optimization) pass manager and install
119 * relevant optimization passes.
120 * \return TRUE for success, FALSE for failure
123 create_pass_manager(struct gallivm_state
*gallivm
)
125 assert(!gallivm
->passmgr
);
126 assert(gallivm
->target
);
128 gallivm
->passmgr
= LLVMCreateFunctionPassManagerForModule(gallivm
->module
);
129 if (!gallivm
->passmgr
)
132 #if GALLIVM_HAVE_CORO
133 gallivm
->cgpassmgr
= LLVMCreatePassManager();
136 * TODO: some per module pass manager with IPO passes might be helpful -
137 * the generated texture functions may benefit from inlining if they are
138 * simple, or constant propagation into them, etc.
141 #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR < 9
142 // Old versions of LLVM get the DataLayout from the pass manager.
143 LLVMAddTargetData(gallivm
->target
, gallivm
->passmgr
);
148 // New ones from the Module.
149 td_str
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
150 LLVMSetDataLayout(gallivm
->module
, td_str
);
154 #if GALLIVM_HAVE_CORO
155 LLVMAddCoroEarlyPass(gallivm
->cgpassmgr
);
156 LLVMAddCoroSplitPass(gallivm
->cgpassmgr
);
157 LLVMAddCoroElidePass(gallivm
->cgpassmgr
);
160 if ((gallivm_perf
& GALLIVM_PERF_NO_OPT
) == 0) {
162 * TODO: Evaluate passes some more - keeping in mind
163 * both quality of generated code and compile times.
166 * NOTE: if you change this, don't forget to change the output
167 * with GALLIVM_DEBUG_DUMP_BC in gallivm_compile_module.
169 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
170 LLVMAddEarlyCSEPass(gallivm
->passmgr
);
171 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
173 * FIXME: LICM is potentially quite useful. However, for some
174 * rather crazy shaders the compile time can reach _hours_ per shader,
175 * due to licm implying lcssa (since llvm 3.5), which can take forever.
176 * Even for sane shaders, the cost of licm is rather high (and not just
177 * due to lcssa, licm itself too), though mostly only in cases when it
178 * can actually move things, so having to disable it is a pity.
179 * LLVMAddLICMPass(gallivm->passmgr);
181 LLVMAddReassociatePass(gallivm
->passmgr
);
182 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
183 LLVMAddConstantPropagationPass(gallivm
->passmgr
);
184 LLVMAddInstructionCombiningPass(gallivm
->passmgr
);
185 LLVMAddGVNPass(gallivm
->passmgr
);
186 #if GALLIVM_HAVE_CORO
187 LLVMAddCoroCleanupPass(gallivm
->passmgr
);
191 /* We need at least this pass to prevent the backends to fail in
194 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
202 * Free gallivm object's LLVM allocations, but not any generated code
203 * nor the gallivm object itself.
206 gallivm_free_ir(struct gallivm_state
*gallivm
)
208 if (gallivm
->passmgr
) {
209 LLVMDisposePassManager(gallivm
->passmgr
);
212 #if GALLIVM_HAVE_CORO
213 if (gallivm
->cgpassmgr
) {
214 LLVMDisposePassManager(gallivm
->cgpassmgr
);
218 if (gallivm
->engine
) {
219 /* This will already destroy any associated module */
220 LLVMDisposeExecutionEngine(gallivm
->engine
);
221 } else if (gallivm
->module
) {
222 LLVMDisposeModule(gallivm
->module
);
225 FREE(gallivm
->module_name
);
228 /* Don't free the TargetData, it's owned by the exec engine */
230 if (gallivm
->target
) {
231 LLVMDisposeTargetData(gallivm
->target
);
235 if (gallivm
->builder
)
236 LLVMDisposeBuilder(gallivm
->builder
);
238 /* The LLVMContext should be owned by the parent of gallivm. */
240 gallivm
->engine
= NULL
;
241 gallivm
->target
= NULL
;
242 gallivm
->module
= NULL
;
243 gallivm
->module_name
= NULL
;
244 gallivm
->cgpassmgr
= NULL
;
245 gallivm
->passmgr
= NULL
;
246 gallivm
->context
= NULL
;
247 gallivm
->builder
= NULL
;
252 * Free LLVM-generated code. Should be done AFTER gallivm_free_ir().
255 gallivm_free_code(struct gallivm_state
*gallivm
)
257 assert(!gallivm
->module
);
258 assert(!gallivm
->engine
);
259 lp_free_generated_code(gallivm
->code
);
260 gallivm
->code
= NULL
;
261 lp_free_memory_manager(gallivm
->memorymgr
);
262 gallivm
->memorymgr
= NULL
;
267 init_gallivm_engine(struct gallivm_state
*gallivm
)
270 enum LLVM_CodeGenOpt_Level optlevel
;
274 if (gallivm_perf
& GALLIVM_PERF_NO_OPT
) {
281 ret
= lp_build_create_jit_compiler_for_module(&gallivm
->engine
,
289 _debug_printf("%s\n", error
);
290 LLVMDisposeMessage(error
);
296 gallivm
->target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
297 if (!gallivm
->target
)
302 * Dump the data layout strings.
305 LLVMTargetDataRef target
= LLVMGetExecutionEngineTargetData(gallivm
->engine
);
307 char *engine_data_layout
;
309 data_layout
= LLVMCopyStringRepOfTargetData(gallivm
->target
);
310 engine_data_layout
= LLVMCopyStringRepOfTargetData(target
);
313 debug_printf("module target data = %s\n", data_layout
);
314 debug_printf("engine target data = %s\n", engine_data_layout
);
318 free(engine_data_layout
);
330 * Allocate gallivm LLVM objects.
331 * \return TRUE for success, FALSE for failure
334 init_gallivm_state(struct gallivm_state
*gallivm
, const char *name
,
335 LLVMContextRef context
)
337 assert(!gallivm
->context
);
338 assert(!gallivm
->module
);
340 if (!lp_build_init())
343 gallivm
->context
= context
;
345 if (!gallivm
->context
)
348 gallivm
->module_name
= NULL
;
350 size_t size
= strlen(name
) + 1;
351 gallivm
->module_name
= MALLOC(size
);
352 if (gallivm
->module_name
) {
353 memcpy(gallivm
->module_name
, name
, size
);
357 gallivm
->module
= LLVMModuleCreateWithNameInContext(name
,
359 if (!gallivm
->module
)
362 gallivm
->builder
= LLVMCreateBuilderInContext(gallivm
->context
);
363 if (!gallivm
->builder
)
366 gallivm
->memorymgr
= lp_get_default_memory_manager();
367 if (!gallivm
->memorymgr
)
370 /* FIXME: MC-JIT only allows compiling one module at a time, and it must be
371 * complete when MC-JIT is created. So defer the MC-JIT engine creation for
375 if (!init_gallivm_engine(gallivm
)) {
380 * MC-JIT engine compiles the module immediately on creation, so we can't
381 * obtain the target data from it. Instead we create a target data layout
384 * The produced layout strings are not precisely the same, but should make
385 * no difference for the kind of optimization passes we run.
387 * For reference this is the layout string on x64:
389 * 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
392 * - http://llvm.org/docs/LangRef.html#datalayout
396 const unsigned pointer_size
= 8 * sizeof(void *);
398 snprintf(layout
, sizeof layout
, "%c-p:%u:%u:%u-i64:64:64-a0:0:%u-s0:%u:%u",
399 #ifdef PIPE_ARCH_LITTLE_ENDIAN
400 'e', // little endian
404 pointer_size
, pointer_size
, pointer_size
, // pointer size, abi alignment, preferred alignment
405 pointer_size
, // aggregate preferred alignment
406 pointer_size
, pointer_size
); // stack objects abi alignment, preferred alignment
408 gallivm
->target
= LLVMCreateTargetData(layout
);
409 if (!gallivm
->target
) {
415 if (!create_pass_manager(gallivm
))
421 gallivm_free_ir(gallivm
);
422 gallivm_free_code(gallivm
);
430 if (gallivm_initialized
)
434 /* LLVMLinkIn* are no-ops at runtime. They just ensure the respective
435 * component is linked at buildtime, which is sufficient for its static
436 * constructors to be called at load time.
438 #if defined(USE_MCJIT)
445 use_mcjit
= debug_get_bool_option("GALLIVM_MCJIT", FALSE
);
451 gallivm_debug
= debug_get_option_gallivm_debug();
454 gallivm_perf
= debug_get_flags_option("GALLIVM_PERF", lp_bld_perf_flags
, 0 );
456 lp_set_target_options();
460 /* For simulating less capable machines */
462 if (debug_get_bool_option("LP_FORCE_SSE2", FALSE
)) {
463 assert(util_cpu_caps
.has_sse2
);
464 util_cpu_caps
.has_sse3
= 0;
465 util_cpu_caps
.has_ssse3
= 0;
466 util_cpu_caps
.has_sse4_1
= 0;
467 util_cpu_caps
.has_sse4_2
= 0;
468 util_cpu_caps
.has_avx
= 0;
469 util_cpu_caps
.has_avx2
= 0;
470 util_cpu_caps
.has_f16c
= 0;
471 util_cpu_caps
.has_fma
= 0;
475 /* AMD Bulldozer AVX's throughput is the same as SSE2; and because using
476 * 8-wide vector needs more floating ops than 4-wide (due to padding), it is
477 * actually more efficient to use 4-wide vectors on this processor.
480 * - http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
482 if (util_cpu_caps
.has_avx
&&
483 util_cpu_caps
.has_intel
) {
484 lp_native_vector_width
= 256;
486 /* Leave it at 128, even when no SIMD extensions are available.
487 * Really needs to be a multiple of 128 so can fit 4 floats.
489 lp_native_vector_width
= 128;
492 lp_native_vector_width
= debug_get_num_option("LP_NATIVE_VECTOR_WIDTH",
493 lp_native_vector_width
);
495 if (lp_native_vector_width
<= 128) {
496 /* Hide AVX support, as often LLVM AVX intrinsics are only guarded by
497 * "util_cpu_caps.has_avx" predicate, and lack the
498 * "lp_native_vector_width > 128" predicate. And also to ensure a more
499 * consistent behavior, allowing one to test SSE2 on AVX machines.
500 * XXX: should not play games with util_cpu_caps directly as it might
501 * get used for other things outside llvm too.
503 util_cpu_caps
.has_avx
= 0;
504 util_cpu_caps
.has_avx2
= 0;
505 util_cpu_caps
.has_f16c
= 0;
506 util_cpu_caps
.has_fma
= 0;
508 if ((LLVM_VERSION_MAJOR
== 3 && LLVM_VERSION_MINOR
< 4) || !use_mcjit
) {
509 /* AVX2 support has only been tested with LLVM 3.4, and it requires
511 util_cpu_caps
.has_avx2
= 0;
514 #ifdef PIPE_ARCH_PPC_64
515 /* Set the NJ bit in VSCR to 0 so denormalized values are handled as
516 * specified by IEEE standard (PowerISA 2.06 - Section 6.3). This guarantees
517 * that some rounding and half-float to float handling does not round
519 * XXX: should eventually follow same logic on all platforms.
520 * Right now denorms get explicitly disabled (but elsewhere) for x86,
521 * whereas ppc64 explicitly enables them...
523 if (util_cpu_caps
.has_altivec
) {
524 unsigned short mask
[] = { 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF,
525 0xFFFF, 0xFFFF, 0xFFFE, 0xFFFF };
536 gallivm_initialized
= TRUE
;
544 * Create a new gallivm_state object.
546 struct gallivm_state
*
547 gallivm_create(const char *name
, LLVMContextRef context
)
549 struct gallivm_state
*gallivm
;
551 gallivm
= CALLOC_STRUCT(gallivm_state
);
553 if (!init_gallivm_state(gallivm
, name
, context
)) {
564 * Destroy a gallivm_state object.
567 gallivm_destroy(struct gallivm_state
*gallivm
)
569 gallivm_free_ir(gallivm
);
570 gallivm_free_code(gallivm
);
576 * Validate a function.
577 * Verification is only done with debug builds.
580 gallivm_verify_function(struct gallivm_state
*gallivm
,
583 /* Verify the LLVM IR. If invalid, dump and abort */
585 if (LLVMVerifyFunction(func
, LLVMPrintMessageAction
)) {
586 lp_debug_dump_value(func
);
592 if (gallivm_debug
& GALLIVM_DEBUG_IR
) {
593 /* Print the LLVM IR to stderr */
594 lp_debug_dump_value(func
);
602 * This does IR optimization on all functions in the module.
605 gallivm_compile_module(struct gallivm_state
*gallivm
)
608 int64_t time_begin
= 0;
610 assert(!gallivm
->compiled
);
612 if (gallivm
->builder
) {
613 LLVMDisposeBuilder(gallivm
->builder
);
614 gallivm
->builder
= NULL
;
617 /* Dump bitcode to a file */
618 if (gallivm_debug
& GALLIVM_DEBUG_DUMP_BC
) {
620 assert(gallivm
->module_name
);
621 snprintf(filename
, sizeof(filename
), "ir_%s.bc", gallivm
->module_name
);
622 LLVMWriteBitcodeToFile(gallivm
->module
, filename
);
623 debug_printf("%s written\n", filename
);
624 debug_printf("Invoke as \"opt %s %s | llc -O%d %s%s\"\n",
625 gallivm_debug
& GALLIVM_PERF_NO_OPT
? "-mem2reg" :
626 "-sroa -early-cse -simplifycfg -reassociate "
627 "-mem2reg -constprop -instcombine -gvn",
628 filename
, gallivm_debug
& GALLIVM_PERF_NO_OPT
? 0 : 2,
629 (LLVM_VERSION_MAJOR
> 3 || (LLVM_VERSION_MAJOR
== 3 && LLVM_VERSION_MINOR
>= 5)) ? "[-mcpu=<-mcpu option>] " : "",
630 "[-mattr=<-mattr option(s)>]");
633 if (gallivm_debug
& GALLIVM_DEBUG_PERF
)
634 time_begin
= os_time_get();
636 #if GALLIVM_HAVE_CORO
637 LLVMRunPassManager(gallivm
->cgpassmgr
, gallivm
->module
);
639 /* Run optimization passes */
640 LLVMInitializeFunctionPassManager(gallivm
->passmgr
);
641 func
= LLVMGetFirstFunction(gallivm
->module
);
644 debug_printf("optimizing func %s...\n", LLVMGetValueName(func
));
647 /* Disable frame pointer omission on debug/profile builds */
648 /* XXX: And workaround http://llvm.org/PR21435 */
649 #if (LLVM_VERSION_MAJOR > 3 || (LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 7)) && \
650 (defined(DEBUG) || defined(PROFILE) || \
651 defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
652 LLVMAddTargetDependentFunctionAttr(func
, "no-frame-pointer-elim", "true");
653 LLVMAddTargetDependentFunctionAttr(func
, "no-frame-pointer-elim-non-leaf", "true");
656 LLVMRunFunctionPassManager(gallivm
->passmgr
, func
);
657 func
= LLVMGetNextFunction(func
);
659 LLVMFinalizeFunctionPassManager(gallivm
->passmgr
);
661 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
662 int64_t time_end
= os_time_get();
663 int time_msec
= (int)((time_end
- time_begin
) / 1000);
664 assert(gallivm
->module_name
);
665 debug_printf("optimizing module %s took %d msec\n",
666 gallivm
->module_name
, time_msec
);
670 /* Setting the module's DataLayout to an empty string will cause the
671 * ExecutionEngine to copy to the DataLayout string from its target
672 * machine to the module. As of LLVM 3.8 the module and the execution
673 * engine are required to have the same DataLayout.
675 * We must make sure we do this after running the optimization passes,
676 * because those passes need a correct datalayout string. For example,
677 * if those optimization passes see an empty datalayout, they will assume
678 * this is a little endian target and will do optimizations that break big
681 * TODO: This is just a temporary work-around. The correct solution is
682 * for gallivm_init_state() to create a TargetMachine and pull the
683 * DataLayout from there. Currently, the TargetMachine used by llvmpipe
684 * is being implicitly created by the EngineBuilder in
685 * lp_build_create_jit_compiler_for_module()
687 LLVMSetDataLayout(gallivm
->module
, "");
688 assert(!gallivm
->engine
);
689 if (!init_gallivm_engine(gallivm
)) {
693 assert(gallivm
->engine
);
697 if (gallivm_debug
& GALLIVM_DEBUG_ASM
) {
698 LLVMValueRef llvm_func
= LLVMGetFirstFunction(gallivm
->module
);
702 * Need to filter out functions which don't have an implementation,
703 * such as the intrinsics. May not be sufficient in case of IPO?
704 * LLVMGetPointerToGlobal() will abort otherwise.
706 if (!LLVMIsDeclaration(llvm_func
)) {
707 void *func_code
= LLVMGetPointerToGlobal(gallivm
->engine
, llvm_func
);
708 lp_disassemble(llvm_func
, func_code
);
710 llvm_func
= LLVMGetNextFunction(llvm_func
);
716 LLVMValueRef llvm_func
= LLVMGetFirstFunction(gallivm
->module
);
719 if (!LLVMIsDeclaration(llvm_func
)) {
720 void *func_code
= LLVMGetPointerToGlobal(gallivm
->engine
, llvm_func
);
721 lp_profile(llvm_func
, func_code
);
723 llvm_func
= LLVMGetNextFunction(llvm_func
);
732 gallivm_jit_function(struct gallivm_state
*gallivm
,
736 func_pointer jit_func
;
737 int64_t time_begin
= 0;
739 assert(gallivm
->compiled
);
740 assert(gallivm
->engine
);
742 if (gallivm_debug
& GALLIVM_DEBUG_PERF
)
743 time_begin
= os_time_get();
745 code
= LLVMGetPointerToGlobal(gallivm
->engine
, func
);
747 jit_func
= pointer_to_func(code
);
749 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
750 int64_t time_end
= os_time_get();
751 int time_msec
= (int)(time_end
- time_begin
) / 1000;
752 debug_printf(" jitting func %s took %d msec\n",
753 LLVMGetValueName(func
), time_msec
);