1 /**************************************************************************
3 * Copyright 2010 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 SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
18 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
19 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
20 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 * The above copyright notice and this permission notice (including the
23 * next paragraph) shall be included in all copies or substantial portions
26 **************************************************************************/
30 * The purpose of this module is to expose LLVM functionality not available
31 * through the C++ bindings.
35 // Undef these vars just to silence warnings
36 #undef PACKAGE_BUGREPORT
39 #undef PACKAGE_TARNAME
40 #undef PACKAGE_VERSION
45 #include <llvm/Config/llvm-config.h>
47 #if LLVM_VERSION_MAJOR < 7
48 // Workaround http://llvm.org/PR23628
49 #pragma push_macro("DEBUG")
53 #include <llvm/Config/llvm-config.h>
54 #include <llvm-c/Core.h>
55 #include <llvm-c/Support.h>
56 #include <llvm-c/ExecutionEngine.h>
57 #include <llvm/Target/TargetOptions.h>
58 #include <llvm/ExecutionEngine/ExecutionEngine.h>
59 #include <llvm/ADT/Triple.h>
60 #include <llvm/Analysis/TargetLibraryInfo.h>
61 #include <llvm/ExecutionEngine/SectionMemoryManager.h>
62 #include <llvm/Support/CommandLine.h>
63 #include <llvm/Support/Host.h>
64 #include <llvm/Support/PrettyStackTrace.h>
65 #include <llvm/ExecutionEngine/ObjectCache.h>
66 #include <llvm/Support/TargetSelect.h>
68 #if LLVM_VERSION_MAJOR < 11
69 #include <llvm/IR/CallSite.h>
71 #include <llvm/IR/IRBuilder.h>
72 #include <llvm/IR/Module.h>
73 #include <llvm/Support/CBindingWrapping.h>
75 #include <llvm/Config/llvm-config.h>
76 #if LLVM_USE_INTEL_JITEVENTS
77 #include <llvm/ExecutionEngine/JITEventListener.h>
80 #if LLVM_VERSION_MAJOR < 7
81 // Workaround http://llvm.org/PR23628
82 #pragma pop_macro("DEBUG")
85 #include "c11/threads.h"
86 #include "os/os_thread.h"
87 #include "pipe/p_config.h"
88 #include "util/u_debug.h"
89 #include "util/u_cpu_detect.h"
91 #include "lp_bld_misc.h"
92 #include "lp_bld_debug.h"
96 class LLVMEnsureMultithreaded
{
98 LLVMEnsureMultithreaded()
100 if (!LLVMIsMultithreaded()) {
101 LLVMStartMultithreaded();
106 static LLVMEnsureMultithreaded lLVMEnsureMultithreaded
;
110 static once_flag init_native_targets_once_flag
= ONCE_FLAG_INIT
;
112 static void init_native_targets()
114 // If we have a native target, initialize it to ensure it is linked in and
115 // usable by the JIT.
116 llvm::InitializeNativeTarget();
118 llvm::InitializeNativeTargetAsmPrinter();
120 llvm::InitializeNativeTargetDisassembler();
123 char *env_llc_options
= getenv("GALLIVM_LLC_OPTIONS");
124 if (env_llc_options
) {
126 char *options
[64] = {(char *) "llc"}; // Warning without cast
128 for (n
= 0, option
= strtok(env_llc_options
, " "); option
; n
++, option
= strtok(NULL
, " ")) {
129 options
[n
+ 1] = option
;
131 if (gallivm_debug
& (GALLIVM_DEBUG_IR
| GALLIVM_DEBUG_ASM
| GALLIVM_DEBUG_DUMP_BC
)) {
132 debug_printf("llc additional options (%d):\n", n
);
133 for (int i
= 1; i
<= n
; i
++)
134 debug_printf("\t%s\n", options
[i
]);
137 LLVMParseCommandLineOptions(n
+ 1, options
, NULL
);
144 lp_set_target_options(void)
146 /* The llvm target registry is not thread-safe, so drivers and gallium frontends
147 * that want to initialize targets should use the lp_set_target_options()
148 * function to safely initialize targets.
150 * LLVM targets should be initialized before the driver or gallium frontend tries
151 * to access the registry.
153 call_once(&init_native_targets_once_flag
, init_native_targets
);
157 LLVMTargetLibraryInfoRef
158 gallivm_create_target_library_info(const char *triple
)
160 return reinterpret_cast<LLVMTargetLibraryInfoRef
>(
161 new llvm::TargetLibraryInfoImpl(
162 llvm::Triple(triple
)));
167 gallivm_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info
)
169 delete reinterpret_cast<
170 llvm::TargetLibraryInfoImpl
175 typedef llvm::RTDyldMemoryManager BaseMemoryManager
;
179 * Delegating is tedious but the default manager class is hidden in an
180 * anonymous namespace in LLVM, so we cannot just derive from it to change
183 class DelegatingJITMemoryManager
: public BaseMemoryManager
{
186 virtual BaseMemoryManager
*mgr() const = 0;
190 * From RTDyldMemoryManager
192 virtual uint8_t *allocateCodeSection(uintptr_t Size
,
195 llvm::StringRef SectionName
) {
196 return mgr()->allocateCodeSection(Size
, Alignment
, SectionID
,
199 virtual uint8_t *allocateDataSection(uintptr_t Size
,
202 llvm::StringRef SectionName
,
204 return mgr()->allocateDataSection(Size
, Alignment
, SectionID
,
208 virtual void registerEHFrames(uint8_t *Addr
, uint64_t LoadAddr
, size_t Size
) {
209 mgr()->registerEHFrames(Addr
, LoadAddr
, Size
);
211 #if LLVM_VERSION_MAJOR >= 5
212 virtual void deregisterEHFrames() {
213 mgr()->deregisterEHFrames();
216 virtual void deregisterEHFrames(uint8_t *Addr
, uint64_t LoadAddr
, size_t Size
) {
217 mgr()->deregisterEHFrames(Addr
, LoadAddr
, Size
);
220 virtual void *getPointerToNamedFunction(const std::string
&Name
,
221 bool AbortOnFailure
=true) {
222 return mgr()->getPointerToNamedFunction(Name
, AbortOnFailure
);
224 virtual bool finalizeMemory(std::string
*ErrMsg
= 0) {
225 return mgr()->finalizeMemory(ErrMsg
);
231 * Delegate memory management to one shared manager for more efficient use
232 * of memory than creating a separate pool for each LLVM engine.
233 * Keep generated code until freeGeneratedCode() is called, instead of when
234 * memory manager is destroyed, which happens during engine destruction.
235 * This allows additional memory savings as we don't have to keep the engine
236 * around in order to use the code.
237 * All methods are delegated to the shared manager except destruction and
238 * deallocating code. For the latter we just remember what needs to be
239 * deallocated later. The shared manager is deleted once it is empty.
241 class ShaderMemoryManager
: public DelegatingJITMemoryManager
{
243 BaseMemoryManager
*TheMM
;
245 struct GeneratedCode
{
246 typedef std::vector
<void *> Vec
;
247 Vec FunctionBody
, ExceptionTable
;
248 BaseMemoryManager
*TheMM
;
250 GeneratedCode(BaseMemoryManager
*MM
) {
260 BaseMemoryManager
*mgr() const {
266 ShaderMemoryManager(BaseMemoryManager
* MM
) {
268 code
= new GeneratedCode(MM
);
271 virtual ~ShaderMemoryManager() {
273 * 'code' is purposely not deleted. It is the user's responsibility
274 * to call getGeneratedCode() and freeGeneratedCode().
278 struct lp_generated_code
*getGeneratedCode() {
279 return (struct lp_generated_code
*) code
;
282 static void freeGeneratedCode(struct lp_generated_code
*code
) {
283 delete (GeneratedCode
*) code
;
286 virtual void deallocateFunctionBody(void *Body
) {
287 // remember for later deallocation
288 code
->FunctionBody
.push_back(Body
);
292 class LPObjectCache
: public llvm::ObjectCache
{
295 struct lp_cached_code
*cache_out
;
297 LPObjectCache(struct lp_cached_code
*cache
) {
304 void notifyObjectCompiled(const llvm::Module
*M
, llvm::MemoryBufferRef Obj
) {
305 const std::string ModuleID
= M
->getModuleIdentifier();
307 fprintf(stderr
, "CACHE ALREADY HAS MODULE OBJECT\n");
309 cache_out
->data_size
= Obj
.getBufferSize();
310 cache_out
->data
= malloc(cache_out
->data_size
);
311 memcpy(cache_out
->data
, Obj
.getBufferStart(), cache_out
->data_size
);
314 virtual std::unique_ptr
<llvm::MemoryBuffer
> getObject(const llvm::Module
*M
) {
315 if (cache_out
->data_size
) {
316 return llvm::MemoryBuffer::getMemBuffer(llvm::StringRef((const char *)cache_out
->data
, cache_out
->data_size
), "", false);
324 * Same as LLVMCreateJITCompilerForModule, but:
325 * - allows using MCJIT and enabling AVX feature where available.
326 * - set target options
329 * - llvm/lib/ExecutionEngine/ExecutionEngineBindings.cpp
330 * - llvm/tools/lli/lli.cpp
331 * - http://markmail.org/message/ttkuhvgj4cxxy2on#query:+page:1+mid:aju2dggerju3ivd3+state:results
335 lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef
*OutJIT
,
336 lp_generated_code
**OutCode
,
337 struct lp_cached_code
*cache_out
,
339 LLVMMCJITMemoryManagerRef CMM
,
343 using namespace llvm
;
346 EngineBuilder
builder(std::unique_ptr
<Module
>(unwrap(M
)));
349 * LLVM 3.1+ haven't more "extern unsigned llvm::StackAlignmentOverride" and
350 * friends for configuring code generation options, like stack alignment.
352 TargetOptions options
;
353 #if defined(PIPE_ARCH_X86)
354 options
.StackAlignmentOverride
= 4;
357 builder
.setEngineKind(EngineKind::JIT
)
359 .setTargetOptions(options
)
360 .setOptLevel((CodeGenOpt::Level
)OptLevel
);
364 * MCJIT works on Windows, but currently only through ELF object format.
366 * XXX: We could use `LLVM_HOST_TRIPLE "-elf"` but LLVM_HOST_TRIPLE has
367 * different strings for MinGW/MSVC, so better play it safe and be
371 LLVMSetTarget(M
, "x86_64-pc-win32-elf");
373 LLVMSetTarget(M
, "i686-pc-win32-elf");
377 llvm::SmallVector
<std::string
, 16> MAttrs
;
379 #if LLVM_VERSION_MAJOR >= 4 && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) || defined(PIPE_ARCH_ARM))
380 /* llvm-3.3+ implements sys::getHostCPUFeatures for Arm
381 * and llvm-3.7+ for x86, which allows us to enable/disable
382 * code generation based on the results of cpuid on these
385 llvm::StringMap
<bool> features
;
386 llvm::sys::getHostCPUFeatures(features
);
388 for (StringMapIterator
<bool> f
= features
.begin();
391 MAttrs
.push_back(((*f
).second
? "+" : "-") + (*f
).first().str());
393 #elif defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
395 * We need to unset attributes because sometimes LLVM mistakenly assumes
396 * certain features are present given the processor name.
398 * https://bugs.freedesktop.org/show_bug.cgi?id=92214
399 * http://llvm.org/PR25021
400 * http://llvm.org/PR19429
401 * http://llvm.org/PR16721
403 MAttrs
.push_back(util_cpu_caps
.has_sse
? "+sse" : "-sse" );
404 MAttrs
.push_back(util_cpu_caps
.has_sse2
? "+sse2" : "-sse2" );
405 MAttrs
.push_back(util_cpu_caps
.has_sse3
? "+sse3" : "-sse3" );
406 MAttrs
.push_back(util_cpu_caps
.has_ssse3
? "+ssse3" : "-ssse3" );
407 MAttrs
.push_back(util_cpu_caps
.has_sse4_1
? "+sse4.1" : "-sse4.1");
408 MAttrs
.push_back(util_cpu_caps
.has_sse4_2
? "+sse4.2" : "-sse4.2");
410 * AVX feature is not automatically detected from CPUID by the X86 target
411 * yet, because the old (yet default) JIT engine is not capable of
412 * emitting the opcodes. On newer llvm versions it is and at least some
413 * versions (tested with 3.3) will emit avx opcodes without this anyway.
415 MAttrs
.push_back(util_cpu_caps
.has_avx
? "+avx" : "-avx");
416 MAttrs
.push_back(util_cpu_caps
.has_f16c
? "+f16c" : "-f16c");
417 MAttrs
.push_back(util_cpu_caps
.has_fma
? "+fma" : "-fma");
418 MAttrs
.push_back(util_cpu_caps
.has_avx2
? "+avx2" : "-avx2");
419 /* disable avx512 and all subvariants */
420 MAttrs
.push_back("-avx512cd");
421 MAttrs
.push_back("-avx512er");
422 MAttrs
.push_back("-avx512f");
423 MAttrs
.push_back("-avx512pf");
424 MAttrs
.push_back("-avx512bw");
425 MAttrs
.push_back("-avx512dq");
426 MAttrs
.push_back("-avx512vl");
428 #if defined(PIPE_ARCH_ARM)
429 if (!util_cpu_caps
.has_neon
) {
430 MAttrs
.push_back("-neon");
431 MAttrs
.push_back("-crypto");
432 MAttrs
.push_back("-vfp2");
436 #if defined(PIPE_ARCH_PPC)
437 MAttrs
.push_back(util_cpu_caps
.has_altivec
? "+altivec" : "-altivec");
438 #if (LLVM_VERSION_MAJOR < 4)
440 * Make sure VSX instructions are disabled
442 * https://llvm.org/bugs/show_bug.cgi?id=25503#c7 (fixed in 3.8.1)
443 * https://llvm.org/bugs/show_bug.cgi?id=26775 (fixed in 3.8.1)
444 * https://llvm.org/bugs/show_bug.cgi?id=33531 (fixed in 4.0)
445 * https://llvm.org/bugs/show_bug.cgi?id=34647 (llc performance on certain unusual shader IR; intro'd in 4.0, pending as of 5.0)
447 if (util_cpu_caps
.has_altivec
) {
448 MAttrs
.push_back("-vsx");
452 * Bug 25503 is fixed, by the same fix that fixed
453 * bug 26775, in versions of LLVM later than 3.8 (starting with 3.8.1).
454 * BZ 33531 actually comprises more than one bug, all of
455 * which are fixed in LLVM 4.0.
457 * With LLVM 4.0 or higher:
458 * Make sure VSX instructions are ENABLED (if supported), unless
459 * VSX instructions are explicitly enabled/disabled via GALLIVM_VSX=1 or 0.
461 if (util_cpu_caps
.has_altivec
) {
462 MAttrs
.push_back(util_cpu_caps
.has_vsx
? "+vsx" : "-vsx");
467 builder
.setMAttrs(MAttrs
);
469 if (gallivm_debug
& (GALLIVM_DEBUG_IR
| GALLIVM_DEBUG_ASM
| GALLIVM_DEBUG_DUMP_BC
)) {
470 int n
= MAttrs
.size();
472 debug_printf("llc -mattr option(s): ");
473 for (int i
= 0; i
< n
; i
++)
474 debug_printf("%s%s", MAttrs
[i
].c_str(), (i
< n
- 1) ? "," : "");
479 StringRef MCPU
= llvm::sys::getHostCPUName();
481 * The cpu bits are no longer set automatically, so need to set mcpu manually.
482 * Note that the MAttrs set above will be sort of ignored (since we should
483 * not set any which would not be set by specifying the cpu anyway).
484 * It ought to be safe though since getHostCPUName() should include bits
485 * not only from the cpu but environment as well (for instance if it's safe
486 * to use avx instructions which need OS support). According to
487 * http://llvm.org/bugs/show_bug.cgi?id=19429 however if I understand this
488 * right it may be necessary to specify older cpu (or disable mattrs) though
489 * when not using MCJIT so no instructions are generated which the old JIT
490 * can't handle. Not entirely sure if we really need to do anything yet.
493 #ifdef PIPE_ARCH_PPC_64
495 * Large programs, e.g. gnome-shell and firefox, may tax the addressability
496 * of the Medium code model once dynamically generated JIT-compiled shader
497 * programs are linked in and relocated. Yet the default code model as of
498 * LLVM 8 is Medium or even Small.
499 * The cost of changing from Medium to Large is negligible:
500 * - an additional 8-byte pointer stored immediately before the shader entrypoint;
501 * - change an add-immediate (addis) instruction to a load (ld).
503 builder
.setCodeModel(CodeModel::Large
);
505 #if UTIL_ARCH_LITTLE_ENDIAN
507 * Versions of LLVM prior to 4.0 lacked a table entry for "POWER8NVL",
508 * resulting in (big-endian) "generic" being returned on
509 * little-endian Power8NVL systems. The result was that code that
510 * attempted to load the least significant 32 bits of a 64-bit quantity
511 * from memory loaded the wrong half. This resulted in failures in some
513 * .../arb_gpu_shader_fp64/execution/conversion/frag-conversion-explicit-double-uint
515 if (MCPU
== "generic")
519 builder
.setMCPU(MCPU
);
520 if (gallivm_debug
& (GALLIVM_DEBUG_IR
| GALLIVM_DEBUG_ASM
| GALLIVM_DEBUG_DUMP_BC
)) {
521 debug_printf("llc -mcpu option: %s\n", MCPU
.str().c_str());
524 ShaderMemoryManager
*MM
= NULL
;
525 BaseMemoryManager
* JMM
= reinterpret_cast<BaseMemoryManager
*>(CMM
);
526 MM
= new ShaderMemoryManager(JMM
);
527 *OutCode
= MM
->getGeneratedCode();
529 builder
.setMCJITMemoryManager(std::unique_ptr
<RTDyldMemoryManager
>(MM
));
530 MM
= NULL
; // ownership taken by std::unique_ptr
532 ExecutionEngine
*JIT
;
534 JIT
= builder
.create();
537 LPObjectCache
*objcache
= new LPObjectCache(cache_out
);
538 JIT
->setObjectCache(objcache
);
539 cache_out
->jit_obj_cache
= (void *)objcache
;
542 #if LLVM_USE_INTEL_JITEVENTS
543 JITEventListener
*JEL
= JITEventListener::createIntelJITEventListener();
544 JIT
->RegisterJITEventListener(JEL
);
550 lp_free_generated_code(*OutCode
);
553 *OutError
= strdup(Error
.c_str());
560 lp_free_generated_code(struct lp_generated_code
*code
)
562 ShaderMemoryManager::freeGeneratedCode(code
);
566 LLVMMCJITMemoryManagerRef
567 lp_get_default_memory_manager()
569 BaseMemoryManager
*mm
;
570 mm
= new llvm::SectionMemoryManager();
571 return reinterpret_cast<LLVMMCJITMemoryManagerRef
>(mm
);
576 lp_free_memory_manager(LLVMMCJITMemoryManagerRef memorymgr
)
578 delete reinterpret_cast<BaseMemoryManager
*>(memorymgr
);
582 lp_free_objcache(void *objcache_ptr
)
584 LPObjectCache
*objcache
= (LPObjectCache
*)objcache_ptr
;
588 extern "C" LLVMValueRef
589 lp_get_called_value(LLVMValueRef call
)
591 return LLVMGetCalledValue(call
);
595 lp_is_function(LLVMValueRef v
)
597 return LLVMGetValueKind(v
) == LLVMFunctionValueKind
;