LLVMBool
lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
lp_generated_code **OutCode,
LLVMModuleRef M,
LLVMMCJITMemoryManagerRef CMM,
unsigned OptLevel,
int useMCJIT,
char **OutError)
{
using namespace llvm;
std::string Error;
#if HAVE_LLVM >= 0x0306
EngineBuilder builder(std::unique_ptr<Module>(unwrap(M)));
#else
EngineBuilder builder(unwrap(M));
#endif
/**
* LLVM 3.1+ haven't more "extern unsigned llvm::StackAlignmentOverride" and
* friends for configuring code generation options, like stack alignment.
*/
TargetOptions options;
#if defined(PIPE_ARCH_X86)
options.StackAlignmentOverride = 4;
#if HAVE_LLVM < 0x0304
options.RealignStack = true;
#endif
#endif
#if defined(DEBUG) && HAVE_LLVM < 0x0307
options.JITEmitDebugInfo = true;
#endif
/* XXX: Workaround http://llvm.org/PR21435 */
#if defined(DEBUG) || defined(PROFILE) || \
(HAVE_LLVM >= 0x0303 && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)))
#if HAVE_LLVM < 0x0304
options.NoFramePointerElimNonLeaf = true;
#endif
#if HAVE_LLVM < 0x0307
options.NoFramePointerElim = true;
#endif
#endif
builder.setEngineKind(EngineKind::JIT)
.setErrorStr(&Error)
.setTargetOptions(options)
.setOptLevel((CodeGenOpt::Level)OptLevel);
if (useMCJIT) {
#if HAVE_LLVM < 0x0306
builder.setUseMCJIT(true);
#endif
#ifdef _WIN32
/*
* MCJIT works on Windows, but currently only through ELF object format.
*
* XXX: We could use `LLVM_HOST_TRIPLE "-elf"` but LLVM_HOST_TRIPLE has
* different strings for MinGW/MSVC, so better play it safe and be
* explicit.
*/
# ifdef _WIN64
LLVMSetTarget(M, "x86_64-pc-win32-elf");
# else
LLVMSetTarget(M, "i686-pc-win32-elf");
# endif
#endif
}
llvm::SmallVector<std::string, 16> MAttrs;
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
/*
* We need to unset attributes because sometimes LLVM mistakenly assumes
* certain features are present given the processor name.
*
* https://bugs.freedesktop.org/show_bug.cgi?id=92214
* http://llvm.org/PR25021
* http://llvm.org/PR19429
* http://llvm.org/PR16721
*/
MAttrs.push_back(util_cpu_caps.has_sse ? "+sse" : "-sse" );
MAttrs.push_back(util_cpu_caps.has_sse2 ? "+sse2" : "-sse2" );
MAttrs.push_back(util_cpu_caps.has_sse3 ? "+sse3" : "-sse3" );
MAttrs.push_back(util_cpu_caps.has_ssse3 ? "+ssse3" : "-ssse3" );
#if HAVE_LLVM >= 0x0304
MAttrs.push_back(util_cpu_caps.has_sse4_1 ? "+sse4.1" : "-sse4.1");
#else
MAttrs.push_back(util_cpu_caps.has_sse4_1 ? "+sse41" : "-sse41" );
#endif
#if HAVE_LLVM >= 0x0304
MAttrs.push_back(util_cpu_caps.has_sse4_2 ? "+sse4.2" : "-sse4.2");
#else
MAttrs.push_back(util_cpu_caps.has_sse4_2 ? "+sse42" : "-sse42" );
#endif
/*
* AVX feature is not automatically detected from CPUID by the X86 target
* yet, because the old (yet default) JIT engine is not capable of
* emitting the opcodes. On newer llvm versions it is and at least some
* versions (tested with 3.3) will emit avx opcodes without this anyway.
*/
MAttrs.push_back(util_cpu_caps.has_avx ? "+avx" : "-avx");
MAttrs.push_back(util_cpu_caps.has_f16c ? "+f16c" : "-f16c");
if (HAVE_LLVM >= 0x0304) {
MAttrs.push_back(util_cpu_caps.has_fma ? "+fma" : "-fma");
} else {
/*
* The old JIT in LLVM 3.3 has a bug encoding llvm.fmuladd.f32 and
* llvm.fmuladd.v2f32 intrinsics when FMA is available.
*/
MAttrs.push_back("-fma");
}
MAttrs.push_back(util_cpu_caps.has_avx2 ? "+avx2" : "-avx2");
/* disable avx512 and all subvariants */
#if HAVE_LLVM >= 0x0304
MAttrs.push_back("-avx512cd");
MAttrs.push_back("-avx512er");
MAttrs.push_back("-avx512f");
MAttrs.push_back("-avx512pf");
#endif
#if HAVE_LLVM >= 0x0305
MAttrs.push_back("-avx512bw");
MAttrs.push_back("-avx512dq");
MAttrs.push_back("-avx512vl");
#endif
#endif
#if defined(PIPE_ARCH_PPC)
MAttrs.push_back(util_cpu_caps.has_altivec ? "+altivec" : "-altivec");
#if HAVE_LLVM >= 0x0304
/*
* Make sure VSX instructions are disabled
* See LLVM bug https://llvm.org/bugs/show_bug.cgi?id=25503#c7
*/
if (util_cpu_caps.has_altivec) {
MAttrs.push_back("-vsx");
}
#endif
#endif
builder.setMAttrs(MAttrs);
#if HAVE_LLVM >= 0x0305
StringRef MCPU = llvm::sys::getHostCPUName();
/*
* The cpu bits are no longer set automatically, so need to set mcpu manually.
* Note that the MAttrs set above will be sort of ignored (since we should
* not set any which would not be set by specifying the cpu anyway).
* It ought to be safe though since getHostCPUName() should include bits
* not only from the cpu but environment as well (for instance if it's safe
* to use avx instructions which need OS support). According to
* http://llvm.org/bugs/show_bug.cgi?id=19429 however if I understand this
* right it may be necessary to specify older cpu (or disable mattrs) though
* when not using MCJIT so no instructions are generated which the old JIT
* can't handle. Not entirely sure if we really need to do anything yet.
*/
builder.setMCPU(MCPU);
#endif
ShaderMemoryManager *MM = NULL;
if (useMCJIT) {
BaseMemoryManager* JMM = reinterpret_cast<BaseMemoryManager*>(CMM);
MM = new ShaderMemoryManager(JMM);
*OutCode = MM->getGeneratedCode();
#if HAVE_LLVM >= 0x0306
builder.setMCJITMemoryManager(std::unique_ptr<RTDyldMemoryManager>(MM));
MM = NULL; // ownership taken by std::unique_ptr
#elif HAVE_LLVM > 0x0303
builder.setMCJITMemoryManager(MM);
#else
builder.setJITMemoryManager(MM);
#endif
} else {
#if HAVE_LLVM < 0x0306
BaseMemoryManager* JMM = reinterpret_cast<BaseMemoryManager*>(CMM);
MM = new ShaderMemoryManager(JMM);
*OutCode = MM->getGeneratedCode();
builder.setJITMemoryManager(MM);
#else
assert(0);
#endif
}
ExecutionEngine *JIT;
JIT = builder.create();
#if LLVM_USE_INTEL_JITEVENTS
JITEventListener *JEL = JITEventListener::createIntelJITEventListener();
JIT->RegisterJITEventListener(JEL);
#endif
if (JIT) {
*OutJIT = wrap(JIT);
return 0;
}
lp_free_generated_code(*OutCode);
*OutCode = 0;
delete MM;
*OutError = strdup(Error.c_str());
return 1;
}
LLVMBool
lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
lp_generated_code **OutCode,
LLVMModuleRef M,
LLVMMCJITMemoryManagerRef CMM,
unsigned OptLevel,
int useMCJIT,
char **OutError)
{
using namespace llvm;
std::string Error;
#if HAVE_LLVM >= 0x0306
EngineBuilder builder(std::unique_ptr<Module>(unwrap(M)));
#else
EngineBuilder builder(unwrap(M));
#endif
/**
* LLVM 3.1+ haven't more "extern unsigned llvm::StackAlignmentOverride" and
* friends for configuring code generation options, like stack alignment.
*/
TargetOptions options;
#if defined(PIPE_ARCH_X86)
options.StackAlignmentOverride = 4;
#if HAVE_LLVM < 0x0304
options.RealignStack = true;
#endif
#endif
#if defined(DEBUG) && HAVE_LLVM < 0x0307
options.JITEmitDebugInfo = true;
#endif
/* XXX: Workaround http://llvm.org/PR21435 */
#if defined(DEBUG) || defined(PROFILE) || \
(HAVE_LLVM >= 0x0303 && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)))
#if HAVE_LLVM < 0x0304
options.NoFramePointerElimNonLeaf = true;
#endif
#if HAVE_LLVM < 0x0307
options.NoFramePointerElim = true;
#endif
#endif
builder.setEngineKind(EngineKind::JIT)
.setErrorStr(&Error)
.setTargetOptions(options)
.setOptLevel((CodeGenOpt::Level)OptLevel);
if (useMCJIT) {
#if HAVE_LLVM < 0x0306
builder.setUseMCJIT(true);
#endif
#ifdef _WIN32
/*
* MCJIT works on Windows, but currently only through ELF object format.
*
* XXX: We could use `LLVM_HOST_TRIPLE "-elf"` but LLVM_HOST_TRIPLE has
* different strings for MinGW/MSVC, so better play it safe and be
* explicit.
*/
# ifdef _WIN64
LLVMSetTarget(M, "x86_64-pc-win32-elf");
# else
LLVMSetTarget(M, "i686-pc-win32-elf");
# endif
#endif
}
llvm::SmallVector<std::string, 16> MAttrs;
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
#if HAVE_LLVM >= 0x0400
/* llvm-3.7+ implements sys::getHostCPUFeatures for x86,
* which allows us to enable/disable code generation based
* on the results of cpuid.
*/
llvm::StringMap<bool> features;
llvm::sys::getHostCPUFeatures(features);
for (StringMapIterator<bool> f = features.begin();
f != features.end();
++f) {
MAttrs.push_back(((*f).second ? "+" : "-") + (*f).first().str());
}
#else
/*
* We need to unset attributes because sometimes LLVM mistakenly assumes
* certain features are present given the processor name.
*
* https://bugs.freedesktop.org/show_bug.cgi?id=92214
* http://llvm.org/PR25021
* http://llvm.org/PR19429
* http://llvm.org/PR16721
*/
MAttrs.push_back(util_cpu_caps.has_sse ? "+sse" : "-sse" );
MAttrs.push_back(util_cpu_caps.has_sse2 ? "+sse2" : "-sse2" );
MAttrs.push_back(util_cpu_caps.has_sse3 ? "+sse3" : "-sse3" );
MAttrs.push_back(util_cpu_caps.has_ssse3 ? "+ssse3" : "-ssse3" );
#if HAVE_LLVM >= 0x0304
MAttrs.push_back(util_cpu_caps.has_sse4_1 ? "+sse4.1" : "-sse4.1");
#else
MAttrs.push_back(util_cpu_caps.has_sse4_1 ? "+sse41" : "-sse41" );
#endif
#if HAVE_LLVM >= 0x0304
MAttrs.push_back(util_cpu_caps.has_sse4_2 ? "+sse4.2" : "-sse4.2");
#else
MAttrs.push_back(util_cpu_caps.has_sse4_2 ? "+sse42" : "-sse42" );
#endif
/*
* AVX feature is not automatically detected from CPUID by the X86 target
* yet, because the old (yet default) JIT engine is not capable of
* emitting the opcodes. On newer llvm versions it is and at least some
* versions (tested with 3.3) will emit avx opcodes without this anyway.
*/
MAttrs.push_back(util_cpu_caps.has_avx ? "+avx" : "-avx");
MAttrs.push_back(util_cpu_caps.has_f16c ? "+f16c" : "-f16c");
if (HAVE_LLVM >= 0x0304) {
MAttrs.push_back(util_cpu_caps.has_fma ? "+fma" : "-fma");
} else {
/*
* The old JIT in LLVM 3.3 has a bug encoding llvm.fmuladd.f32 and
* llvm.fmuladd.v2f32 intrinsics when FMA is available.
*/
MAttrs.push_back("-fma");
}
MAttrs.push_back(util_cpu_caps.has_avx2 ? "+avx2" : "-avx2");
/* disable avx512 and all subvariants */
#if HAVE_LLVM >= 0x0304
MAttrs.push_back("-avx512cd");
MAttrs.push_back("-avx512er");
MAttrs.push_back("-avx512f");
MAttrs.push_back("-avx512pf");
#endif
#if HAVE_LLVM >= 0x0305
MAttrs.push_back("-avx512bw");
MAttrs.push_back("-avx512dq");
MAttrs.push_back("-avx512vl");
#endif
#endif
#endif
#if defined(PIPE_ARCH_PPC)
MAttrs.push_back(util_cpu_caps.has_altivec ? "+altivec" : "-altivec");
#if (HAVE_LLVM >= 0x0304)
#if (HAVE_LLVM < 0x0400)
/*
* Make sure VSX instructions are disabled
* See LLVM bugs:
* https://llvm.org/bugs/show_bug.cgi?id=25503#c7 (fixed in 3.8.1)
* https://llvm.org/bugs/show_bug.cgi?id=26775 (fixed in 3.8.1)
* https://llvm.org/bugs/show_bug.cgi?id=33531 (fixed in 4.0)
* 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)
*/
if (util_cpu_caps.has_altivec) {
MAttrs.push_back("-vsx");
}
#else
/*
* Bug 25503 is fixed, by the same fix that fixed
* bug 26775, in versions of LLVM later than 3.8 (starting with 3.8.1).
* BZ 33531 actually comprises more than one bug, all of
* which are fixed in LLVM 4.0.
*
* With LLVM 4.0 or higher:
* Make sure VSX instructions are ENABLED, unless
* a) the entire -mattr option is overridden via GALLIVM_MATTRS, or
* b) VSX instructions are explicitly enabled/disabled via GALLIVM_VSX=1 or 0.
*/
if (util_cpu_caps.has_altivec) {
char *env_mattrs = getenv("GALLIVM_MATTRS");
if (env_mattrs) {
MAttrs.push_back(env_mattrs);
}
else {
boolean enable_vsx = true;
char *env_vsx = getenv("GALLIVM_VSX");
if (env_vsx && env_vsx[0] == '0') {
enable_vsx = false;
}
if (enable_vsx)
MAttrs.push_back("+vsx");
else
MAttrs.push_back("-vsx");
}
}
#endif
#endif
#endif
builder.setMAttrs(MAttrs);
if (gallivm_debug & (GALLIVM_DEBUG_IR | GALLIVM_DEBUG_ASM | GALLIVM_DEBUG_DUMP_BC)) {
int n = MAttrs.size();
if (n > 0) {
debug_printf("llc -mattr option(s): ");
for (int i = 0; i < n; i++)
debug_printf("%s%s", MAttrs[i].c_str(), (i < n - 1) ? "," : "");
debug_printf("\n");
}
}
#if HAVE_LLVM >= 0x0305
StringRef MCPU = llvm::sys::getHostCPUName();
/*
* The cpu bits are no longer set automatically, so need to set mcpu manually.
* Note that the MAttrs set above will be sort of ignored (since we should
* not set any which would not be set by specifying the cpu anyway).
* It ought to be safe though since getHostCPUName() should include bits
* not only from the cpu but environment as well (for instance if it's safe
* to use avx instructions which need OS support). According to
* http://llvm.org/bugs/show_bug.cgi?id=19429 however if I understand this
* right it may be necessary to specify older cpu (or disable mattrs) though
* when not using MCJIT so no instructions are generated which the old JIT
* can't handle. Not entirely sure if we really need to do anything yet.
*/
#if defined(PIPE_ARCH_LITTLE_ENDIAN) && defined(PIPE_ARCH_PPC_64)
/*
* Versions of LLVM prior to 4.0 lacked a table entry for "POWER8NVL",
* resulting in (big-endian) "generic" being returned on
* little-endian Power8NVL systems. The result was that code that
* attempted to load the least significant 32 bits of a 64-bit quantity
* from memory loaded the wrong half. This resulted in failures in some
* Piglit tests, e.g.
* .../arb_gpu_shader_fp64/execution/conversion/frag-conversion-explicit-double-uint
*/
if (MCPU == "generic")
MCPU = "pwr8";
#endif
builder.setMCPU(MCPU);
if (gallivm_debug & (GALLIVM_DEBUG_IR | GALLIVM_DEBUG_ASM | GALLIVM_DEBUG_DUMP_BC)) {
debug_printf("llc -mcpu option: %s\n", MCPU.str().c_str());
}
#endif
ShaderMemoryManager *MM = NULL;
if (useMCJIT) {
BaseMemoryManager* JMM = reinterpret_cast<BaseMemoryManager*>(CMM);
MM = new ShaderMemoryManager(JMM);
*OutCode = MM->getGeneratedCode();
#if HAVE_LLVM >= 0x0306
builder.setMCJITMemoryManager(std::unique_ptr<RTDyldMemoryManager>(MM));
MM = NULL; // ownership taken by std::unique_ptr
#elif HAVE_LLVM > 0x0303
builder.setMCJITMemoryManager(MM);
#else
builder.setJITMemoryManager(MM);
#endif
} else {
#if HAVE_LLVM < 0x0306
BaseMemoryManager* JMM = reinterpret_cast<BaseMemoryManager*>(CMM);
MM = new ShaderMemoryManager(JMM);
*OutCode = MM->getGeneratedCode();
builder.setJITMemoryManager(MM);
#else
assert(0);
#endif
}
ExecutionEngine *JIT;
JIT = builder.create();
#if LLVM_USE_INTEL_JITEVENTS
JITEventListener *JEL = JITEventListener::createIntelJITEventListener();
JIT->RegisterJITEventListener(JEL);
#endif
if (JIT) {
*OutJIT = wrap(JIT);
return 0;
}
lp_free_generated_code(*OutCode);
*OutCode = 0;
delete MM;
*OutError = strdup(Error.c_str());
return 1;
}
LLVMBool
lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
lp_generated_code **OutCode,
LLVMModuleRef M,
unsigned OptLevel,
int useMCJIT,
char **OutError)
{
using namespace llvm;
std::string Error;
EngineBuilder builder(unwrap(M));
/**
* LLVM 3.1+ haven't more "extern unsigned llvm::StackAlignmentOverride" and
* friends for configuring code generation options, like stack alignment.
*/
TargetOptions options;
#if defined(PIPE_ARCH_X86)
options.StackAlignmentOverride = 4;
#if HAVE_LLVM < 0x0304
options.RealignStack = true;
#endif
#endif
#if defined(DEBUG)
options.JITEmitDebugInfo = true;
#endif
#if defined(DEBUG) || defined(PROFILE)
#if HAVE_LLVM < 0x0304
options.NoFramePointerElimNonLeaf = true;
#endif
options.NoFramePointerElim = true;
#endif
builder.setEngineKind(EngineKind::JIT)
.setErrorStr(&Error)
.setTargetOptions(options)
.setOptLevel((CodeGenOpt::Level)OptLevel);
if (useMCJIT) {
builder.setUseMCJIT(true);
#ifdef _WIN32
/*
* MCJIT works on Windows, but currently only through ELF object format.
*/
std::string targetTriple = llvm::sys::getProcessTriple();
targetTriple.append("-elf");
unwrap(M)->setTargetTriple(targetTriple);
#endif
}
llvm::SmallVector<std::string, 1> MAttrs;
if (util_cpu_caps.has_avx) {
/*
* AVX feature is not automatically detected from CPUID by the X86 target
* yet, because the old (yet default) JIT engine is not capable of
* emitting the opcodes. On newer llvm versions it is and at least some
* versions (tested with 3.3) will emit avx opcodes without this anyway.
*/
MAttrs.push_back("+avx");
if (util_cpu_caps.has_f16c) {
MAttrs.push_back("+f16c");
}
builder.setMAttrs(MAttrs);
}
#if HAVE_LLVM >= 0x0305
StringRef MCPU = llvm::sys::getHostCPUName();
/*
* The cpu bits are no longer set automatically, so need to set mcpu manually.
* Note that the MAttrs set above will be sort of ignored (since we should
* not set any which would not be set by specifying the cpu anyway).
* It ought to be safe though since getHostCPUName() should include bits
* not only from the cpu but environment as well (for instance if it's safe
* to use avx instructions which need OS support). According to
* http://llvm.org/bugs/show_bug.cgi?id=19429 however if I understand this
* right it may be necessary to specify older cpu (or disable mattrs) though
* when not using MCJIT so no instructions are generated which the old JIT
* can't handle. Not entirely sure if we really need to do anything yet.
*/
builder.setMCPU(MCPU);
#endif
ShaderMemoryManager *MM = new ShaderMemoryManager();
*OutCode = MM->getGeneratedCode();
builder.setJITMemoryManager(MM);
ExecutionEngine *JIT;
#if HAVE_LLVM >= 0x0302
JIT = builder.create();
#else
/*
* Workaround http://llvm.org/PR12833
*/
StringRef MArch = "";
StringRef MCPU = "";
Triple TT(unwrap(M)->getTargetTriple());
JIT = builder.create(builder.selectTarget(TT, MArch, MCPU, MAttrs));
#endif
if (JIT) {
*OutJIT = wrap(JIT);
return 0;
}
lp_free_generated_code(*OutCode);
*OutCode = 0;
delete MM;
*OutError = strdup(Error.c_str());
return 1;
}