Reputation: 597
The following lengthy C program generates a simple LLVM module containing a function that merely calls llvm.x86.sse41.round.ps
. It emits the bitcode file and then runs the code generated by LLVM. My question is how do I find out target triple and instruction extensions like SSE or AVX of the host machine and how do I add this information to the LLVM module or how do I otherwise tell it to the LLVM execution engine. Here is, what I do:
$ cat ctest/avx-instruction-selection.c
#include <llvm-c/Core.h>
#include <llvm-c/Target.h>
#include <llvm-c/ExecutionEngine.h>
#include <llvm-c/BitWriter.h>
#include <llvm-c/Transforms/Scalar.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#if 1
const int vectorSize = 4;
const char* roundName = "llvm.x86.sse41.round.ps";
#else
const int vectorSize = 8;
const char* roundName = "llvm.x86.avx.round.ps.256";
#endif
int main ()
{
LLVMModuleRef module;
LLVMExecutionEngineRef execEngine;
LLVMTargetDataRef targetData;
LLVMTypeRef floatType, vectorType, ptrType, voidType, funcType, roundType, int32Type;
LLVMValueRef func, roundFunc;
LLVMValueRef param, loaded, const1, callRound;
LLVMBuilderRef builder;
LLVMBasicBlockRef block;
const int false = 0;
LLVMInitializeX86TargetInfo();
LLVMInitializeX86Target();
LLVMInitializeX86TargetMC();
module = LLVMModuleCreateWithName("_module");
LLVMSetTarget(module, "x86_64-unknown-linux-gnu");
floatType = LLVMFloatType();
vectorType = LLVMVectorType(floatType, vectorSize);
ptrType = LLVMPointerType(vectorType, 0);
voidType = LLVMVoidType();
LLVMTypeRef roundParams[] = { ptrType };
roundType = LLVMFunctionType(voidType, roundParams, 1, false);
func = LLVMAddFunction(module, "round", roundType);
LLVMSetLinkage(func, LLVMExternalLinkage);
builder = LLVMCreateBuilder();
block = LLVMAppendBasicBlock(func, "_L1");
LLVMPositionBuilderAtEnd(builder, block);
param = LLVMGetParam(func, 0);
loaded = LLVMBuildLoad(builder, param, "");
int32Type = LLVMIntType(32);
LLVMTypeRef funcParams[] = { vectorType, int32Type } ;
funcType = LLVMFunctionType(vectorType, funcParams, 2, false);
roundFunc = LLVMAddFunction(module, roundName, funcType);
LLVMSetLinkage(roundFunc, LLVMExternalLinkage);
const1 = LLVMConstInt(int32Type, 1, false);
LLVMValueRef callParams [] = { loaded, const1 } ;
callRound = LLVMBuildCall(builder, roundFunc, callParams, 2, "");
LLVMSetInstructionCallConv(callRound, 0);
LLVMAddInstrAttribute(callRound, 0, 0);
LLVMBuildStore(builder, callRound, param);
LLVMBuildRetVoid(builder);
LLVMWriteBitcodeToFile(module, "round-avx.bc");
char *errorMsg;
LLVMCreateExecutionEngineForModule(&execEngine, module, &errorMsg);
targetData = LLVMGetExecutionEngineTargetData(execEngine);
size_t vectorSize0 = LLVMStoreSizeOfType(targetData, vectorType);
size_t vectorAlign = LLVMABIAlignmentOfType(targetData, vectorType);
float vector[vectorSize];
printf("%lx, size %lx, align %lx\n", (size_t)vector, vectorSize0, vectorAlign);
LLVMGenericValueRef genericVector = LLVMCreateGenericValueOfPointer(vector);
LLVMGenericValueRef runParams[] = { genericVector } ;
LLVMRunFunction(execEngine, func, 1, runParams);
return 0;
}
$ gcc -Wall -o ctest/avx-instruction-selection ctest/avx-instruction-selection.c `/usr/lib/llvm-3.4/bin/llvm-config --cflags --ldflags` -lLLVM-3.4
$ ctest/avx-instruction-selection
7fff590431c0, size 10, align 10
$ ls round-avx.bc
round-avx.bc
$ llvm-dis -o - round-avx.bc
; ModuleID = 'round-avx.bc'
target triple = "x86_64-unknown-linux-gnu"
define void @round(<4 x float>*) {
_L1:
%1 = load <4 x float>* %0
%2 = call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %1, i32 1)
store <4 x float> %2, <4 x float>* %0
ret void
}
; Function Attrs: nounwind readnone
declare <4 x float> @llvm.x86.sse41.round.ps(<4 x float>, i32) #0
attributes #0 = { nounwind readnone }
$ gcc -Wall -o ctest/avx-instruction-selection ctest/avx-instruction-selection.c `/usr/lib/llvm-3.5/bin/llvm-config --cflags --ldflags` -lLLVM-3.5
$ ctest/avx-instruction-selection
7ffed6170350, size 10, align 10
LLVM ERROR: Cannot select: intrinsic %llvm.x86.sse41.round.ps
$ gcc -Wall -o ctest/avx-instruction-selection ctest/avx-instruction-selection.c `/usr/lib/llvm-3.6/bin/llvm-config --cflags --ldflags` -lLLVM-3.6
$ ctest/avx-instruction-selection
7ffeae91eb40, size 10, align 10
LLVM ERROR: Target does not support MC emission!
$ gcc -Wall -o ctest/avx-instruction-selection ctest/avx-instruction-selection.c `/usr/lib/llvm-3.7/bin/llvm-config --cflags --ldflags` -lLLVM-3.7
$ ctest/avx-instruction-selection
7fffb6464ea0, size 10, align 10
LLVM ERROR: Target does not support MC emission!
$ gcc -Wall -o ctest/avx-instruction-selection ctest/avx-instruction-selection.c `/usr/lib/llvm-3.8/bin/llvm-config --cflags --ldflags` -lLLVM-3.8
$ ctest/avx-instruction-selection
7ffd5e233000, size 10, align 10
LLVM ERROR: Target does not support MC emission!
Summarized: With LLVM-3.4 the example works, with LLVM-3.5 the intrinsic function round.ps
cannot be found and LLVM-3.6 and later say something about MC emissions that I do not understand.
As I understand, LLVM-3.5 does not find the round.ps
intrinsic and I guess that it cannot find it because I have not told it about the existing SSE extension. When running llc
I can add the option -mattr=sse4.1
but how can I tell it to the execution engine?
Second question: How can I find out about the available instruction extensions like SSE of the host machine via the LLVM-C API? On x86 I can call the CPUID instruction but is there a way that works uniformly on all platforms and can LLVM assist detection of extensions?
Third question: I have hard-coded the target triple into the C code. How can I find out the host target-triple via the LLVM-C API?
Last question: What about this MC emission error?
Upvotes: 2
Views: 2586
Reputation: 597
After trying around a lot I think the answer is as follows:
Replace the lines
LLVMInitializeX86TargetInfo();
LLVMInitializeX86Target();
LLVMInitializeX86TargetMC();
by
LLVMInitializeNativeTarget();
LLVMInitializeNativeAsmPrinter();
LLVMInitializeNativeAsmParser();
Replace the call of LLVMCreateExecutionEngineForModule
by a call to the custom function LLVMCreateExecutionEngineForModuleCPU
. It is the original implementation of LLVMCreateExecutionEngineForModule
plus a call of setMCPU
.
#define LLVM_VERSION (LLVM_VERSION_MAJOR * 100 + LLVM_VERSION_MINOR)
LLVMBool LLVMCreateExecutionEngineForModuleCPU
(LLVMExecutionEngineRef *OutEE,
LLVMModuleRef M,
char **OutError) {
std::string Error;
#if LLVM_VERSION < 306
EngineBuilder builder(unwrap(M));
#else
EngineBuilder builder(std::unique_ptr<Module>(unwrap(M)));
#endif
builder.setEngineKind(EngineKind::Either)
.setMCPU(sys::getHostCPUName().data())
.setErrorStr(&Error);
if (ExecutionEngine *EE = builder.create()){
*OutEE = wrap(EE);
return 0;
}
*OutError = strdup(Error.c_str());
return 1;
}
I should also add
float vector[vectorSize] __attribute__((aligned(32)));
in order to align the array for AVX vectors.
According to an answer in the thread crash JIT with AVX intrinsics LLVMRunFunction
is restricted to main
-like prototypes (apparently only in MCJIT). Thus we should also replace the LLVMRunFunction
stuff by
void (*funcPtr) (float *);
funcPtr = LLVMGetPointerToGlobal(execEngine, func);
funcPtr(vector);
Upvotes: 1