Question

以下冗长的C程序生成一个简单的LLVM模块，其中包含仅调用llvm.x86.sse41.round.ps的函数。它会发出bitcode文件，然后运行LLVM生成的代码。我的问题是如何找到目标三元组和指令扩展，如主机的SSE或AVX，以及如何将此信息添加到LLVM模块或如何将其告知LLVM执行引擎。这就是我的所作所为：

$ 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!

总结：使用LLVM-3.4，该示例有效，LLVM-3.5无法找到内部函数round.ps，LLVM-3.6及更高版本说明了我不理解的MC发射。

据我所知，LLVM-3.5找不到round.ps内在函数，我猜它找不到它，因为我还没有告诉它现有的SSE扩展。运行llc时，我可以添加选项-mattr=sse4.1，但如何将其告知执行引擎？

第二个问题：如何通过LLVM-C API了解主机的SSE等可用指令扩展？在x86上，我可以调用CPUID指令，但是有一种方法可以在所有平台上统一工作，并且LLVM可以帮助检测扩展吗？

第三个问题：我已将目标三元组硬编码到C代码中。如何通过LLVM-C API找出主机目标三元组？

最后一个问题：MC发射误差怎么样？

Answer 1

经过多次尝试后，我认为答案如下：

替换

行

LLVMInitializeX86TargetInfo();
LLVMInitializeX86Target();
LLVMInitializeX86TargetMC();

通过

LLVMInitializeNativeTarget();
LLVMInitializeNativeAsmPrinter();
LLVMInitializeNativeAsmParser();

通过调用自定义函数LLVMCreateExecutionEngineForModule来替换LLVMCreateExecutionEngineForModuleCPU的来电。这是LLVMCreateExecutionEngineForModule的原始实现，加上setMCPU的调用。

#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;
}

我还应该添加

float vector[vectorSize] __attribute__((aligned(32)));

为了对齐AVX向量的数组。

根据帖子中的答案crash JIT with AVX intrinsics LLVMRunFunction仅限于main - 就像原型一样（显然只在MCJIT中）。因此，我们还应该用

替换LLVMRunFunction内容

void (*funcPtr) (float *);
funcPtr = LLVMGetPointerToGlobal(execEngine, func);
funcPtr(vector);

在LLVM-C API中确定和设置主机目标三元组和指令扩展

1 个答案: