Merge changes from topic \'fix-struct-layout\'
am: ea61366
* commit 'ea613661f50e49d9a8074a6342c2091a1bc55d3e':
Translate GEP instructions on structs to GEP on int8*
X86: Fix kernel Input/Output GetElementPtr offset issue
Change-Id: If350936ccc606dbb38cfe1fe766765aada7137b3
diff --git a/include/bcc/Compiler.h b/include/bcc/Compiler.h
index a0925b8..1097bad 100644
--- a/include/bcc/Compiler.h
+++ b/include/bcc/Compiler.h
@@ -68,7 +68,9 @@
kErrInvalidSource,
- kIllegalGlobalFunction
+ kIllegalGlobalFunction,
+
+ kErrInvalidTargetMachine
};
static const char *GetErrorString(enum ErrorCode pErrCode);
@@ -116,6 +118,8 @@
// all RenderScript functions. Returns error if any external function that is
// not in this whitelist is callable from the script.
enum ErrorCode screenGlobalFunctions(Script &pScript);
+
+ void translateGEPs(Script &pScript);
};
} // end namespace bcc
diff --git a/include/bcc/Config/Config.h b/include/bcc/Config/Config.h
index 5c209fd..8293827 100644
--- a/include/bcc/Config/Config.h
+++ b/include/bcc/Config/Config.h
@@ -81,6 +81,13 @@
#define DEFAULT_X86_TRIPLE_STRING "i686-unknown-linux"
#define DEFAULT_X86_64_TRIPLE_STRING "x86_64-unknown-linux"
+// Custom DataLayout string for X86 with i64 and f64 set to match the ARM32
+// alignment requirement of 64-bits.
+#define X86_CUSTOM_DL_STRING "e-m:e-p:32:32-i64:64-f64:64:64-f80:32-n8:16:32-S128"
+// Default DataLayout string for X86. Present to detect future LLVM datalayout
+// changes so X86_CUSTOM_DL_STRING above can be modified appropriately.
+#define X86_DEFAULT_DL_STRING "e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128"
+
#if defined(DEFAULT_ARM_CODEGEN)
#define DEFAULT_TARGET_TRIPLE_STRING DEFAULT_ARM_TRIPLE_STRING
#elif defined(DEFAULT_ARM64_CODEGEN)
diff --git a/include/bcc/Renderscript/RSTransforms.h b/include/bcc/Renderscript/RSTransforms.h
index 66353a3..04e8460 100644
--- a/include/bcc/Renderscript/RSTransforms.h
+++ b/include/bcc/Renderscript/RSTransforms.h
@@ -47,6 +47,8 @@
llvm::ModulePass * createRSAddDebugInfoPass();
+llvm::FunctionPass *createRSX86TranslateGEPPass();
+
} // end namespace bcc
#endif // BCC_RS_TRANSFORMS_H
diff --git a/lib/Core/Compiler.cpp b/lib/Core/Compiler.cpp
index fe16b71..a9f1df0 100644
--- a/lib/Core/Compiler.cpp
+++ b/lib/Core/Compiler.cpp
@@ -32,6 +32,7 @@
#include <llvm/Transforms/Vectorize.h>
#include "bcc/Assert.h"
+#include "bcc/Config/Config.h"
#include "bcc/Renderscript/RSScript.h"
#include "bcc/Renderscript/RSTransforms.h"
#include "bcc/Script.h"
@@ -74,6 +75,8 @@
return "Error loading input bitcode";
case kIllegalGlobalFunction:
return "Use of undefined external function";
+ case kErrInvalidTargetMachine:
+ return "Invalid/unexpected llvm::TargetMachine.";
}
// This assert should never be reached as the compiler verifies that the
@@ -251,6 +254,15 @@
return kErrInvalidSource;
}
+ if (getTargetMachine().getTargetTriple().getArch() == llvm::Triple::x86) {
+ // Detect and fail if TargetMachine datalayout is different than what we
+ // expect. This is to detect changes in default target layout for x86 and
+ // update X86_CUSTOM_DL_STRING in include/bcc/Config/Config.h appropriately.
+ if (dl.getStringRepresentation().compare(X86_DEFAULT_DL_STRING) != 0) {
+ return kErrInvalidTargetMachine;
+ }
+ }
+
// Sanitize module's target information.
module.setTargetTriple(getTargetMachine().getTargetTriple().str());
module.setDataLayout(getTargetMachine().createDataLayout());
@@ -453,3 +465,11 @@
return kSuccess;
}
+
+void Compiler::translateGEPs(Script &pScript) {
+ llvm::legacy::PassManager pPM;
+ pPM.add(createRSX86TranslateGEPPass());
+
+ // Materialization done in screenGlobalFunctions above.
+ pPM.run(pScript.getSource().getModule());
+}
diff --git a/lib/Renderscript/Android.mk b/lib/Renderscript/Android.mk
index b406be7..e392994 100644
--- a/lib/Renderscript/Android.mk
+++ b/lib/Renderscript/Android.mk
@@ -34,7 +34,8 @@
RSScreenFunctionsPass.cpp \
RSStubsWhiteList.cpp \
RSScriptGroupFusion.cpp \
- RSX86CallConvPass.cpp
+ RSX86CallConvPass.cpp \
+ RSX86TranslateGEPPass.cpp
#=====================================================================
# Device Static Library: libbccRenderscript
diff --git a/lib/Renderscript/RSCompilerDriver.cpp b/lib/Renderscript/RSCompilerDriver.cpp
index 0b63ea2..1dfc699 100644
--- a/lib/Renderscript/RSCompilerDriver.cpp
+++ b/lib/Renderscript/RSCompilerDriver.cpp
@@ -23,6 +23,7 @@
#include <llvm/Support/CommandLine.h>
#include <llvm/Support/Path.h>
#include <llvm/Support/raw_ostream.h>
+#include <llvm/Target/TargetMachine.h>
#include "bcinfo/BitcodeWrapper.h"
#include "bcc/Assert.h"
@@ -126,6 +127,19 @@
return Compiler::kErrInvalidSource;
}
+ // For (32-bit) x86, translate GEPs on structs or arrays of structs to GEPs on
+ // int8* with byte offsets. This is to ensure that layout of structs with
+ // 64-bit scalar fields matches frontend-generated code that adheres to ARM
+ // data layout.
+ //
+ // The translation is done before RenderScript runtime library is linked
+ // (during LinkRuntime below) to ensure that RenderScript-driver-provided
+ // structs (like Allocation_t) don't get forced into using the ARM layout
+ // rules.
+ if (mCompiler.getTargetMachine().getTargetTriple().getArch() == llvm::Triple::x86) {
+ mCompiler.translateGEPs(pScript);
+ }
+
//===--------------------------------------------------------------------===//
// Link RS script with Renderscript runtime.
//===--------------------------------------------------------------------===//
diff --git a/lib/Renderscript/RSKernelExpand.cpp b/lib/Renderscript/RSKernelExpand.cpp
index 7337a30..aa139f2 100644
--- a/lib/Renderscript/RSKernelExpand.cpp
+++ b/lib/Renderscript/RSKernelExpand.cpp
@@ -665,8 +665,9 @@
// ArgIter - iterator pointing to first input of the UNexpanded function
// NumInputs - number of inputs (NOT number of ARGUMENTS)
//
- // InBufPtrs[] - this function sets each array element to point to the first
- // cell of the corresponding input allocation
+ // InTypes[] - this function saves input type, they will be used in ExpandInputsBody().
+ // InBufPtrs[] - this function sets each array element to point to the first cell / byte
+ // (byte for x86, cell for other platforms) of the corresponding input allocation
// InStructTempSlots[] - this function sets each array element either to nullptr
// or to the result of an alloca (for the case where the
// calling convention dictates that a value must be passed
@@ -677,6 +678,7 @@
llvm::MDNode *TBAAPointer,
llvm::Function::arg_iterator ArgIter,
const size_t NumInputs,
+ llvm::SmallVectorImpl<llvm::Type *> &InTypes,
llvm::SmallVectorImpl<llvm::Value *> &InBufPtrs,
llvm::SmallVectorImpl<llvm::Value *> &InStructTempSlots) {
bccAssert(NumInputs <= RS_KERNEL_INPUT_LIMIT);
@@ -712,12 +714,25 @@
static_cast<int32_t>(InputIndex)}));
llvm::Value *InBufPtrAddr = Builder.CreateInBoundsGEP(Arg_p, InBufPtrGEP, "input_buf.gep");
llvm::LoadInst *InBufPtr = Builder.CreateLoad(InBufPtrAddr, "input_buf");
- llvm::Value *CastInBufPtr = Builder.CreatePointerCast(InBufPtr, InType, "casted_in");
+
+ llvm::Value *CastInBufPtr = nullptr;
+ if (Module->getTargetTriple() != DEFAULT_X86_TRIPLE_STRING) {
+ CastInBufPtr = Builder.CreatePointerCast(InBufPtr, InType, "casted_in");
+ } else {
+ // The disagreement between module and x86 target machine datalayout
+ // causes mismatched input/output data offset between slang reflected
+ // code and bcc codegen for GetElementPtr. To solve this issue, skip the
+ // cast to InType and leave CastInBufPtr as an int8_t*. The buffer is
+ // later indexed with an explicit byte offset computed based on
+ // X86_CUSTOM_DL_STRING and then bitcast it to actual input type.
+ CastInBufPtr = InBufPtr;
+ }
if (gEnableRsTbaa) {
InBufPtr->setMetadata("tbaa", TBAAPointer);
}
+ InTypes.push_back(InType);
InBufPtrs.push_back(CastInBufPtr);
}
@@ -732,6 +747,8 @@
// Arg_x1 - first X coordinate to be processed by the expanded function
// TBAAAllocation - metadata for marking loads of input values out of allocations
// NumInputs -- number of inputs (NOT number of ARGUMENTS)
+ // InTypes[] - this function uses the saved input types in ExpandInputsLoopInvariant()
+ // to convert the pointer of byte InPtr to its real type.
// InBufPtrs[] - this function consumes the information produced by ExpandInputsLoopInvariant()
// InStructTempSlots[] - this function consumes the information produced by ExpandInputsLoopInvariant()
// IndVar - value of loop induction variable (X coordinate) for a given loop iteration
@@ -742,16 +759,32 @@
llvm::Value *Arg_x1,
llvm::MDNode *TBAAAllocation,
const size_t NumInputs,
+ const llvm::SmallVectorImpl<llvm::Type *> &InTypes,
const llvm::SmallVectorImpl<llvm::Value *> &InBufPtrs,
const llvm::SmallVectorImpl<llvm::Value *> &InStructTempSlots,
llvm::Value *IndVar,
llvm::SmallVectorImpl<llvm::Value *> &RootArgs) {
llvm::Value *Offset = Builder.CreateSub(IndVar, Arg_x1);
+ llvm::Type *Int32Ty = llvm::Type::getInt32Ty(*Context);
for (size_t Index = 0; Index < NumInputs; ++Index) {
- llvm::Value *InPtr = Builder.CreateInBoundsGEP(InBufPtrs[Index], Offset);
- llvm::Value *Input;
+ llvm::Value *InPtr = nullptr;
+ if (Module->getTargetTriple() != DEFAULT_X86_TRIPLE_STRING) {
+ InPtr = Builder.CreateInBoundsGEP(InBufPtrs[Index], Offset);
+ } else {
+ // Treat x86 input buffer as byte[], get indexed pointer with explicit
+ // byte offset computed using a datalayout based on
+ // X86_CUSTOM_DL_STRING, then bitcast it to actual input type.
+ llvm::DataLayout DL(X86_CUSTOM_DL_STRING);
+ llvm::Type *InTy = InTypes[Index];
+ uint64_t InStep = DL.getTypeAllocSize(InTy->getPointerElementType());
+ llvm::Value *OffsetInBytes = Builder.CreateMul(Offset, llvm::ConstantInt::get(Int32Ty, InStep));
+ InPtr = Builder.CreateInBoundsGEP(InBufPtrs[Index], OffsetInBytes);
+ InPtr = Builder.CreatePointerCast(InPtr, InTy);
+ }
+
+ llvm::Value *Input;
llvm::LoadInst *InputLoad = Builder.CreateLoad(InPtr, "input");
if (gEnableRsTbaa) {
@@ -793,6 +826,9 @@
}
llvm::DataLayout DL(Module);
+ if (Module->getTargetTriple() == DEFAULT_X86_TRIPLE_STRING) {
+ DL.reset(X86_CUSTOM_DL_STRING);
+ }
llvm::Function *ExpandedFunction =
createEmptyExpandedForEachKernel(Function->getName());
@@ -923,6 +959,10 @@
// TODO: Refactor this to share functionality with ExpandOldStyleForEach.
llvm::DataLayout DL(Module);
+ if (Module->getTargetTriple() == DEFAULT_X86_TRIPLE_STRING) {
+ DL.reset(X86_CUSTOM_DL_STRING);
+ }
+ llvm::Type *Int32Ty = llvm::Type::getInt32Ty(*Context);
llvm::Function *ExpandedFunction =
createEmptyExpandedForEachKernel(Function->getName());
@@ -1003,9 +1043,20 @@
OutBasePtr->setMetadata("tbaa", TBAAPointer);
}
- CastedOutBasePtr = Builder.CreatePointerCast(OutBasePtr, OutTy, "casted_out");
+ if (Module->getTargetTriple() != DEFAULT_X86_TRIPLE_STRING) {
+ CastedOutBasePtr = Builder.CreatePointerCast(OutBasePtr, OutTy, "casted_out");
+ } else {
+ // The disagreement between module and x86 target machine datalayout
+ // causes mismatched input/output data offset between slang reflected
+ // code and bcc codegen for GetElementPtr. To solve this issue, skip the
+ // cast to OutTy and leave CastedOutBasePtr as an int8_t*. The buffer
+ // is later indexed with an explicit byte offset computed based on
+ // X86_CUSTOM_DL_STRING and then bitcast it to actual output type.
+ CastedOutBasePtr = OutBasePtr;
+ }
}
+ llvm::SmallVector<llvm::Type*, 8> InTypes;
llvm::SmallVector<llvm::Value*, 8> InBufPtrs;
llvm::SmallVector<llvm::Value*, 8> InStructTempSlots;
@@ -1030,7 +1081,7 @@
if (NumInPtrArguments > 0) {
ExpandInputsLoopInvariant(Builder, LoopHeader, Arg_p, TBAAPointer, ArgIter, NumInPtrArguments,
- InBufPtrs, InStructTempSlots);
+ InTypes, InBufPtrs, InStructTempSlots);
}
// Populate the actual call to kernel().
@@ -1043,7 +1094,18 @@
llvm::Value *OutPtr = nullptr;
if (CastedOutBasePtr) {
llvm::Value *OutOffset = Builder.CreateSub(IV, Arg_x1);
- OutPtr = Builder.CreateInBoundsGEP(CastedOutBasePtr, OutOffset);
+
+ if (Module->getTargetTriple() != DEFAULT_X86_TRIPLE_STRING) {
+ OutPtr = Builder.CreateInBoundsGEP(CastedOutBasePtr, OutOffset);
+ } else {
+ // Treat x86 output buffer as byte[], get indexed pointer with explicit
+ // byte offset computed using a datalayout based on
+ // X86_CUSTOM_DL_STRING, then bitcast it to actual output type.
+ uint64_t OutStep = DL.getTypeAllocSize(OutTy->getPointerElementType());
+ llvm::Value *OutOffsetInBytes = Builder.CreateMul(OutOffset, llvm::ConstantInt::get(Int32Ty, OutStep));
+ OutPtr = Builder.CreateInBoundsGEP(CastedOutBasePtr, OutOffsetInBytes);
+ OutPtr = Builder.CreatePointerCast(OutPtr, OutTy);
+ }
if (PassOutByPointer) {
RootArgs.push_back(OutPtr);
@@ -1054,7 +1116,7 @@
if (NumInPtrArguments > 0) {
ExpandInputsBody(Builder, Arg_x1, TBAAAllocation, NumInPtrArguments,
- InBufPtrs, InStructTempSlots, IV, RootArgs);
+ InTypes, InBufPtrs, InStructTempSlots, IV, RootArgs);
}
finishArgList(RootArgs, CalleeArgs, CalleeArgsContextIdx, *Function, Builder);
@@ -1141,6 +1203,9 @@
ALOGV("Expanding simple reduce kernel %s", Function->getName().str().c_str());
llvm::DataLayout DL(Module);
+ if (Module->getTargetTriple() == DEFAULT_X86_TRIPLE_STRING) {
+ DL.reset(X86_CUSTOM_DL_STRING);
+ }
// TBAA Metadata
llvm::MDNode *TBAARenderScriptDistinct, *TBAARenderScript, *TBAAAllocation;
@@ -1431,15 +1496,16 @@
ExpandSpecialArguments(Signature, IndVar, Arg_p, Builder, CalleeArgs,
[](){}, LoopHeader->getTerminator());
+ llvm::SmallVector<llvm::Type*, 8> InTypes;
llvm::SmallVector<llvm::Value*, 8> InBufPtrs;
llvm::SmallVector<llvm::Value*, 8> InStructTempSlots;
ExpandInputsLoopInvariant(Builder, LoopHeader, Arg_p, TBAAPointer, AccumulatorArgIter, NumInputs,
- InBufPtrs, InStructTempSlots);
+ InTypes, InBufPtrs, InStructTempSlots);
// Populate the actual call to the original accumulator.
llvm::SmallVector<llvm::Value*, 8> RootArgs;
RootArgs.push_back(Arg_accum);
- ExpandInputsBody(Builder, Arg_x1, TBAAAllocation, NumInputs, InBufPtrs, InStructTempSlots,
+ ExpandInputsBody(Builder, Arg_x1, TBAAAllocation, NumInputs, InTypes, InBufPtrs, InStructTempSlots,
IndVar, RootArgs);
finishArgList(RootArgs, CalleeArgs, CalleeArgsContextIdx, *FnAccumulator, Builder);
Builder.CreateCall(FnAccumulator, RootArgs);
diff --git a/lib/Renderscript/RSX86TranslateGEPPass.cpp b/lib/Renderscript/RSX86TranslateGEPPass.cpp
new file mode 100644
index 0000000..75fc2ed
--- /dev/null
+++ b/lib/Renderscript/RSX86TranslateGEPPass.cpp
@@ -0,0 +1,181 @@
+/*
+ * Copyright 2016, The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "bcc/Assert.h"
+#include "bcc/Config/Config.h"
+#include "bcc/Support/Log.h"
+#include "bcc/Renderscript/RSTransforms.h"
+
+#include <cstdlib>
+
+#include <llvm/IR/Function.h>
+#include <llvm/IR/Instructions.h>
+#include <llvm/IR/Module.h>
+#include <llvm/Pass.h>
+#include <llvm/IR/GetElementPtrTypeIterator.h>
+
+namespace { // anonymous namespace
+
+/* This pass translates GEPs that index into structs or arrays of structs to
+ * GEPs with an int8* operand and a byte offset. This translation is done to
+ * enforce on x86 the ARM alignment rule that 64-bit scalars be 8-byte aligned
+ * for structs with such scalars.
+ */
+class RSX86TranslateGEPPass : public llvm::FunctionPass {
+private:
+ static char ID;
+ llvm::LLVMContext *Context;
+ const llvm::DataLayout DL;
+
+ // Walk a GEP instruction and return true if any type indexed is a struct.
+ bool GEPIndexesStructType(const llvm::GetElementPtrInst *GEP) {
+ for (llvm::gep_type_iterator GTI = gep_type_begin(GEP),
+ GTE = gep_type_end(GEP);
+ GTI != GTE; ++GTI) {
+ if (llvm::dyn_cast<llvm::StructType>(*GTI)) {
+ return true;
+ }
+ }
+ return false;
+ }
+
+ // Helper method to add two llvm::Value parameters
+ llvm::Value *incrementOffset(llvm::Value *accum, llvm::Value *incr,
+ llvm::Instruction *InsertBefore) {
+ if (accum == nullptr)
+ return incr;
+ return llvm::BinaryOperator::CreateAdd(accum, incr, "", InsertBefore);
+ }
+
+ // Compute the byte offset for a GEP from the GEP's base pointer operand.
+ // Based on visitGetElementPtrInst in llvm/lib/Transforms/Scalar/SROA.cpp.
+ // The difference is that this function handles non-constant array indices and
+ // constructs a sequence of instructions to calculate the offset. These
+ // instructions might not be the most efficient way to calculate this offset,
+ // but we rely on subsequent optimizations to do necessary fold/combine.
+ llvm::Value *computeGEPOffset(llvm::GetElementPtrInst *GEP) {
+ llvm::Value *Offset = nullptr;
+
+ for (llvm::gep_type_iterator GTI = gep_type_begin(GEP),
+ GTE = gep_type_end(GEP);
+ GTI != GTE; ++GTI) {
+ if (llvm::StructType *STy = llvm::dyn_cast<llvm::StructType>(*GTI)) {
+ llvm::ConstantInt *OpC = llvm::dyn_cast<llvm::ConstantInt>(GTI.getOperand());
+ if (!OpC) {
+ ALOGE("Operand for struct type is not constant!");
+ bccAssert(false);
+ }
+
+ // Offset = Offset + EltOffset for index into a struct
+ const llvm::StructLayout *SL = DL.getStructLayout(STy);
+ unsigned EltOffset = SL->getElementOffset(OpC->getZExtValue());
+ llvm::Value *Incr = llvm::ConstantInt::get(
+ llvm::Type::getInt32Ty(*Context), EltOffset);
+ Offset = incrementOffset(Offset, Incr, GEP);
+ } else {
+ // Offset = Offset + Index * EltSize for index into an array or a vector
+ llvm::Value *EltSize = llvm::ConstantInt::get(
+ llvm::Type::getInt32Ty(*Context),
+ DL.getTypeAllocSize(GTI.getIndexedType()));
+ llvm::Value *Incr = llvm::BinaryOperator::CreateMul(
+ GTI.getOperand() /* Index */,
+ EltSize, "", GEP);
+ Offset = incrementOffset(Offset, Incr, GEP);
+ }
+ }
+
+ return Offset;
+ }
+
+ void translateGEP(llvm::GetElementPtrInst *GEP) {
+ // cast GEP pointer operand to int8*
+ llvm::CastInst *Int8Ptr = llvm::CastInst::CreatePointerCast(
+ GEP->getPointerOperand(),
+ llvm::Type::getInt8PtrTy(*Context),
+ "to.int8ptr",
+ GEP);
+ llvm::Value *Indices[1] = {computeGEPOffset(GEP)};
+
+ // index into the int8* based on the byte offset
+ llvm::GetElementPtrInst *Int8PtrGEP = llvm::GetElementPtrInst::Create(
+ llvm::Type::getInt8Ty(*Context), Int8Ptr, llvm::makeArrayRef(Indices),
+ "int8ptr.indexed", GEP);
+ Int8PtrGEP->setIsInBounds(GEP->isInBounds());
+
+ // cast the indexed int8* back to the type of the original GEP
+ llvm::CastInst *OutCast = llvm::CastInst::CreatePointerCast(
+ Int8PtrGEP, GEP->getType(), "to.orig.geptype", GEP);
+
+ GEP->replaceAllUsesWith(OutCast);
+ }
+
+public:
+ RSX86TranslateGEPPass()
+ : FunctionPass (ID), DL(X86_CUSTOM_DL_STRING) {
+ }
+
+ virtual void getAnalysisUsage(llvm::AnalysisUsage &AU) const override {
+ // This pass is run in isolation in a separate pass manager. So setting
+ // AnalysisUsage is unnecessary. Set just for completeness.
+ AU.setPreservesCFG();
+ }
+
+ virtual bool runOnFunction(llvm::Function &F) override {
+ bool changed = false;
+ Context = &F.getParent()->getContext();
+
+ // To avoid updating/deleting instructions while walking a BasicBlock's instructions,
+ // collect the GEPs that need to be translated and process them
+ // subsequently.
+ std::vector<llvm::GetElementPtrInst *> GEPsToHandle;
+
+ for (auto &BB: F) {
+ for (auto &I: BB) {
+ if (auto *GEP = llvm::dyn_cast<llvm::GetElementPtrInst>(&I)) {
+ if (GEPIndexesStructType(GEP)) {
+ GEPsToHandle.push_back(GEP);
+ }
+ }
+ }
+ }
+
+ for (auto *GEP: GEPsToHandle) {
+ // Translate GEPs and erase them
+ translateGEP(GEP);
+ changed = true;
+ GEP->eraseFromParent();
+ }
+
+ return changed;
+ }
+
+ virtual const char *getPassName() const override {
+ return "Translate GEPs on structs, intended for x86 target";
+ }
+};
+
+}
+
+char RSX86TranslateGEPPass::ID = 0;
+
+namespace bcc {
+
+llvm::FunctionPass *
+createRSX86TranslateGEPPass() {
+ return new RSX86TranslateGEPPass();
+}
+
+}