GetElementPtrInst *Res =
GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
Res->setIsInBounds(GEP.isInBounds());
- return Res;
+ if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace())
+ return Res;
+ // Insert Res, and create an addrspacecast.
+ // e.g.,
+ // GEP (addrspacecast i8 addrspace(1)* X to [0 x i8]*), i32 0, ...
+ // ->
+ // %0 = GEP i8 addrspace(1)* X, ...
+ // addrspacecast i8 addrspace(1)* %0 to i8*
+ return new AddrSpaceCastInst(Builder->Insert(Res), GEP.getType());
}
if (ArrayType *XATy =
// to an array of the same type as the destination pointer
// array. Because the array type is never stepped over (there
// is a leading zero) we can fold the cast into this GEP.
- GEP.setOperand(0, StrippedPtr);
- return &GEP;
+ if (StrippedPtrTy->getAddressSpace() == GEP.getAddressSpace()) {
+ GEP.setOperand(0, StrippedPtr);
+ return &GEP;
+ }
+ // Cannot replace the base pointer directly because StrippedPtr's
+ // address space is different. Instead, create a new GEP followed by
+ // an addrspacecast.
+ // e.g.,
+ // GEP (addrspacecast [10 x i8] addrspace(1)* X to [0 x i8]*),
+ // i32 0, ...
+ // ->
+ // %0 = GEP [10 x i8] addrspace(1)* X, ...
+ // addrspacecast i8 addrspace(1)* %0 to i8*
+ SmallVector<Value*, 8> Idx(GEP.idx_begin(), GEP.idx_end());
+ Value *NewGEP = GEP.isInBounds() ?
+ Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
+ Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
+ return new AddrSpaceCastInst(NewGEP, GEP.getType());
}
}
}
-; RUN: opt < %s -instcombine -S
+; RUN: opt < %s -instcombine -S | FileCheck %s
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
target triple = "x86_64-pc-win32"
ret void
}
+@array = internal addrspace(3) global [256 x float] zeroinitializer, align 4
+@scalar = internal addrspace(3) global float 0.000000e+00, align 4
+
+define void @keep_necessary_addrspacecast(i64 %i, float** %out0, float** %out1) {
+entry:
+; CHECK-LABEL: @keep_necessary_addrspacecast
+ %0 = getelementptr [256 x float]* addrspacecast ([256 x float] addrspace(3)* @array to [256 x float]*), i64 0, i64 %i
+; CHECK: addrspacecast float addrspace(3)* %{{[0-9]+}} to float*
+ %1 = getelementptr [0 x float]* addrspacecast (float addrspace(3)* @scalar to [0 x float]*), i64 0, i64 %i
+; CHECK: addrspacecast float addrspace(3)* %{{[0-9]+}} to float*
+ store float* %0, float** %out0, align 4
+ store float* %1, float** %out1, align 4
+ ret void
+}
+