[GVN] Fix a crash bug w/non-integral pointers and memtransfers

Problem is very similiar to the one fixed for memsets in r354399, we try to coerce a value to non-integral type, and then crash while try to do so.  Since we shouldn't be doing such coercions to start with, easy fix.  From inspection, I see two other cases which look to be similiar and will follow up with most test cases and fixes if confirmed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@354403 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Utils/VNCoercion.cpp b/lib/Transforms/Utils/VNCoercion.cpp
index 8e21472f2c1b9b438dd33aaa7df6678ab4411468..1aedce75ca6a6d905876a4076846831b089a8445 100644 (file)
--- a/lib/Transforms/Utils/VNCoercion.cpp
+++ b/lib/Transforms/Utils/VNCoercion.cpp
@@ -300,6 +300,11 @@ int analyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
   if (Offset == -1)
     return Offset;
 
+  // Don't coerce non-integral pointers to integers or vice versa.
+  if (DL.isNonIntegralPointerType(LoadTy->getScalarType()))
+    // TODO: Can allow nullptrs from constant zeros
+    return -1;
+
   unsigned AS = Src->getType()->getPointerAddressSpace();
   // Otherwise, see if we can constant fold a load from the constant with the
   // offset applied as appropriate.

diff --git a/test/Transforms/GVN/non-integral-pointers.ll b/test/Transforms/GVN/non-integral-pointers.ll
index c5a281e142367556dd2537d75d8fc61b8a2530a8..fc7f74d3796d837de9e58a6a769c2f523206f3a0 100644 (file)
--- a/test/Transforms/GVN/non-integral-pointers.ll
+++ b/test/Transforms/GVN/non-integral-pointers.ll
@@ -154,3 +154,62 @@ define i8 addrspace(4)* @forward_store_zero(i8 addrspace(4)* addrspace(4)* %loc)
   %ref = load i8 addrspace(4)*, i8 addrspace(4)* addrspace(4)* %loc
   ret i8 addrspace(4)* %ref
 }
+
+@NonZeroConstant = constant <4 x i64> <i64 3, i64 3, i64 3, i64 3>
+@ZeroConstant = constant <4 x i64> zeroinitializer
+
+
+; Can't forward as the load might be dead.  (Pretend we wrote out the alwaysfalse idiom above.)
+define i8 addrspace(4)* @neg_forward_memcopy(i8 addrspace(4)* addrspace(4)* %loc) {
+; CHECK-LABEL: @neg_forward_memcopy(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[LOC_BC:%.*]] = bitcast i8 addrspace(4)* addrspace(4)* [[LOC:%.*]] to i8 addrspace(4)*
+; CHECK-NEXT:    call void @llvm.memcpy.p4i8.p0i8.i64(i8 addrspace(4)* align 4 [[LOC_BC]], i8* bitcast (<4 x i64>* @NonZeroConstant to i8*), i64 8, i1 false)
+; CHECK-NEXT:    [[REF:%.*]] = load i8 addrspace(4)*, i8 addrspace(4)* addrspace(4)* [[LOC]]
+; CHECK-NEXT:    ret i8 addrspace(4)* [[REF]]
+;
+entry:
+  %loc.bc = bitcast i8 addrspace(4)* addrspace(4)* %loc to i8 addrspace(4)*
+  %src.bc = bitcast <4 x i64>* @NonZeroConstant to i8*
+  call void @llvm.memcpy.p4i8.p0i8.i64(i8 addrspace(4)* align 4 %loc.bc, i8* %src.bc, i64 8, i1 false)
+  %ref = load i8 addrspace(4)*, i8 addrspace(4)* addrspace(4)* %loc
+  ret i8 addrspace(4)* %ref
+}
+
+define <1 x i8 addrspace(4)*> @neg_forward_memcpy_vload(<1 x i8 addrspace(4)*> addrspace(4)* %loc) {
+; CHECK-LABEL: @neg_forward_memcpy_vload(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[LOC_BC:%.*]] = bitcast <1 x i8 addrspace(4)*> addrspace(4)* [[LOC:%.*]] to i8 addrspace(4)*
+; CHECK-NEXT:    call void @llvm.memcpy.p4i8.p0i8.i64(i8 addrspace(4)* align 4 [[LOC_BC]], i8* bitcast (<4 x i64>* @NonZeroConstant to i8*), i64 8, i1 false)
+; CHECK-NEXT:    [[REF:%.*]] = load <1 x i8 addrspace(4)*>, <1 x i8 addrspace(4)*> addrspace(4)* [[LOC]]
+; CHECK-NEXT:    ret <1 x i8 addrspace(4)*> [[REF]]
+;
+entry:
+  %loc.bc = bitcast <1 x i8 addrspace(4)*> addrspace(4)* %loc to i8 addrspace(4)*
+  %src.bc = bitcast <4 x i64>* @NonZeroConstant to i8*
+  call void @llvm.memcpy.p4i8.p0i8.i64(i8 addrspace(4)* align 4 %loc.bc, i8* %src.bc, i64 8, i1 false)
+  %ref = load <1 x i8 addrspace(4)*>, <1 x i8 addrspace(4)*> addrspace(4)* %loc
+  ret <1 x i8 addrspace(4)*> %ref
+}
+
+
+; Can forward since we can do so w/o breaking types
+; TODO: missed optimization
+define i8 addrspace(4)* @forward_memcpy_zero(i8 addrspace(4)* addrspace(4)* %loc) {
+; CHECK-LABEL: @forward_memcpy_zero(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[LOC_BC:%.*]] = bitcast i8 addrspace(4)* addrspace(4)* [[LOC:%.*]] to i8 addrspace(4)*
+; CHECK-NEXT:    call void @llvm.memcpy.p4i8.p0i8.i64(i8 addrspace(4)* align 4 [[LOC_BC]], i8* bitcast (<4 x i64>* @ZeroConstant to i8*), i64 8, i1 false)
+; CHECK-NEXT:    [[REF:%.*]] = load i8 addrspace(4)*, i8 addrspace(4)* addrspace(4)* [[LOC]]
+; CHECK-NEXT:    ret i8 addrspace(4)* [[REF]]
+;
+entry:
+  %loc.bc = bitcast i8 addrspace(4)* addrspace(4)* %loc to i8 addrspace(4)*
+  %src.bc = bitcast <4 x i64>* @ZeroConstant to i8*
+  call void @llvm.memcpy.p4i8.p0i8.i64(i8 addrspace(4)* align 4 %loc.bc, i8* %src.bc, i64 8, i1 false)
+  %ref = load i8 addrspace(4)*, i8 addrspace(4)* addrspace(4)* %loc
+  ret i8 addrspace(4)* %ref
+}
+
+
+declare void @llvm.memcpy.p4i8.p0i8.i64(i8 addrspace(4)* nocapture, i8* nocapture, i64, i1) nounwind