[ValueTracking][SVE] Fix getOffsetFromIndex for scalable vector.

[huihuiz]

Summary

Return None if GEP index type is scalable vector. Size of scalable vectors
are multiplied by a runtime constant.

Avoid transforming:

%a = bitcast i8* %p to <vscale x 16 x i8>*
%tmp0 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i64 0
store <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8>* %tmp0
%tmp1 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i64 1
store <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8>* %tmp1

into:

%a = bitcast i8* %p to <vscale x 16 x i8>*
%tmp0 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i64 0
%1 = bitcast <vscale x 16 x i8>* %tmp0 to i8*
call void @llvm.memset.p0i8.i64(i8* align 16 %1, i8 0, i64 32, i1 false)

[efriedma]

llvm/lib/Analysis/ValueTracking.cpp

6157

If we're calling getTypeAllocSize anyway, can we just check whether it returns a scalable TypeSize?

[huihuiz]

Address review comments.

llvm/lib/Analysis/ValueTracking.cpp

6157

Thanks Eli for the feedback! Yes, that sounds better.

[sdesmalen]

Thanks for this patch! Looks good to me.

llvm/lib/Analysis/ValueTracking.cpp

switch (R.Flavor) {
     case SPF_UMAX:
       Lower = *C;
       break;
     case SPF_SMIN:
       Lower = APInt::getSignedMinValue(BitWidth);
       Upper = *C + 1;
       break;
     case SPF_SMAX:
       Lower = *C;
       Upper = APInt::getSignedMaxValue(BitWidth) + 1;
       break;
     default:
       break;
   }
 }
 
 ConstantRange llvm::computeConstantRange(const Value *V, bool UseInstrInfo) {
if (OpC->isZero())
       continue; // No offset.
 
     // Handle struct indices, which add their field offset to the pointer.
     if (StructType *STy = GTI.getStructTypeOrNull()) {
       Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
       continue;
     }
 
-    // Otherwise, we have a sequential type like an array or vector.  Multiply
-    // the index by the ElementSize.
-    uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
-    Offset += Size * OpC->getSExtValue();
+    // Otherwise, we have a sequential type like an array or fixed-length
+    // vector. Multiply the index by the ElementSize.
+    TypeSize Size = DL.getTypeAllocSize(GTI.getIndexedType());
+    if (Size.isScalable())
+      return None;
+    Offset += Size.getFixedSize() * OpC->getSExtValue();
   }
 
   return Offset;
 }
 
 Optional<int64_t> llvm::isPointerOffset(const Value *Ptr1, const Value *Ptr2,
                                         const DataLayout &DL) {
   Ptr1 = Ptr1->stripPointerCasts();

llvm/test/Transforms/MemCpyOpt/vscale-memset.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -memcpyopt -dce -S | FileCheck %s
+
+; Negative test
+; Check this test is not transformed into memset, or cause a compiler warning
+; warning: Compiler has made implicit assumption that TypeSize is not scalable. This may or may not lead to broken code.
+
+define void @foo(i8* %p) {
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:    [[A:%.*]] = bitcast i8* [[P:%.*]] to <vscale x 16 x i8>*
+; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* [[A]], i64 0
+; CHECK-NEXT:    store <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8>* [[TMP0]]
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* [[A]], i64 1
+; CHECK-NEXT:    store <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8>* [[TMP1]]
+; CHECK-NEXT:    ret void
+;
+  %a = bitcast i8* %p to <vscale x 16 x i8>*
+  %tmp0 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i64 0
+  store <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8>* %tmp0
+  %tmp1 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i64 1
+  store <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8>* %tmp1
+  ret void
+}
+
+; Positive test
+
+define void @memset_vscale_index_zero(i8* %p, i8 %z) {
+; CHECK-LABEL: @memset_vscale_index_zero(
+; CHECK-NEXT:    [[A:%.*]] = bitcast i8* [[P:%.*]] to <vscale x 16 x i8>*
+; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* [[A]], i32 0, i32 0
+; CHECK-NEXT:    call void @llvm.memset.p0i8.i64(i8* align 1 [[TMP0]], i8 [[Z:%.*]], i64 17, i1 false)
+; CHECK-NEXT:    ret void
+;
+  %a = bitcast i8* %p to <vscale x 16 x i8>*
+  %tmp0 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 0
+  store i8 %z, i8* %tmp0
+  %tmp1 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 1
+  store i8 %z, i8* %tmp1
+  %tmp2 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 2
+  store i8 %z, i8* %tmp2
+  %tmp3 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 3
+  store i8 %z, i8* %tmp3
+  %tmp4 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 4
+  store i8 %z, i8* %tmp4
+  %tmp5 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 5
+  store i8 %z, i8* %tmp5
+  %tmp6 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 6
+  store i8 %z, i8* %tmp6
+  %tmp7 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 7
+  store i8 %z, i8* %tmp7
+  %tmp8 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 8
+  store i8 %z, i8* %tmp8
+  %tmp9 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 9
+  store i8 %z, i8* %tmp9
+  %tmp10 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 10
+  store i8 %z, i8* %tmp10
+  %tmp11 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 11
+  store i8 %z, i8* %tmp11
+  %tmp12 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 12
+  store i8 %z, i8* %tmp12
+  %tmp13 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 13
+  store i8 %z, i8* %tmp13
+  %tmp14 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 14
+  store i8 %z, i8* %tmp14
+  %tmp15 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 15
+  store i8 %z, i8* %tmp15
+  %tmp16 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 0, i32 16
+  store i8 %z, i8* %tmp16
+  ret void
+}
+
+define void @memset_vscale_index_nonzero(i8* %p, i8 %z) {
+; CHECK-LABEL: @memset_vscale_index_nonzero(
+; CHECK-NEXT:    [[A:%.*]] = bitcast i8* [[P:%.*]] to <vscale x 16 x i8>*
+; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* [[A]], i32 1, i32 0
+; CHECK-NEXT:    call void @llvm.memset.p0i8.i64(i8* align 1 [[TMP0]], i8 [[Z:%.*]], i64 17, i1 false)
+; CHECK-NEXT:    ret void
+;
+  %a = bitcast i8* %p to <vscale x 16 x i8>*
+  %tmp0 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 0
+  store i8 %z, i8* %tmp0
+  %tmp1 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 1
+  store i8 %z, i8* %tmp1
+  %tmp2 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 2
+  store i8 %z, i8* %tmp2
+  %tmp3 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 3
+  store i8 %z, i8* %tmp3
+  %tmp4 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 4
+  store i8 %z, i8* %tmp4
+  %tmp5 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 5
+  store i8 %z, i8* %tmp5
+  %tmp6 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 6
+  store i8 %z, i8* %tmp6
+  %tmp7 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 7
+  store i8 %z, i8* %tmp7
+  %tmp8 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 8
+  store i8 %z, i8* %tmp8
+  %tmp9 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 9
+  store i8 %z, i8* %tmp9
+  %tmp10 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 10
+  store i8 %z, i8* %tmp10
+  %tmp11 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 11
+  store i8 %z, i8* %tmp11
+  %tmp12 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 12
+  store i8 %z, i8* %tmp12
+  %tmp13 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 13
+  store i8 %z, i8* %tmp13
+  %tmp14 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 14
+  store i8 %z, i8* %tmp14
+  %tmp15 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 15
+  store i8 %z, i8* %tmp15
+  %tmp16 = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %a, i32 1, i32 16
+  store i8 %z, i8* %tmp16
+  ret void
+}