[ConstantFold][SVE] Fix constant fold for FoldReinterpretLoadFromConstPtr.

[huihuiz]

Summary

Bail out early for scalable vectors. As global variables are not expected
to be scalable.

Use explicit call of getFixedSize() to assert on places where scalable size
doesn't make sense.

[huihuiz]

Current upstream crash with : include/llvm/Support/TypeSize.h:126: uint64_t llvm::TypeSize::getFixedSize() const: Assertion `!IsScalable && "Request for a fixed size on a scalable object"' failed.

test.ll

define <vscale x 2 x double> @load() {
  %r = load <vscale x 2 x double>, <vscale x 2 x double>* getelementptr (<vscale x 2 x double>, <vscale x 2 x double>* null, i64 1)
  ret <vscale x 2 x double> %r
}

run: opt -S -constprop test.ll -o -

[sdesmalen]

llvm/lib/Analysis/ConstantFolding.cpp

522

It seems a bit arbitrary to special case this condition, given that the remainder of the function also doesn't work for scalable vectors (scalable vectors cannot be global variables according to the LangRef).

It probably makes more sense to bail out at the beginning of this function and to query the fixed-size explicit using getTypeSizeInBits().getFixed().

[huihuiz]

Thanks Sander for the review!

[sdesmalen]

Thanks, LGTM!

llvm/lib/Analysis/ConstantFolding.cpp

   }
 
   // Otherwise, unknown initializer type.
   return false;
 }
 
 Constant *FoldReinterpretLoadFromConstPtr(Constant *C, Type *LoadTy,
                                           const DataLayout &DL) {
+  // Bail out early. Not expect to load from scalable global variable.
+  if (LoadTy->isVectorTy() && LoadTy->getVectorIsScalable())
+    return nullptr;
+
   auto *PTy = cast<PointerType>(C->getType());
   auto *IntType = dyn_cast<IntegerType>(LoadTy);
 
   // If this isn't an integer load we can't fold it directly.
   if (!IntType) {
     unsigned AS = PTy->getAddressSpace();
 
     // If this is a float/double load, we can try folding it as an int32/64 load
     // and then bitcast the result.  This can be useful for union cases.  Note
     // that address spaces don't matter here since we're not going to result in
     // an actual new load.
     Type *MapTy;
     if (LoadTy->isHalfTy())
       MapTy = Type::getInt16Ty(C->getContext());
     else if (LoadTy->isFloatTy())
       MapTy = Type::getInt32Ty(C->getContext());
     else if (LoadTy->isDoubleTy())
       MapTy = Type::getInt64Ty(C->getContext());
     else if (LoadTy->isVectorTy()) {
-      MapTy = PointerType::getIntNTy(C->getContext(),
-                                     DL.getTypeSizeInBits(LoadTy));
+      MapTy = PointerType::getIntNTy(
+          C->getContext(), DL.getTypeSizeInBits(LoadTy).getFixedSize());
     } else
       return nullptr;
 
     C = FoldBitCast(C, MapTy->getPointerTo(AS), DL);
     if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, MapTy, DL)) {
       if (Res->isNullValue() && !LoadTy->isX86_MMXTy())
         // Materializing a zero can be done trivially without a bitcast
         return Constant::getNullValue(LoadTy);

     return nullptr;
 
   auto *GV = dyn_cast<GlobalVariable>(GVal);
   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() ||
       !GV->getInitializer()->getType()->isSized())
     return nullptr;
 
   int64_t Offset = OffsetAI.getSExtValue();
-  int64_t InitializerSize = DL.getTypeAllocSize(GV->getInitializer()->getType());
+  int64_t InitializerSize =
+      DL.getTypeAllocSize(GV->getInitializer()->getType()).getFixedSize();
 
   // If we're not accessing anything in this constant, the result is undefined.
   if (Offset <= -1 * static_cast<int64_t>(BytesLoaded))
     return UndefValue::get(IntType);
 
   // If we're not accessing anything in this constant, the result is undefined.
   if (Offset >= InitializerSize)
     return UndefValue::get(IntType);

llvm/test/Analysis/ConstantFolding/vscale.ll

 define <vscale x 4 x i32> @shufflevector() {
 ; CHECK-LABEL: @shufflevector(
 ; CHECK-NEXT:    ret <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> undef, i32 1, i32 0), <vscale x 4 x i32> undef, <vscale x 4 x i32> zeroinitializer)
 ;
   %i = insertelement <vscale x 4 x i32> undef, i32 1, i32 0
   %i2 = shufflevector <vscale x 4 x i32> %i, <vscale x 4 x i32> undef, <vscale x 4 x i32> zeroinitializer
   ret <vscale x 4 x i32> %i2
 }
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Memory Access and Addressing Operations
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+define <vscale x 2 x double> @load() {
+; CHECK-LABEL: @load(
+; CHECK-NEXT:    [[R:%.*]] = load <vscale x 2 x double>, <vscale x 2 x double>* getelementptr (<vscale x 2 x double>, <vscale x 2 x double>* null, i64 1)
+; CHECK-NEXT:    ret <vscale x 2 x double> [[R]]
+;
+  %r = load <vscale x 2 x double>, <vscale x 2 x double>* getelementptr (<vscale x 2 x double>, <vscale x 2 x double>* null, i64 1)
+  ret <vscale x 2 x double> %r
+}
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; Conversion Operations
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 define <vscale x 4 x float> @bitcast() {
 ; CHECK-LABEL: @bitcast(
 ; CHECK-NEXT:    ret <vscale x 4 x float> bitcast (<vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> undef, i32 1, i32 0), <vscale x 4 x i32> undef, <vscale x 4 x i32> zeroinitializer) to <vscale x 4 x float>)
 ;