[SVE] Reject vector struct indexes for scalable vector GEPs

[efriedma]

Summary

It's messy to pattern-match, and completely unnecessary: scalar indexes work equally well.

See also discussion on D81620 and D82061.

[sdesmalen]

LGTM

[ctetreau]

The language reference needs to get updated to account for this change. I think it's fine if that happens in a separate commit.

llvm/lib/IR/Type.cpp

Type *StructType::getTypeAtIndex(const Value *V) const {
   return getElementType(Idx);
 }
 
 bool StructType::indexValid(const Value *V) const {
   // Structure indexes require (vectors of) 32-bit integer constants.  In the
   // vector case all of the indices must be equal.
   if (!V->getType()->isIntOrIntVectorTy(32))
     return false;
+  if (isa<ScalableVectorType>(V->getType()))
+    return false;
   const Constant *C = dyn_cast<Constant>(V);
   if (C && V->getType()->isVectorTy())
     C = C->getSplatValue();
   const ConstantInt *CU = dyn_cast_or_null<ConstantInt>(C);
   return CU && CU->getZExtValue() < getNumElements();
 }
 
 //===----------------------------------------------------------------------===//

llvm/test/Assembler/getelementptr_vscale_struct.ll

+; RUN: not llvm-as < %s >/dev/null 2> %t
+; RUN: FileCheck %s < %t
+; Test that a scalable vector struct index is rejected.
+
+; CHECK: invalid getelementptr indices
+
+define <vscale x 2 x i32*> @test7(<vscale x 2 x {i32, i32}*> %a) {
+  %w = getelementptr {i32, i32}, <vscale x 2 x {i32, i32}*> %a, <vscale x 2 x i32> zeroinitializer, <vscale x 2 x i32> zeroinitializer
+  ret <vscale x 2 x i32*> %w
+}