[WebAssembly] Disable offset folding for function addresses

[aheejin]

Summary

Wasm does not support function addresses with offsets, but isel can
generate folded SDValues in the form of (@func + offset) without this
patch.

Fixes https://bugs.llvm.org/show_bug.cgi?id=43133.

[aheejin]

Change target triple to wasm32-unknown-unknown (Doesn't matter here though)

[sbc100]

I love how precise this patch is. Does exactly what we need.

llvm/test/CodeGen/WebAssembly/function-pointer-offset.ll

19 ↗

(On Diff #365888)

Seems odd not to have any checks here other than the successful compile.. but maybe just the fact that we don't see the fatal error is enough? Perhaps mention that if folding did occur a fatal error would be generated in the backend?

[aheejin]

Add CHECK lines

[aheejin]

llvm/test/CodeGen/WebAssembly/function-pointer-offset.ll

19 ↗

(On Diff #365888)

Yeah this shouldn't crash. But added some CHECK lines too, and added a description that this shouldn't crash in MCInstLower.

[sbc100]

Maybe add Fixes: https://bugs.llvm.org/show_bug.cgi?id=43133 to the description?

lgtm but maybe wait for final review from @dschuff

[aheejin]

More comment

[aheejin]

function-pointer-offset.ll -> function-offset.ll

[aheejin]

function-offset.ll -> function-addr-offset.ll

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.h

private:
   bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
                              unsigned AS,
                              Instruction *I = nullptr) const override;
   bool allowsMisalignedMemoryAccesses(EVT, unsigned AddrSpace, Align Alignment,
                                       MachineMemOperand::Flags Flags,
                                       bool *Fast) const override;
   bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
   bool isVectorLoadExtDesirable(SDValue ExtVal) const override;
+  bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
                          EVT VT) const override;
   bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
                           MachineFunction &MF,
                           unsigned Intrinsic) const override;
 
   void computeKnownBitsForTargetNode(const SDValue Op, KnownBits &Known,
                                      const APInt &DemandedElts,

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

 bool WebAssemblyTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
   EVT ExtT = ExtVal.getValueType();
   EVT MemT = cast<LoadSDNode>(ExtVal->getOperand(0))->getValueType(0);
   return (ExtT == MVT::v8i16 && MemT == MVT::v8i8) ||
          (ExtT == MVT::v4i32 && MemT == MVT::v4i16) ||
          (ExtT == MVT::v2i64 && MemT == MVT::v2i32);
 }
 
+bool WebAssemblyTargetLowering::isOffsetFoldingLegal(
+    const GlobalAddressSDNode *GA) const {
+  // Wasm doesn't support function addresses with offsets
+  const GlobalValue *GV = GA->getGlobal();
+  return isa<Function>(GV) ? false : TargetLowering::isOffsetFoldingLegal(GA);
+}
+
 EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL,
                                                   LLVMContext &C,
                                                   EVT VT) const {
   if (VT.isVector())
     return VT.changeVectorElementTypeToInteger();
 
   // So far, all branch instructions in Wasm take an I32 condition.
   // The default TargetLowering::getSetCCResultType returns the pointer size,

llvm/test/CodeGen/WebAssembly/function-addr-offset.ll

+; RUN: llc < %s -verify-machineinstrs | FileCheck %s
+
+; Wasm does not currently support function addresses with offsets, so we
+; shouldn't try to create a folded SDNode like (function + offset). This is a
+; regression test for the folding bug and this should not crash in MCInstLower.
+
+target triple = "wasm32-unknown-unknown"
+
+; 'hidden' here should be present to reproduce the bug
+declare hidden void @ham(i8*)
+
+define void @bar(i8* %ptr) {
+bb1:
+  br i1 undef, label %bb3, label %bb2
+
+bb2:
+  ; While lowering this switch, isel creates (@ham + 1) expression as a course
+  ; of range optimization for switch, and tries to fold the expression, but
+  ; wasm does not support with function addresses with offsets. This folding
+  ; should be disabled.
+  ; CHECK:      i32.const  ham
+  ; CHECK-NEXT: i32.const  1
+  ; CHECK-NEXT: i32.add
+  switch i32 ptrtoint (void (i8*)* @ham to i32), label %bb4 [
+    i32 -1, label %bb3
+    i32 0, label %bb3
+  ]
+
+bb3:
+  unreachable
+
+bb4:
+  %tmp = load i8, i8* %ptr
+  unreachable
+}