[AArch64][FastISel] Fallback on atomic stlr/cas with non-reg operands.

This has been a latent bug for almost 10 years, but is relatively
hard to trigger, needing an address operand that isn't handled
by getRegForValue (in the test here, constexpr casts).
When that happens, it returns 0, which FastISel happily uses
as a register operand, all the way to asm, where we either get
a crash on an invalid register, or a silently corrupt instruction.

Unfortunately, FastISel is still enabled at -O0 for at least
ILP32/arm64_32.

Author: ahmedbougacha

llvm/lib/Target/AArch64/AArch64FastISel.cpp

@@ -2190,6 +2190,8 @@ bool AArch64FastISel::selectStore(const Instruction *I) {
     if (isReleaseOrStronger(Ord)) {
       // The STLR addressing mode only supports a base reg; pass that directly.
       Register AddrReg = getRegForValue(PtrV);
+      if (!AddrReg)
+        return false;
       return emitStoreRelease(VT, SrcReg, AddrReg,
                               createMachineMemOperandFor(I));
     }
@@ -5070,12 +5072,16 @@ bool AArch64FastISel::selectAtomicCmpXchg(const AtomicCmpXchgInst *I) {
 
   const MCInstrDesc &II = TII.get(Opc);
 
-  const Register AddrReg = constrainOperandRegClass(
-      II, getRegForValue(I->getPointerOperand()), II.getNumDefs());
-  const Register DesiredReg = constrainOperandRegClass(
-      II, getRegForValue(I->getCompareOperand()), II.getNumDefs() + 1);
-  const Register NewReg = constrainOperandRegClass(
-      II, getRegForValue(I->getNewValOperand()), II.getNumDefs() + 2);
+  Register AddrReg = getRegForValue(I->getPointerOperand());
+  Register DesiredReg = getRegForValue(I->getCompareOperand());
+  Register NewReg = getRegForValue(I->getNewValOperand());
+
+  if (!AddrReg || !DesiredReg || !NewReg)
+    return false;
+
+  AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
+  DesiredReg = constrainOperandRegClass(II, DesiredReg, II.getNumDefs() + 1);
+  NewReg = constrainOperandRegClass(II, NewReg, II.getNumDefs() + 2);
 
   const Register ResultReg1 = createResultReg(ResRC);
   const Register ResultReg2 = createResultReg(&AArch64::GPR32RegClass);

llvm/test/CodeGen/AArch64/fast-isel-atomic-fallback.ll

@@ -0,0 +1,53 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=arm64_32-apple-darwin -O0 -fast-isel -verify-machineinstrs \
+; RUN:   -aarch64-enable-atomic-cfg-tidy=0 -aarch64-enable-collect-loh=0 \
+; RUN:   < %s | FileCheck %s
+
+; FastISel doesn't support cstexprs as operands here, but make
+; sure it knows to fallback, at least.
+
+define void @atomic_store_cstexpr_addr(i32 %val) #0 {
+; CHECK-LABEL: atomic_store_cstexpr_addr:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    adrp x8, _g@PAGE
+; CHECK-NEXT:    add x8, x8, _g@PAGEOFF
+; CHECK-NEXT:    ; kill: def $w1 killed $w8 killed $x8
+; CHECK-NEXT:    adrp x8, _g@PAGE
+; CHECK-NEXT:    add x8, x8, _g@PAGEOFF
+; CHECK-NEXT:    stlr w0, [x8]
+; CHECK-NEXT:    ret
+  store atomic i32 %val, ptr inttoptr (i32 ptrtoint (ptr @g to i32) to ptr) release, align 4
+  ret void
+}
+
+define i32 @cmpxchg_cstexpr_addr(i32 %cmp, i32 %new, ptr %ps) #0 {
+; CHECK-LABEL: cmpxchg_cstexpr_addr:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    adrp x9, _g@PAGE
+; CHECK-NEXT:    add x9, x9, _g@PAGEOFF
+; CHECK-NEXT:    ; kill: def $w0 killed $w9 killed $x9
+; CHECK-NEXT:    adrp x10, _g@PAGE
+; CHECK-NEXT:    add x10, x10, _g@PAGEOFF
+; CHECK-NEXT:  LBB1_1: ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    ldaxr w0, [x10]
+; CHECK-NEXT:    cmp w0, w8
+; CHECK-NEXT:    b.ne LBB1_3
+; CHECK-NEXT:  ; %bb.2: ; in Loop: Header=BB1_1 Depth=1
+; CHECK-NEXT:    stlxr w9, w1, [x10]
+; CHECK-NEXT:    cbnz w9, LBB1_1
+; CHECK-NEXT:  LBB1_3:
+; CHECK-NEXT:    subs w8, w0, w8
+; CHECK-NEXT:    cset w8, eq
+; CHECK-NEXT:    and w8, w8, #0x1
+; CHECK-NEXT:    str w8, [x2]
+; CHECK-NEXT:    ret
+  %tmp0 = cmpxchg ptr inttoptr (i32 ptrtoint (ptr @g to i32) to ptr), i32 %cmp, i32 %new seq_cst seq_cst
+  %tmp1 = extractvalue { i32, i1 } %tmp0, 0
+  %tmp2 = extractvalue { i32, i1 } %tmp0, 1
+  %tmp3 = zext i1 %tmp2 to i32
+  store i32 %tmp3, ptr %ps
+  ret i32 %tmp1
+}
+
+@g = global i32 0