miscompile by arm64 backend

[regehr]

this function should return 1:

define i32 @f() {
  %1 = sub i8 -66, 0
  %2 = icmp ugt i8 -31, %1
  %3 = select i1 %2, i32 1, i32 0
  ret i32 %3
}

however, the arm64 backend (in LLVM 14 and also top of tree) has it returning 0:

regehr@john-home:~$ ~/llvm-project/for-alive/bin/llc foo.ll -o - -march=aarch64
	.text
	.file	"foo.ll"
	.globl	f                               // -- Begin function f
	.p2align	2
	.type	f,@function
f:                                      // @f
	.cfi_startproc
// %bb.0:
	mov	w0, wzr
	ret
.Lfunc_end0:
	.size	f, .Lfunc_end0-f
	.cfi_endproc
                                        // -- End function
	.section	".note.GNU-stack","",@progbits
regehr@john-home:~$ 

cc @ornata @nunoplopes @ryan-berger @nbushehri @zhengyang92 @aqjune @Hatsunespica

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[topperc]

@sparker-arm looks like the TypePromotion pass.

[topperc]

I think PromoteTree shouldn't sext the first operand for SafeWrap instructions. Something like this

diff --git a/llvm/lib/CodeGen/TypePromotion.cpp b/llvm/lib/CodeGen/TypePromotion.cpp
index 48454a5047dd..0caf763c7ff0 100644
--- a/llvm/lib/CodeGen/TypePromotion.cpp
+++ b/llvm/lib/CodeGen/TypePromotion.cpp
@@ -484,7 +484,7 @@ void IRPromoter::PromoteTree() {
         continue;
 
       if (auto *Const = dyn_cast<ConstantInt>(Op)) {
-        Constant *NewConst = SafeWrap.contains(I)
+        Constant *NewConst = (SafeWrap.contains(I) && i == 1)
                                  ? ConstantExpr::getSExt(Const, ExtTy)
                                  : ConstantExpr::getZExt(Const, ExtTy);
         I->setOperand(i, NewConst);

[davemgreen]

That sounds good to me, I don't think the pass handles non-canonical IR very well at the moment.

I can put together a patch, if you don't have one already.