assigned to @rotateright

Bugpoint reduction:

opt -instcombine

; ModuleID = ''
source_filename = "test.c"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

@​a = external dso_local global i32, align 4

define dso_local void @​b() local_unnamed_addr #​0 {
entry:
br label %for.cond

for.cond: ; preds = %for.body, %entry
%storemerge = phi i32 [ 7, %entry ], [ %add, %for.body ]
br i1 icmp uge (i32* inttoptr (i64 2 to i32*), i32* @​a), label %for.body, label %for.end

for.body: ; preds = %for.cond
%0 = load i32, i32* @​a, align 4
%div = sdiv i32 %0, undef
%add = add nsw i32 %div, 1
br label %for.cond

for.end: ; preds = %for.cond
%cmp6 = icmp ne i32 %storemerge, 4
%conv = zext i1 %cmp6 to i32
%and7 = and i32 %storemerge, %conv
%sub = sub nsw i32 %storemerge, %and7
%cmp9 = icmp ne i32 %sub, 4
%conv10 = zext i1 %cmp9 to i32
%sub11 = sub nsw i32 %conv10, %sub
%and14 = and i32 %sub11, 1
%1 = select i1 false, i32 0, i32 %and14
store i32 %1, i32* @​a, align 4
ret void
}

attributes #​0 = { "tune-cpu"="generic" }

!llvm.ident = !{#0}

!​0 = !{!"clang version 13.0.0"}

The bug appears to be in InstCombine's SimplifyDemandedBits of a select. We may have conflicting motives in changing a select's constant operands into the least set bits vs. matching a cmp's operand (another min/max hack that could disappear with canonicalization to intrinsics).

So we're ping-ponging between:
IC: Visiting: %sub11 = select i1 %cmp9, i32 1, i32 6
IC: Visiting: %sub11 = select i1 %cmp9, i32 1, i32 4

I think there's a small fix to avoid that - will try to reduce further and confirm.

This should fix the bug in trunk:
https://reviews.llvm.org/rG9502061bcc86

I'm marking this as blocking the 12.0 release to see if we can still get it in there.

As noted in the description, this was visible in the 11.0 release, but not before that. The bug was likely introduced with:
https://reviews.llvm.org/rGa59cc5e128f09ec

Merged: 692808e