Preserving nsw/nuw for x-(-A)=>x+A is unsafe

[llvmbot]

Bugzilla Link	20189
Resolution	FIXED
Resolved on	Jul 30, 2014 23:50
Version	trunk
OS	All
Reporter	LLVM Bugzilla Contributor
CC	@majnemer,@nunoplopes,@regehr

Extended Description

When optimizing x-(-A) to x+A, the current logic of preserving nsw/nuw has a hole.

For example, the current logic transforms "x -nsw (0-A)" to "x +nsw A". However, when x = a = INT_MIN = -2^31, "x -nsw (0 - A)" is defined but "x +nsw A"
returns a poison value. The following test case exposes the bug.

define i32 @​test37(i32 %A, i32 %x) {
%B = sub i32 0, %A
%C = sub nsw i32 %x, %B
ret i32 %C
; CHECK-LABEL: @​test37(
; CHECK: %C = add i32 %x, %A
; CHECK: ret i32 %C
}

To fix this issue, we need to consider the nsw/nuw of both minuses in x-(-A). If both of them are marked nsw/nuw, we can preserve the nsw/nuw in the result x+A.

[llvmbot]

assigned to @majnemer

[majnemer]

Fixed with r214384 and r214385.