missed optz'n in mul+div

[llvmbot]

Bugzilla Link	20681
Resolution	FIXED
Resolved on	Aug 16, 2014 03:55
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@majnemer

Extended Description

I noticed this when looking at some vectorized code. We turned the shift+divide into vector shift and vector divide. It was doing x<<2/12, which is really just x/3.

Testcases:

define i32 @​test1(i32 %a) {
%A = shl nsw i32 %a, 2
%B = sdiv i32 %A, 12
; CHECK: %B = sdiv i32 %a, 3
ret i32 %B
}

define i32 @​test2(i32 %a) {
%A = mul nsw i32 %a, 3
%B = sdiv i32 %A, 12
; CHECK: %B = sdiv i32 %a, 4 ;; we don't simplify that to ashr?
ret i32 %B
}

define i32 @​test3(i32 %a) {
%A = shl nuw i32 %a, 2
%B = udiv i32 %A, 12
; CHECK: %B = udiv i32 %a, 3
ret i32 %B
}

define i32 @​test4(i32 %a) {
%A = mul nuw i32 %a, 3
%B = udiv i32 %A, 12
; CHECK: %B = lshr i32 %a, 2
ret i32 %B
}

[llvmbot]

assigned to @majnemer

[majnemer]

Fixed in r215815.