[PATCH] D26818: Bug fix and a proposed improvement in integer division expansion

[Dmitri Shtilman via llvm-commits]

Hi all,

Another improvement in integer division sw expansion and a bug fix.

The special case code in udiv expansion is currently broken, namely the check for divisor==1 :

// ; special-cases:
// ;   %ret0_1      = icmp eq i32 %divisor, 0
// ;   %ret0_2      = icmp eq i32 %dividend, 0
// ;   %ret0_3      = or i1 %ret0_1, %ret0_2
// ;   %tmp0        = tail call i32 @llvm.ctlz.i32(i32 %divisor, i1 true)
// ;   %tmp1        = tail call i32 @llvm.ctlz.i32(i32 %dividend, i1 true)
// ;   %sr          = sub nsw i32 %tmp0, %tmp1
// ;   %ret0_4      = icmp ugt i32 %sr, 31
// ;   %ret0        = or i1 %ret0_3, %ret0_4
// ;   %retDividend = icmp eq i32 %sr, 31
// ;   %retVal      = select i1 %ret0, i32 0, i32 %dividend
// ;   %earlyRet    = or i1 %ret0, %retDividend
// ;   br i1 %earlyRet, label %end, label %bb1

The intention must have been to use %tmp0, not %sr:

// ;   %retDividend = icmp eq i32 %tmp0, 31

The following patch fixes this special case and also extends it to cover divisions by a power-of-2 which will be routed through a simple shift-right instead of going through the full division loop. Division by 1 will also go through the shift-right, yielding the dividend [shifted by 0]. Power-of-2 divisors are identified through adding trailing zeroes to the already calculated leading zeroes to verify that only one bit is set.

Please review.
Dmitri

https://reviews.llvm.org/D26818
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161130/a5be2f85/attachment.html>