https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82498
Bug ID: 82498
Summary: Missed optimization for x86 rotate instruction
Product: gcc
Version: 7.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: lloyd at randombit dot net
Target Milestone: ---
GCC doesn't seem to realize that x86 masks the high bits in the rol/ror
instructions. GCC 7.2.0 on x86-64 compiles this function, which is attempting
to do a 32-bit rotate without invoking undefined behavior
#include <stdint.h>
uint32_t rotate_left(uint32_t input, uint8_t rot)
{
if(rot == 0)
return input;
rot %= 8 * sizeof(uint32_t);
return static_cast<uint32_t>((input << rot) | (input >>
(8*sizeof(uint32_t)-rot)));;
}
Into
movl %esi, %ecx # rot, rot
movl %edi, %eax # input, tmp97
andl $31, %ecx #, rot
roll %cl, %eax # rot, tmp97
testb %sil, %sil # rot
cmove %edi, %eax # tmp97,, input, <retval>
The `andl` is unnecessary as the machine will mask the rotation amount for us.
In addition the testb/cmov pair can be omitted. Overall this resulted in a ~15%
slowdown in some code using many variable rotations (CAST-128 cipher being used
in an OpenPGP library).
Some related (but not quite the same, and supposedly fixed) issues: 57157 59100