clang does not properly optimize xor expressions

[avg-I]

Bugzilla Link	36622
Version	5.0
OS	FreeBSD
CC	@RKSimon,@zygoloid,@rotateright

Extended Description

It appears that in some code sequences Clang fails to recognize that x xor C xor C == x where C is a compile-time constant.

For instance:

uint32_t combine(uint16_t x, uint16_t y) { uint32_t r = 0x5a5a7777;

    r ^= (r ^ x) & 0xffff;
    r ^= ((r >> 16) ^ y) << 16;

    return (r);

}

In this code the initial value of r is not important and the code should be equivalent to:

uint32_t combine(uint16_t x, uint16_t y) { return (x | ((uint32_t)y << 16)); }

But even with -O3 Clang produces the following assembler code:

    orl     $1515847680, %edi       # imm = 0x5A5A0000
    xorl    $23130, %esi            # imm = 0x5A5A
    shll    $16, %esi
    xorl    %edi, %esi
    movl    %esi, %eax

---

It's easy to see that 0x5A5A will always cancel out through double application via xor, but Clang does not recognize that.

[llvmbot]

Maybe this can be fixed in reassociate or instcombine.

There are two zext getting in the way:

define dso_local i32 @​combine(unsigned short, unsigned short)(i16 zeroext, i16 zeroext) local_unnamed_addr #​0 { %3 = zext i16 %0 to i32 %4 = or i32 %3, 1515847680 %5 = xor i16 %1, 23130 %6 = zext i16 %5 to i32 %7 = shl nuw i32 %6, 16 %8 = xor i32 %7, %4 ret i32 %8 }

If you change the function argument to be:

uint32_t combine(uint32_t x, uint32_t y)

it produces optimal code, I think:

define dso_local i32 @​combine(unsigned int, unsigned int)(i32, i32) local_unnamed_addr #​0 { %3 = shl i32 %1, 16 %4 = and i32 %0, 65535 %5 = or i32 %4, %3 ret i32 %5 }

or equivalently, in asm:

combine(unsigned int, unsigned int): # @​combine(unsigned int, unsigned int) shl esi, 16 movzx eax, di or eax, esi ret