On Wed, Sep 2, 2020 at 11:17 AM <peterz@xxxxxxxxxxxxx> wrote: > > On Wed, Sep 02, 2020 at 07:50:36AM +0200, Uros Bizjak wrote: > > On Tue, Sep 1, 2020 at 9:12 PM Jason A. Donenfeld <Jason@xxxxxxxxx> wrote: > > > > > > On Tue, Sep 1, 2020 at 8:13 PM Jason A. Donenfeld <Jason@xxxxxxxxx> wrote: > > > > operands are the same. Also, have you seen any measurable differences > > > > when benching this? I can stick it into kbench9000 to see if you > > > > haven't looked yet. > > > > > > On a Skylake server (Xeon Gold 5120), I'm unable to see any measurable > > > difference with this, at all, no matter how much I median or mean or > > > reduce noise by disabling interrupts. > > > > > > One thing that sticks out is that all the replacements of r8-r15 by > > > their %r8d-r15d counterparts still have the REX prefix, as is > > > necessary to access those registers. The only ones worth changing, > > > then, are the legacy registers, and on a whole, this amounts to only > > > 48 bytes of difference. > > > > The patch implements one of x86 target specific optimizations, > > performed by gcc. The optimization results in code size savings of one > > byte, where REX prefix is omitted with legacy registers, but otherwise > > should have no measurable runtime effect. Since gcc applies this > > optimization universally to all integer registers, I took the same > > approach and implemented the same change to legacy and REX registers. > > As measured above, 48 bytes saved is a good result for such a trivial > > optimization. > > Could we instead implement this optimization in GAS ? Then we can leave > the code as-is. I don't think that e.g. "xorq %rax,%rax" should universally be translated to "xorl %eax,%eax" in the assembler. Perhaps the author expected exactly 3 bytes (to align the code or similar), and the assembler would change the length to 2 bytes behind his back, breaking the expectations. Uros.
