[Valgrind-developers] [PATCH 1/2] x86: optimize XCHG to MOV for same-register forms

[H.J. Lu]

On Wed, Jul 8, 2026 at 2:04 PM Jan Beulich <jbeulich@suse.com> wrote:
>
> On 08.07.2026 08:00, H.J. Lu wrote:
> > On Wed, Jul 8, 2026 at 1:54 PM Jan Beulich <jbeulich@suse.com> wrote:
> >>
> >> On 07.07.2026 23:18, H.J. Lu wrote:
> >>> On Tue, Jul 7, 2026 at 10:35 PM Jan Beulich <jbeulich@suse.com> wrote:
> >>>> On 07.07.2026 16:21, Sam James wrote:
> >>>>> Just what you keep saying?
> >>>>>
> >>>>> "Assembler optimization is intended for hand-written code, not to be
> >>>>> used on compiler output." which then gives us licence to dispose with
> >>>>> bugs reports like this, as opposed to the ambiguity right now.
> >>>>
> >>>> I fear I wouldn't be happy with making such a statement in doc. For one,
> >>>> "intended" is weak enough that people may still think the options are
> >>>> worthwhile to use on compiler output. Plus there's the issue with inline
> >>>> assembly, which imo can plausibly be subject to optimization. Yet at
> >>>> this time we have no way to have optimization "kick in" only on those
> >>>> portions.
> >>>>
> >>>> I could perhaps live with a yet weaker version of what you suggest:
> >>>>
> >>>> "Assembler optimization is intended primarily for hand-written code.  If
> >>>
> >>> I disagree.  I added -O to assembler for
> >>>
> >>> On x86, some instructions have alternate shorter encodings:
> >>>
> >>> 1. When the upper 32 bits of destination registers of
> >>>
> >>> andq $imm31, %r64
> >>> testq $imm31, %r64
> >>> xorq %r64, %r64
> >>> subq %r64, %r64
> >>>
> >>> known to be zero, we can encode them without the REX_W bit:
> >>>
> >>> andl $imm31, %r32
> >>> testl $imm31, %r32
> >>> xorl %r32, %r32
> >>> subl %r32, %r32
> >>>
> >>> This optimization is enabled with -O, -O2 and -Os.
> >>> 2. Since 0xb0 mov with 32-bit destination registers zero-extends 32-bit
> >>> immediate to 64-bit destination register, we can use it to encode 64-bit
> >>> mov with 32-bit immediates.  This optimization is enabled with -O, -O2
> >>> and -Os.
> >>> 3. Since the upper bits of destination registers of VEX128 and EVEX128
> >>> instructions are extended to zero, if all bits of destination registers
> >>> of AVX256 or AVX512 instructions are zero, we can use VEX128 or EVEX128
> >>> encoding to encode AVX256 or AVX512 instructions.  When 2 source
> >>> registers are identical, AVX256 and AVX512 andn and xor instructions:
> >>>
> >>> VOP %reg, %reg, %dest_reg
> >>>
> >>> can be encoded with
> >>>
> >>> VOP128 %reg, %reg, %dest_reg
> >>>
> >>> This optimization is enabled with -O2 and -Os.
> >>> 4. 16-bit, 32-bit and 64-bit register tests with immediate may be
> >>> encoded as 8-bit register test with immediate.  This optimization is
> >>> enabled with -Os.
> >>>
> >>> These optimizations were intended for compiler generated assembly
> >>> codes.
> >>
> >> Why would that be? I.e. why would the compiler emit sub-optimal code,
> >> for the assembler to tidy after it?
> >
> > Compilers may miss the optimal encoding.
>
> Which can and imo should be fixed there.
>

I first proposed to optimize SSE move in GCC in Feb, 2019:

https://gcc.gnu.org/pipermail/gcc-patches/2019-February/516941.html

It finally went in Mar, 2020:

commit 5358e8f5800daa0012fc9d06705d64bbb21fa07b
Author: H.J. Lu <hjl.tools@gmail.com>
Date:   Thu Mar 5 16:45:05 2020 -0800

    i386: Properly encode vector registers in vector move

Changing compiler is a very long process.   Doing the same thing
in x86 assembler is much easier and it also works with released
versions of GCC which don't have such optimized encoding.

--
H.J.