Re: freeze for sbcl-2.6.8

Forgive me for chimming in, but a question, since you are already
talking about xstore/xrstor and want more performance. Can we switch to
"compact" versions of those (xsavec & co)? These, potentially
write/restore less data, in cases where saving/restoring xmm/ymm/zmm is
enabled.

I learned about it when I was patching for zmm. I didn't dare to try
to change that, but instead patched in support to what was already in.

As I understand it, when full store is needed (zmm registers in use),
than the performance is roughly the same, but if zmm/ymm/xmm regs do not
need to be saved becuase they weren't written since the last clear, they
won't be saved either.

Compacted stores/restores are supported on Linux since kernel 4.4, year
2016, and on Windows since version 10 some release number in 2015.
________________________________
Från: Douglas Katzman via Sbcl-devel <[email protected]>
Skickat: den 22 augusti 2026 18:07
Till: Stas Boukarev <[email protected]>
Kopia: Christophe Rhodes <[email protected]>; [email protected] <[email protected]>
Ämne: Re: [Sbcl-devel] freeze for sbcl-2.6.8

I have the fix for that crash.   I mistakenly thought that FPR-SAVE/FPR-RESTORE no longer need to preserve RDX when using XRSTOR but they still do - but only for #+win32, so I'll add that.
However, Gemini claims the following theoretical performance gain by simply avoiding XSAVE/XRSTOR entirely in the YMM case (not the avx512 of course) which is a very interesting point -

  1.  ~4x Faster: 16 * VMOVAPS runs in ~15–20 cycles vs 80+ cycles for XSAVE/XRSTOR.

Do you think I can commit this register changing patch + its fix for now, and we'll leave the above suggestion for next release (despite also being a valid fix for my crash) ?

_______________________________________________
Sbcl-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/sbcl-devel