On Thu, Oct 01, 2026 at 06:41:20AM +0200, Christophe Leroy (CS GROUP) wrote: > Hi, > > Le 30/09/2026 à 17:16, Jason A. Donenfeld a écrit : > > On Wed, Sep 30, 2026 at 5:13 PM Nathan Chancellor <[email protected]> wrote: > >> > >> On Wed, Sep 30, 2026 at 04:44:29PM +0200, Jason A. Donenfeld wrote: > >>> Nathan, would this be okay with you? > >>> https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgit.zx2c4.com%2Flinux-rng%2Fcommit%2F%3Fid%3Dd216701724b7d8209ff42150658ad5c712bdb503&data=05%7C02%7Cchristophe.leroy2%40cs-soprasteria.com%7C8f69935bffb344bf0ddc08df1f05e358%7C8b87af7d86474dc78df45f69a2011bb5%7C0%7C0%7C639263782324080837%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=BA7Fyr0%2FynggvBnR%2FpiCvNCQHtNBagqvSb0i11AvoLU%3D&reserved=0 > >> > >> Can you stick > >> > >> Cc: [email protected] # v6.12+ > >> Closes: > >> https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FClangBuiltLinux%2Flinux%2Fissues%2F2183&data=05%7C02%7Cchristophe.leroy2%40cs-soprasteria.com%7C8f69935bffb344bf0ddc08df1f05e358%7C8b87af7d86474dc78df45f69a2011bb5%7C0%7C0%7C639263782324110679%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=6C8%2FAEQoh%2B8pfgfYQkL7qf5Ue%2F5sAGtfqRptzsSG0CM%3D&reserved=0 > >> > >> on that? Otherwise, looks good to me, that's basically what I had for my > >> v3 locally. > > > > Sure, done. Also removed the now-unused array_size.h include. > > I'm still very sceptic with this patch. You are degrading the behaviour > with GCC for a problem with CLANG. Why ? > > Before the patch, with both GCC 13 and GCC 16 on powerpc32 I get a > pretty standard optimised loop that clears words 4 by 4 (with auto > increment of pointer) which is the most optimal on powerpc: > > 3f0: 39 00 00 0c li r8,12 > 3f4: 35 08 ff fc addic. r8,r8,-4 > 3f8: 91 49 00 04 stw r10,4(r9) > 3fc: 91 49 00 08 stw r10,8(r9) > 400: 91 49 00 0c stw r10,12(r9) > 404: 95 49 00 10 stwu r10,16(r9) > 408: 40 82 ff ec bne 3f4 <__c_kernel_getrandom+0x3f4> > > With the patch, > > With GCC 13 I get a very suboptimal loop copying bytes one by one > > 3d8: 39 40 00 34 li r10,52 > ... > 3e4: 39 20 00 00 li r9,0 > 3e8: 7d 49 03 a6 mtctr r10 > 3ec: 9d 3e 00 01 stbu r9,1(r30) > 3f0: 42 00 ff fc bdnz 3ec <__c_kernel_getrandom+0x3ec> > > With GCC 16 I get something a bit better but not as good as before, it > is a loop clearing words only one by one and incrementing pointer with > an additional insn instead of using auto-increment instruction stwu. > > 3e0: 39 40 00 0d li r10,13 > ... > 3f0: 7d 49 03 a6 mtctr r10 > 3f4: 91 3f 00 00 stw r9,0(r31) > 3f8: 3b ff 00 04 addi r31,r31,4 > 3fc: 42 00 ff f8 bdnz 3f4 <__c_kernel_getrandom+0x3f4> > > Please restrict the patch to clang builds.
Darn. Yea. The naive memset kills optimizations. Okay, new strategy: - on clang, pass `-mllvm -max-store-memset=4294967295` - on gcc, pass `-finline-stringops=memset` And keep the same code. (Or, better, see if those options generate good code with b7bad082e113640fc81200ff869e5c2d7a9c29a2 reverted; I would prefer that simpler initializer.) Nathan, do these work? Jason
