Re: SIGILL when CPUTYPE set to anyting witjh avx and CFLAGS -O2

On Fri, Dec 29, 2023 at 09:12:35PM +0100, Harry Schmalzbauer wrote:
> On 12/29/23 18:00, Dimitry Andric wrote:
> > On 29 Dec 2023, at 11:14, Harry Schmalzbauer <freebsd@omnilan.de> wrote:
> > > 
> > > at least since 14-release, I can't compile base/ports with CPUTYPE post nehalem.  Any '-march' declaring AVX(*) capability leads to SIGILLing binaries with default -O2 CFLAGS.
> ..
> > > 
> > > If I compile without CFLAGS -O2 ("CFLAGS= -pipe"), -march=skylake-avx512 the binary doens't crash with SIGILL.
> > > 
> > > Is CPUTYPE supposed to stay untouched these days?
> > 
> > Obviously not, but are you sure your target CPU is correct? If it is not, the compiler may insert unsupported instructions.
> 
> Ho Dimitry,  thanks for your attention!
> I have to admit that I know much too less about contemporary compilers and
> CPUs likewise, but target CPU matches -march ;-)
> I nailed it down to AVX instructions.  Starting with 'sandybridge', Intel
> CPUs provide avx instruction units, and starting with CPUTYPE?=sandybridge,
> binaries crash on stable/14 (and late 14-current) on the corresponding
> machines.  But only in combination with CFLAGS -O2!
> 
> 
> > In any case, for the above gdb session, try the command "disassemble" to see which particular instruction it crashed on. That may give more information.
> 
> Thanks for the hint!
> This came out:
> Program terminated with signal SIGILL, Illegal instruction.
> Privileged opcode.
> #0  0x000000000041e040 in init_yank ()
> (gdb) disassemble
> Dump of assembler code for function init_yank:
>    0x000000000041e030 <+0>:	push   %rbp
>    0x000000000041e031 <+1>:	mov    %rsp,%rbp
>    0x000000000041e034 <+4>:	vpxor  %xmm0,%xmm0,%xmm0
>    0x000000000041e038 <+8>:	vmovdqa -0x202c10(%rip),%xmm1        # 0x21b430
> => 0x000000000041e040 <+16>:	kxnorw %k0,%k0,%k1
This is AVX512 instruction (some of it subset, probably F), not AVX.

>    0x000000000041e044 <+20>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e04f <+31>:	vmovdqa -0x202b77(%rip),%xmm1        # 0x21b4e0
>    0x000000000041e057 <+39>:	kxnorw %k0,%k0,%k1
>    0x000000000041e05b <+43>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e066 <+54>:	vmovdqa -0x202c5e(%rip),%xmm1        # 0x21b410
>    0x000000000041e06e <+62>:	kxnorw %k0,%k0,%k1
>    0x000000000041e072 <+66>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e07d <+77>:	vmovdqa -0x202cf5(%rip),%xmm1        # 0x21b390
>    0x000000000041e085 <+85>:	kxnorw %k0,%k0,%k1
>    0x000000000041e089 <+89>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e094 <+100>:	vmovdqa -0x202b5c(%rip),%xmm1        #
> 0x21b540
>    0x000000000041e09c <+108>:	kxnorw %k0,%k0,%k1
>    0x000000000041e0a0 <+112>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e0ab <+123>:	vmovdqa -0x202e43(%rip),%xmm1        #
> 0x21b270
>    0x000000000041e0b3 <+131>:	kxnorw %k0,%k0,%k1
>    0x000000000041e0b7 <+135>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e0c2 <+146>:	vmovdqa -0x202d0a(%rip),%xmm1        #
> 0x21b3c0
>    0x000000000041e0ca <+154>:	kxnorw %k0,%k0,%k1
>    0x000000000041e0ce <+158>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e0d9 <+169>:	vmovdqa -0x202ac1(%rip),%xmm1        #
> 0x21b620
>    0x000000000041e0e1 <+177>:	kxnorw %k0,%k0,%k1
>    0x000000000041e0e5 <+181>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e0f0 <+192>:	vmovdqa -0x202d58(%rip),%xmm1        #
> 0x21b3a0
>    0x000000000041e0f8 <+200>:	kxnorw %k0,%k0,%k1
>    0x000000000041e0fc <+204>:	vpscatterdq %ymm0,0x55f180(,%xmm1,1){%k1}
>    0x000000000041e107 <+215>:	movq   $0x0,0x1414ee(%rip)        # 0x55f600
> <y_regs+1152>
>    0x000000000041e112 <+226>:	pop    %rbp
>    0x000000000041e113 <+227>:	vzeroupper
>    0x000000000041e116 <+230>:	ret
> End of assembler dump.
> 
> Example is from 'cc -c -I. -Iproto -DHAVE_CONFIG_H   -DLIBICONV_PLUG
> -I/usr/local/include  -O2 -pipe -march=skylake-avx512  -DLIBICONV_PLUG
> -fstack-protector-strong -fno-strict-aliasing  -D_REENTRANT
> -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=1        -o objects/alloc.o alloc.c'
> with CPU:
> Intel(R) Xeon(R) CPU E3-1275 v6 @ 3.80GHz (3800.00-MHz K8-class CPU)
> 
> When I 1st noticed, I checked also much more recent CPUs - with the same
> result - if CPUTUPE is set to any avx-involving optimization together with
> -O2, binaries crash with SIGILL on Haswell, SkyLake, TigerLake...
> My last world/ports builds (sucessfully with CPUTYPE?=ivybridge) were with
> stable-13 (arround 13.1) - and llvm 13 I think - can't remember exactly.
> 
> Thanks,
> -harry