https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65871
--- Comment #10 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to James Almer from comment #9)
> Created attachment 35804 [details]
> Preprocessed code where a test instruction is still generated
>
> Please look at the attachment.
>
> [jamrial@archVM ~]$ gcc -std=c99 -march=haswell -O3 -c hevc_ps.i
>
> [jamrial@archVM ~]$ objdump --disassemble hevc_ps.o | grep -B2 -A2 bzhi
> 5854: c4 62 22 f7 d6 sarx %r11d,%esi,%r10d
> 5859: c4 62 22 f7 da sarx %r11d,%edx,%r11d
> 585e: c4 e2 70 f5 f6 bzhi %ecx,%esi,%esi
> 5863: 45 89 9e 68 33 00 00 mov %r11d,0x3368(%r14)
> 586a: 41 89 c3 mov %eax,%r11d
> --
> 589d: 85 f6 test %esi,%esi
> 589f: 75 0d jne 58ae
> <ff_hevc_decode_nal_sps+0x1a6e>
> 58a1: c4 e2 70 f5 d2 bzhi %ecx,%edx,%edx
> 58a6: 85 d2 test %edx,%edx
> 58a8: 0f 84 5a 03 00 00 je 5c08
> <ff_hevc_decode_nal_sps+0x1dc8>
[...]
> Not sure why the new model is not working here.
Because of the cost model. Combine pass says:
Trying 2631 -> 2633:
Successfully matched this instruction:
(set (reg:CCZ 17 flags)
(compare:CCZ (zero_extract:SI (reg:SI 2360 [ D.14992 ])
(umin:SI (zero_extend:SI (subreg:QI (reg:SI 301 [ D.14999 ]) 0))
(const_int 32 [0x20]))
(const_int 0 [0]))
(const_int 0 [0])))
rejecting combination of insns 2631 and 2633
original costs 12 + 4 = 16
replacement cost 18
