RFR: 8326378: [PPC64] CodeEntryAlignment too large

[Dean Long]

On Wed, 21 Feb 2024 05:56:22 GMT, Martin Doerr <mdoerr at openjdk.org> wrote:

> I'd like to save space in the code cache. No performance regression observed. See JBS issue for more details.
> 
> Old:
> 
> [Instructions begin]
>  ;; N1181: #	out( B1 ) <- BLOCK HEAD IS JUNK  Freq: 1
>   0x00007fff7cbda880:   nop                                 ;   {no_reloc}
>   0x00007fff7cbda884:   nop
>   0x00007fff7cbda888:   nop
>   0x00007fff7cbda88c:   nop
>   0x00007fff7cbda890:   nop
>   0x00007fff7cbda894:   nop
>   0x00007fff7cbda898:   nop
>   0x00007fff7cbda89c:   nop
>   0x00007fff7cbda8a0:   nop
>   0x00007fff7cbda8a4:   nop
>   0x00007fff7cbda8a8:   nop
>   0x00007fff7cbda8ac:   nop
>   0x00007fff7cbda8b0:   nop
>   0x00007fff7cbda8b4:   nop
>   0x00007fff7cbda8b8:   nop
>   0x00007fff7cbda8bc:   nop
>   0x00007fff7cbda8c0:   nop
>   0x00007fff7cbda8c4:   nop
>   0x00007fff7cbda8c8:   nop
>   0x00007fff7cbda8cc:   nop
>   0x00007fff7cbda8d0:   nop
>   0x00007fff7cbda8d4:   nop
>   0x00007fff7cbda8d8:   nop
>   0x00007fff7cbda8dc:   nop
>   0x00007fff7cbda8e0:   nop
>   0x00007fff7cbda8e4:   nop
>   0x00007fff7cbda8e8:   nop
>   0x00007fff7cbda8ec:   nop
>   0x00007fff7cbda8f0:   nop
> [Entry Point]
>   # {method} {0x00007fff407064e8} 'bogus_test' '()V' in 'TestAlloc'
>   #           [sp+0x90]  (sp of caller)
>   0x00007fff7cbda8f4:   lwz     r11,8(r3)
>   0x00007fff7cbda8f8:   ld      r12,8(r19)
>   0x00007fff7cbda8fc:   td      3,r11,r12               ;trap: ic miss check
> [Verified Entry Point]
> 
> 
> New:
> 
> [Instructions begin]
>  ;; N1181: #	out( B1 ) <- BLOCK HEAD IS JUNK  Freq: 1
>   0x00007fff802f8380:   nop                                 ;   {no_reloc}
>   0x00007fff802f8384:   nop
>   0x00007fff802f8388:   nop
>   0x00007fff802f838c:   nop
>   0x00007fff802f8390:   nop
>   0x00007fff802f8394:   nop
>   0x00007fff802f8398:   nop
>   0x00007fff802f839c:   nop
>   0x00007fff802f83a0:   nop
>   0x00007fff802f83a4:   nop
>   0x00007fff802f83a8:   nop
>   0x00007fff802f83ac:   nop
>   0x00007fff802f83b0:   nop
> [Entry Point]
>   # {method} {0x00007fff487064e8} 'bogus_test' '()V' in 'TestAlloc'
>   #           [sp+0x90]  (sp of caller)
>   0x00007fff802f83b4:   lwz     r11,8(r3)
>   0x00007fff802f83b8:   ld      r12,8(r19)
>   0x00007fff802f83bc:   td      3,r11,r12               ;trap: ic miss check
> [Verified Entry Point]

src/hotspot/cpu/ppc/globals_ppc.hpp line 62:

> 60: 
> 61: // Use large code-entry alignment.
> 62: define_pd_global(uintx, CodeCacheSegmentSize,  128);

Shouldn't also changing CodeCacheSegmentSize to match CodeEntryAlignment save a little more in the code cache?

If we did all alignment on the actual address rather than an offset that is relative, I think we could get rid of some of the constraints that require CodeCacheSegmentSize to be at least as large as CodeEntryAlignment and OptoLoopAlignment.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/17939#discussion_r1498297906