[11u] RFR (S) 8249189: AARCH64: more L2I conversions can be skipped

[Andrew Haley]

On 5/12/21 6:03 PM, Dmitry Chuyko wrote:
> Hello,
> 
> Original RFE: https://bugs.openjdk.java.net/browse/JDK-8249189
> 
> aarch64_ad.m4 and aarch64.ad are decorated different in 11u. 
> Nevertheless changes are very straightforward.
> 
> aarch64_ad.m4
> Reapplied changes for UBFIZ_INSN and it's usages.
> 
> aarch64.ad
> Following the original patch: operand immL_positive_bitmaskI added after 
> operand immI_bitmask, instruct ubfizwIConvI2L and ubfizLConvL2I added 
> between ubfizL and ubfizIConvI2L, instruct ubfizLConvL2Ix added after 
> ubfizIConvI2L.
> 
> Mask type is changed long to keep it in line with the rest of the .ad, 
> and thus the diff structure is close to the original.
> 
> 11u webrev: http://cr.openjdk.java.net/~dchuyko/8249189/webrev.11u.00/
> 
> Testing: compiler/c2/TestSkipLongToIntCast.java, tier1, tier2.
> 
> Generated part of aarch64.ad file matches the output of 'm4 aarch64_ad.m4'.

Is this tested anywhere?

-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671