RFR: 8277621: ARM32: multiple fastdebug failures with "bad AD file" after JDK-8276162

[Dean Long]

On Wed, 8 Dec 2021 03:59:09 GMT, Hao Sun <haosun at openjdk.org> wrote:

> JDK-8276162 introduced an optimization that creates the following two
> kinds of IR shapes:
> 
> 1. `CMoveI (Bool (CmpUL ...) ...)`: conditionally moving two ints based
> on the comparison result of two unsigned longs.
> 
> 2. `CMoveP (Bool (CmpUL ...) ...)`: conditionally moving two pointers
> based on the comparison result of two unsigned longs.
> 
> But the corresponding match rules are missing for arm32. JDK-8277324 and
> JDK-8277753 complemented the missing rules for x86-32. We do the same
> thing to arm32 in this patch.
> 
> For IR shape 1, the missing rules in arm32 are in form of
> "cmovIL_AA_BB_U", where AA can be reg/imm16/immMov(i.e. three variants
> of moving immediate/register), and BB can be LTGE/EQNE/LEGT(i.e. three
> condition code registers for unsigned long comparisons).
> 
> For IR shape 2, the missing rules are in form of "cmovPL_CC_BB_U", where
> CC can be reg/imm, and BB has the same meaning with IR shape 1.
> 
> Minor updates:
> 1. "cmpOpUL/comOpUL_commute" should be used for "cmovLL_AA_BB_U" rules.
> 2. Rename "cmovIL_imm_BB" rules to "cmovIL_imm16_BB".
> 3. Add "cmovIL_immMov_BB" rules.
> 4. Style issue: remove the extra space in the predicate statement for
> cmov* rules.
> 
> Test:
> We ran tier 1~3 tests on arm32 platform. With this patch, "bad AD file"
> errors are gone without introducing test regression.

Marked as reviewed by dlong (Reviewer).

This looks fine.  I think the cmove rules could be consolidated into fewer rules if the flags operand types were collapsed into fewer types likes x64 does, but that's a bigger change and probably deserves its own RFE.

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

PR: https://git.openjdk.java.net/jdk/pull/6759