RFR: 8318509: x86 count_positives intrinsic broken for -XX:AVX3Threshold=0

[Jatin Bhateja]

On Thu, 19 Oct 2023 12:54:47 GMT, Claes Redestad <redestad at openjdk.org> wrote:

> The AVX-512 implementation of count_positives is broken when there are negative values in the tail block, since the len register is repurposed to then re-read as-if it still contained the original value. 
> 
> Proposed fix is to restore len from the value in tmp1, which at this point holds the count of bytes in the tail. 
> 
> Existing TestCountPositives test is pretty exhaustive and just needs to be run explicitly with AVX3Threshold=0 enabled.

src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp line 3888:

> 3886:       mov64(tmp3_aliased, 0xFFFFFFFFFFFFFFFF);
> 3887:       shlxq(tmp3_aliased, tmp3_aliased, tmp1);
> 3888:       notq(tmp3_aliased);

We can replace these two instructions with BZHI under BMI2 check.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/16267#discussion_r1365864031