RFR: 8350177: C2 SuperWord: Integer.numberOfLeadingZeros, numberOfTrailingZeros, reverse and bitCount have input types wrongly turncated for byte and short

[Emanuel Peter]

On Wed, 28 May 2025 07:38:16 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> Hi all,
>> This patch fixes cases in SuperWord when compiling subword types where vectorized code would be given a narrower type than expected, leading to miscompilation due to truncation. This fix is a generalization of the same fix applied for `Integer.reverseBytes` in [JDK-8305324](https://bugs.openjdk.org/browse/JDK-8305324). The patch introduces a check for nodes that are known to tolerate truncation, so that any future cases of subword truncation will avoid creating miscompiled code.
>> 
>> The patch reuses the existing logic to set the type of the vectors to int, which currently disables vectorization for the affected patterns entirely. Once [JDK-8342095](https://bugs.openjdk.org/browse/JDK-8342095) is merged and automatic casting support is added the autovectorizer should automatically insert casts to and from int, maintaining correctness.
>> 
>> I've added an IR test that checks for correctly compiled outputs. Thoughts and reviews would be appreciated!
>
> src/hotspot/share/opto/superword.cpp line 2496:
> 
>> 2494:   int opc = in->Opcode();
>> 2495:   return opc == Op_AddI || opc == Op_SubI || opc == Op_MulI || opc == Op_AndI || opc == Op_OrI || opc == Op_XorI
>> 2496:     || opc == Op_ReverseBytesS || opc == Op_ReverseBytesUS;
> 
> A switch might look nicer here, and be easier to extend later on ;)

This list is a little scary... how do we know that we have all cases in it, and we are not getting regressions because we are missing some?

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

PR Review Comment: https://git.openjdk.org/jdk/pull/25440#discussion_r2111167881