Re: [PATCH v2] combine: Narrow comparison of memory and constant

[Jeff Law <jeffreyalaw@gmail.com>]

On 7/31/23 15:43, Prathamesh Kulkarni via Gcc-patches wrote:
> On Mon, 19 Jun 2023 at 19:59, Stefan Schulze Frielinghaus via
> Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
>>
>> Comparisons between memory and constants might be done in a smaller mode
>> resulting in smaller constants which might finally end up as immediates
>> instead of in the literal pool.
>>
>> For example, on s390x a non-symmetric comparison like
>>    x <= 0x3fffffffffffffff
>> results in the constant being spilled to the literal pool and an 8 byte
>> memory comparison is emitted.  Ideally, an equivalent comparison
>>    x0 <= 0x3f
>> where x0 is the most significant byte of x, is emitted where the
>> constant is smaller and more likely to materialize as an immediate.
>>
>> Similarly, comparisons of the form
>>    x >= 0x4000000000000000
>> can be shortened into x0 >= 0x40.
>>
>> Bootstrapped and regtested on s390x, x64, aarch64, and powerpc64le.
>> Note, the new tests show that for the mentioned little-endian targets
>> the optimization does not materialize since either the costs of the new
>> instructions are higher or they do not match.  Still ok for mainline?
> Hi Stefan,
> Unfortunately this patch (committed in 7cdd0860949c6c3232e6cff1d7ca37bb5234074c)
> caused the following ICE on armv8l-unknown-linux-gnu:
> during RTL pass: combine
> ../../../gcc/libgcc/fixed-bit.c: In function ‘__gnu_saturate1sq’:
> ../../../gcc/libgcc/fixed-bit.c:210:1: internal compiler error: in
> decompose, at rtl.h:2297
>    210 | }
>        | ^
> 0xaa23e3 wi::int_traits<std::pair<rtx_def*, machine_mode>
>> ::decompose(long long*, unsigned int, std::pair<rtx_def*,
> machine_mode> const&)
>          ../../gcc/gcc/rtl.h:2297
[ ... ]
Yea, we're seeing something very similar on nios2-linux-gnu building the 
kernel.

Prathamesh, can you extract the .i file for fixed-bit on armv8 and open 
a bug for this issue, attaching the .i file as well as the right command 
line options necessary to reproduce the failure.  THat way Stefan can 
tackle it with a cross compiler.

Thanks,
jeff