Don't use s0 in Risc-V zcmp test compile

[will-v-pi]

This file currently fails to compile with any compilers, resulting in pico_riscv_gcc always falling through to the non-zcmp arch string

The fix is to change the registers used to s1 and s2 instead

[kilograham]

ha; i was just making the same fix - but with said compiler i still didn't get any cm.push/pop etc.

[Wren6991]

but with said compiler i still didn't get any cm.push/pop etc.

I think the Zcmp codegen itself has regressed in GCC. It looks like prolog/epilog scheduling can cause the push/pop pattern matching to fail (with no performance benefit from the scheduling). You should still get them, but they're rarer than they should be.

Do you still get cm.mva01s and cm.mvsa01?

[lurch]

Is this also the fix for #3073 or is that a similar-but-different problem?

[will-v-pi]

Is this also the fix for #3073 or is that a similar-but-different problem?

Yes, that is this exact problem

[shtirlic]

When I build with Release variant there are only 75 cm.mva01s instructions, but with MinSizeRel i got around ~600 cm. instructions.

[will-v-pi]

The Multi GCC run (https://github.com/raspberrypi/pico-sdk/actions/runs/29430965108/job/87405548355) has passed the RISC-V GCC 14.1.0 builds, so I think that means this is tested now and can be merged

[lurch]

Is it "expected" for the CI to fail on the Pico W builds for GCC 6 and GCC 7? https://github.com/raspberrypi/pico-sdk/actions/runs/29430965108/job/87405548355

[will-v-pi]

Is it "expected" for the CI to fail on the Pico W builds for GCC 6 and GCC 7? https://github.com/raspberrypi/pico-sdk/actions/runs/29430965108/job/87405548355

Yes, I think that's fixed in some way in multi-gcc-develop branch - ping @liamfraser