Regression: Assert `count_id(wire->name) == 0' failed in ../yosys-src/kernel/rtlil.cc:2621 when building Glasgow Interface Explorer applets in 0.56

[joelpmichael]

Version

Yosys 0.56 (git sha1 9c447ad, ccache clang++ 18.1.3 -O3 -flto -flto)

On which OS did this happen?

macOS

Reproduction Steps

As requested by @whitequark this appears to be an issue with Yosys. Tracking as GlasgowEmbedded/glasgow#994

1. Follow Glasgow Interface Explorer installation as per https://glasgow-embedded.org/latest/install.html at commit 8b02130
2. glasgow -v run selftest (or any other applet)

Expected Behavior

Glasgow applet compiles and runs successfully.

Actual Behavior

Build fails with final error:

T: glasgow.hardware.build_plan: build: ERROR: Assert `count_id(wire->name) == 0' failed in ../yosys-src/kernel/rtlil.cc:2621.
E: glasgow.applet.internal.selftest: gateware build failed with exit code 1; see build log above for details

Full compile log:
glasgow-run-selftest.log

I'm not very familiar with yosys internals, so any help with narrowing down where this issue is would be appreciated!

[whitequark]

Repro: selftest.zip

yosys top.il -p synth_ice40

[georgerennie]

This is happening in the AIGER frontend within synth_ice40/abc so its probably the same as #4451 / #4349 again. It looks like @Ravenslofty had proposed adding splitnets to fix it but this didn't actually happen in the end?

[whitequark]

This is the third Yosys release in a row that's unusable for Glasgow. If there's a commit that exposed this bug in the 0.55..0.56 release span (it seems to work for me on 0.55) I'd like to have it reverted.

[whitequark]

It looks from the linked issues that a revert may not be feasible and it has to be dealt with properly. The proposed splitnets looks like a huge hack to me.

[georgerennie]

Repro: selftest.zip

yosys top.il -p synth_ice40

Interestingly this doesn't repo with my build of 0.56 (Yosys 0.56 (git sha1 9c447ad9d, sccache clang++ 20.1.8 -fPIC -O3), x86 linux with pyosys enabled).

This is the third Yosys release in a row that's unusable for Glasgow. If there's a commit that exposed this bug in the 0.55..0.56 release span (it seems to work for me on 0.55) I'd like to have it reverted.

Given this seems to be hitting a relatively longstanding issue with the aiger frontend, I would guess that rather than directly introducing a bug its more likely that this design now ends up on the wrong side of some chaotic behaviour and hits it - I don't see any commits since 0.55 that seem that directly related on first glance?

[whitequark]

Interestingly this doesn't repo with my build of 0.56 (Yosys 0.56 (git sha1 9c447ad9d, sccache clang++ 20.1.8 -fPIC -O3), x86 linux with pyosys enabled).

<suppressed ~1 debug messages>
ERROR: Assert `count_id(wire->name) == 0' failed in ../yosys-src/kernel/rtlil.cc:2621.
$ yowasp-yosys --version
Yosys 0.56 (git sha1 9c447ad9d, ccache clang++ 18.1.3 -O3 -flto -flto)

Not sure why, reproducibility issue?

Given this seems to be hitting a relatively longstanding issue with the aiger frontend, I would guess that rather than directly introducing a bug its more likely that this design now ends up on the wrong side of some chaotic behaviour and hits it - I don't see any commits since 0.55 that seem that directly related on first glance?

I'm still going to bisect it just in case, but yeah seems so.

[whitequark]

I don't see any commits since 0.55 that seem that directly related on first glance?

I started bisecting and I was wrong earlier: 0.55 is also bad. Since 0.54 and 0.55 weren't tested with Glasgow due to #5182 that probably slipped in too...

[georgerennie]

Not sure why, reproducibility issue?

Yeah I'm not sure, also tried a couple builds I had lying around from the past few months (some with pyosys/sanitizers, some not) and none reproed. I was able to repro it with yowasp-yosys. This produces some differing log statements compared to the native build so the input by the time it reaches abc is definitely different (the AIG has 2274 nodes going into abc on the native build vs 2195 on yowasp for example).

logs_5274.zip

[whitequark]

Yeah I'm not sure, also tried a couple builds I had lying around from the past few months (some with pyosys/sanitizers, some not) and none reproed. I was able to repro it with yowasp-yosys.

I can repro it with native debug builds on amd64 going as far back as 0.52.