abi.decode failing with "Stack too deep" for a list of structs when optimizations are enabled

[thedavidmeister]

Description

Abi decoding SomeStruct[] fails with

CompilerError: Stack too deep. Try compiling with `--via-ir` (cli) or the equivalent `viaIR: true` (standard JSON) while enabling the optimizer. Otherwise, try removing local variables. When compiling inline assembly: Variable headStart is 3 slot(s) too deep inside the stack.K

Strangely, wrapping it in a struct with a SomeStruct field on it "fixes" the issue

https://github.com/beehive-innovation/rain-protocol/pull/452/files#diff-e5cf60434f03c13e62d93df0932957c2208a64a15065489486b9aea0d158e723R9

SomeStruct[] memory foo_ = abi.decode(data_, (SomeStruct[]));

fails, as does

struct Foo {
  SomeStruct[] bar;
}

but this passes

struct Foo {
  SomeStruct bar;
  SomeStruct[] baz;
}

The issue only appears with compiler optimizations enabled and appears for both IR and non-IR

Environment

• Compiler version: 0.8.17
• Target EVM version (as per compiler settings):
• Framework/IDE (e.g. Truffle or Remix): hardhat
• EVM execution environment / backend / blockchain client:
• Operating system:

Steps to Reproduce

https://github.com/beehive-innovation/rain-protocol/pull/452/files#diff-e5cf60434f03c13e62d93df0932957c2208a64a15065489486b9aea0d158e723R9

[thedavidmeister]

cc @cameel

[nikola-matic]

Hi @thedavidmeister, and sorry for the late reply. If you have a minimal reproduction scenario, feel free to open another issue; if not, unfortunately, we don't have the time at the moment to sift through a 15k line PR to do so ourselves, especially seeing as we already have a set of such repros specifically for via-ir.

[cameel]

To add more context, it's not that we don't need more repros - please do report bugs when you run into them - just in this particular case we already have #13906, which could end up being the same thing (it's not certain though and a new repro would still be nice to have). At the same time this has been sitting here for a while untouched because, reducing a 15k LOC project into a small example is quite tedious. It's not hard, just needs some effort, and we'd be happy to reopen this if you could get us something more along the lines of #13906 (comment). You can do that by starting with what you have and gradually removing stuff. Removing those imports, contracts, functions, function bodies, comments that can be removed while still triggering the bug. Otherwise it's difficult to even look into the problem because a repro of this size is hellish to debug.

[cameel]

Actually, let's keep this one open, even if the repro is subpar. It's clearly a bug, and we have no clear evidence it's a duplicate so we should at least retry this once we fix #13906.