Problems with C variadics

[bjoernQ]

This relates to esp-rs/esp-wifi-sys#16

In https://github.com/bjoernQ/xtensa-rust-variadics-problem-repo I have some very simple code to reproduce the problem.

What I do there

• Rust calls function in C
• That function calls a function in Rust with variadics

On RISCV it works as expected and outputs

42 "TAG" "FORMAT"
1 = 1
2 = 2
3 = 3
4 = 4
5 = 5
6 = 6
7 = 7
8 = 8
9 = 9

On ESP32 it outputs

42 "TAG" "FORMAT"
1 = 1061160316
2 = 1
3 = 2
4 = 3
5 = 2148373693
6 = 1073594208
7 = 1074266112
8 = 1073414160
9 = 4

[karlri]

I implemented the shortest solution I could come up with. It's not a full solution and it's not implemented in the proper place but it's provided in the case someone finds it useful.

esp-1.72.1.0...karlri:rust:esp-1.72.1.0

Ideally the compiler should be updated to not use the llvm intrinsic va_arg, or the va_arg intrinsic in llvm should be fixed, but I'm not competent enough to implement it where it should be implemented.

[karlri]

I think I traced this bug back to its source.

I believe the bug to originate from here on line 340:
https://github.com/espressif/llvm-project/blob/xtensa_release_16.0.4/llvm/lib/Target/Xtensa/XtensaISelLowering.cpp

Can't Expand ISD::VAARG since we have a custom struct, we need Custom implementation.
It would need to routed around line 1924

  case ISD::VAARG:
    return LowerVAARG(Op, DAG);

finally, the big job would be to implement
SDValue XtensaTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const

[plietar]

I've opened a PR at #201 which should fix this. With it the code in the reproduction repository works as intended, and so does the esp-wifi syslog facility.

[MabezDev]

Closed with @plietar's patch in 1.73.0.1, thanks again @plietar! Note that not all builds are available, but by the end of the day they should be available :).

[thomas725]

Hello @bjoernQ @plietar @MabezDev - your solution for this ticket also solved esp-rs/esp-idf-sys#212 for me - sadly it broke again with esp-rs compiler between 1.84 (still works) and 1.85 (doesn't work again just as it didn't before 1.73.0.1)

Can you help me resolve this? The bug returning seems to coincide with the type c_char alias being changed from i8 to u8.

[MabezDev]

@thomas725 We have it working in 1.93 (we haven't tested every release, but it works there), however its not currently working in the upcoming 1.94 release - investigating it currently.