Help needed in debugging AVR compilation errors

[deadprogram]

I am working on the I2C interface for the AVR. The WIP branch is located here:

hybridgroup@f3c8e28

My problem is that the compile works with the code I just committed, but if I uncomment this line https://github.com/hybridgroup/tinygo/blob/feature/avr-i2c/src/examples/i2clcd/i2clcd.go#L52 in my sample, the compile fails with:

$ make flash-i2clcd TARGET=arduino                                                                                     
./build/tgo build -target arduino -size=short -o build/i2clcd.elf examples/i2clcd
ROLW unimplemented
UNREACHABLE executed at /home/ron/.gvm/pkgsets/go1.11/global/src/github.com/aykevl/llvm/lib/Target/AVR/AVRExpandPseudo$
nsts.cpp:1316!
SIGABRT: abort
PC=0x7f7ec3df0428 m=4 sigcode=18446744073709551610

If I uncomment the following line instead https://github.com/hybridgroup/tinygo/blob/feature/avr-i2c/src/examples/i2clcd/i2clcd.go#L53

I get this error:

$ make flash-i2clcd TARGET=arduino
go build -o build/tgo -i .
./build/tgo build -target arduino -size=short -o build/i2clcd.elf examples/i2clcd
tgo: /home/ron/.gvm/pkgsets/go1.11/global/src/github.com/aykevl/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp:9
039: void llvm::SelectionDAGISel::LowerArguments(const llvm::Function&): Assertion `InVals.size() == Ins.size() && "Low
erFormalArguments didn't emit the correct number of values!"' failed.
SIGABRT: abort
PC=0x7f633a6f8428 m=5 sigcode=18446744073709551610

Perhaps this has to do with not yet implemented features on AVR?

[aykevl]

I have seen the second error more often. It is due to a limitation in the AVR backend (it cannot deal with struct parameters, and []byte is really a {ptr, len, cap} struct). See #20 for details. A temporary (ugly) workaround is to make the inliner more likely to run by increasing the value of the third parameter to Optmize(opt, size, inliner) in main.go.

I haven't seen the first error before but it also looks like an unimplemented feature.

[deadprogram]

Boosted Optimize() 3 param as you suggested. the value 300 compiles (both errors go away) but does not execute as desired.

You are probably not surprised. Digging here to figure out more.

[aykevl]

You could try again now. The second problem is probably fixed with fdfa810. The first problem should really be fixed in upstream LLVM, but might also be fixed as a side effect (who knows).

A trick that might fix the first problem (again, a dirty workaround) is to sleep the same time everywhere. The optimizer will recognize the pattern and do the expensive int64 calculation at compile time.

[aykevl]

Ah now I see. The trouble with the second uncommented line might be related to the memory allocator, which I have never tested on AVR (as it wasn't able to run any significant programs).

[deadprogram]

You were correct that changing all of the calls to time.Sleep() to use the same value let the compiler run without errors and without changing the value of the inliner param to Optimize(). That is probably not a practical long-term solution, of course.

The latest version of master lets the compiler run using various different time values like my original code, with the inline param set to the lower value 50.

[deadprogram]

The latest version of master now is able to run the AVR I2C code with no change to the Optimize() function. I am going to close this issue as the original problem has been addressed. Thank you!