wasm2c: Add big endian support.

[alexrp]

I took a slightly unconventional approach to detecting endianness here. We have no compiler/platform-specific preprocessor checks in the stage1 C code today, and I think that's a property worth maintaining.

With this patch, I'm able to successfully run s390x-linux-gnu-gcc ../stage1/wasm2c.c -o zig-wasm2c && qemu-s390x ./zig-wasm2c ../stage1/zig1.wasm zig1.c && s390x-linux-gnu-gcc -std=c99 -Os zig1.c ../stage1/wasi.c -o zig1 && qemu-s390x ../lib zig1 build-exe -ofmt=c -lc -OReleaseSmall ... -femit-bin=zig2.c.

[andrewrk]

since the same computer will be generating this C code and running it, it would be better to detect endianness and then generate code that does not branch.

in fact that seems to be how @jacobly0 set you up, all you had to do was one detection and populate a bool... why not just do that?

[alexrp]

It just seemed like a bit of a shame to generate target-specific C code when doing the check in the generated code like this means that the output is target-agnostic.

But if that's a property we don't care about, then sure, I can move the check into the generator.

[andrewrk]

The only goal of that code is to bootstrap on the current host, as fast as possible.

[alexrp]

Ok. Any preference for what to do about wasi.c? My inclination would be to have it call into the load/store helpers generated by wasm2c if we do the detection in the generator.

[alexrp]

Ok, switched the approach to checking in the generator.

It just seemed like a bit of a shame to generate target-specific C code when doing the check in the generated code like this means that the output is target-agnostic.

One additional, minor point here is that, while hacking on wasm2c, it's just convenient to know that it doesn't matter what target the generator itself is built for, as it'll produce usable code for any target. It became a bit more convoluted when I had to build zig-wasm2c for s390x and run it under QEMU too.

[andrewrk]

How about making it an overridable preprocessor define?

[alexrp]

How about making it an overridable preprocessor define?

How about giving wasm2c an optional third command line argument to specify endianness? I think that's a tiny bit nicer since it maintains the ability to just invoke ninja zig-wasm2c to build it.

[andrewrk]

Seems fine to me. My main protest is that duration of the bootstrap is the top priority.

[alexrp]

Ok, done.