expose CPU and features as part of the target

[justinbalexander]

When building for a Cortex-M4F there is currently no way to tell LLVM to use the floating point unit on the processor.

With clang you can pass the option -mfpu=fpv4-sp-d16 and -mfloat-abi=hard and it will output the correct assembly and use the FPU.

Using the options "-mllvm -float-abi=hard -mllvm -fpu=fpv4-sp-d16" gives an llvm error unless I am mistaken.

Regardless, having a zig native way to specify those options would be nice.

[andrewrk]

Thanks for typing up the issue. Float-abi = hard should work by specifying gnueabihf, eabihf, or musleabihf` for the target ABI. Is that not working?

As for -fpu=fpv4-sp-d16, I think this is a general problem of exposing CPU args/features. I'm guessing here:

zig/src/codegen.cpp

Lines 7830 to 7835 in f2ca2ac

	target_specific_cpu_args = "";
	target_specific_features = "";
	}
	g->target_machine = LLVMCreateTargetMachine(target_ref, buf_ptr(&g->triple_str),
	target_specific_cpu_args, target_specific_features, opt_level, reloc_mode, LLVMCodeModelDefault);

This is something I haven't looked into yet to know the best way to do it, but I'd accept a pull request that provides a workaround exposing this option to the CLI (and the build system which is mostly in std/build.zig).

[justinbalexander]

I figured that's what the hf was doing, but I couldn't see the end result without the fpu being specified and functional. Thanks.

…

On Mon, Mar 11, 2019 at 3:15 PM Andrew Kelley ***@***.***> wrote: Thanks for typing up the issue. Float-abi = hard should work by specifying gnueabihf, eabihf, or musleabihf` for the target ABI. Is that not working? As for -fpu=fpv4-sp-d16, I think this is a general problem of exposing CPU args/features. I'm guessing here: https://github.com/ziglang/zig/blob/f2ca2ace0943a81e0fc5b59347feac177a71eb6c/src/codegen.cpp#L7830-L7835 This is something I haven't looked into yet to know the best way to do it, but I'd accept a pull request that provides a workaround exposing this option to the CLI (and the build system which is mostly in std/build.zig). — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#2051 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/Alxei4XkGC6-C5QLHh1_rSV4F16mx3WQks5vVrlhgaJpZM4bpczg> .

[andrewrk]

I think this issue may be a bit difficult for drive-by contributors. I'll repeat the comment I made on a pull request here:

Although LLVM is currently the only backend of Zig, the language does not depend on LLVM, and it is planned to support non-LLVM backends. Even though the target cpu and features options need to eventually be given to LLVM in a way that LLVM understands, we still need to expose these options in a way that is independent of LLVM. That means the CLI option cannot be called --llvm-cpu or --llvm- anything. zig targets must list all the possible options. The CLI should be able to reject invalid options without consulting LLVM.

I recognize this is in direct contradiction with my earlier comment, and I apologize for that. Sometimes it takes me a while to realize the path forward.

[andrewrk]

I'm not sure I'll be able to do it in time for 0.5.0, but I'll give this a slight priority bump, at the expense of other issues in the 0.5.0 milestone.

[andrewrk]

I accidentally made a duplicate issue here: #2883

But since that issue has the checklist, I'll make this one refer to that one. I consider this a high priority issue for this milestone.