[RISC-V] Clang behaves differently from gcc when `-mcmodel=large` and triple is riscv-32

[amitch1999]

When compiling something with clang and -march=rv32i -mcmodel=large I get an error: "unsupported argument 'large' to option '-mcmodel=' for target 'riscv32-unknown-elf'"
If I do so with gcc, it seems to just go with medlow code model.

Should this behavior be changed in clang? to match gcc? maybe just a warning is enough

I encountered this when trying to compile newlib with clang, some of my multilibs are 32 bit, and some are 64 bit, but I exported CFLAGS_FOR_TARGET=-mcmodel=large because I want the 64 bit libraries to be compiled with large code model. when building like this with gcc it works fine.
Maybe I can configure newlib somehow to have different flags for 32 bit libs vs 64 bit libs? I did not find something like that.
This behavior might limit clang when building newlib with multilibs.

[topperc]

cc: @kito-cheng

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: None (amitch1999)

When compiling something with clang and `-march=rv32i -mcmodel=large` I get an error: "unsupported argument 'large' to option '-mcmodel=' for target 'riscv32-unknown-elf'" If I do so with gcc, it seems to just go with `medlow` code model.

Should this behavior be changed in clang? to match gcc? maybe just a warning is enough

I encountered this when trying to compile newlib with clang, some of my multilibs are 32 bit, and some are 64 bit, but I exported CFLAGS_FOR_TARGET=-mcmodel=large because I want the 64 bit libraries to be compiled with large code model. when building like this with gcc it works fine.
Maybe I can configure newlib somehow to have different flags for 32 bit libs vs 64 bit libs? I did not find something like that.
This behavior might limit clang when building newlib with multilibs.

[MaskRay]

(Reporting an error looks a reasonable behavior to me. x86 gcc -m32 -mcmodel=large has a similar error. It's possible
that GCC should be fixed instead.)

[lenary]

I think the error is reasonable and correct. If a user asks for the mcmodel=large on rv32, then we shouldn't just ignore what they asked and do something else; we should tell the user they asked for something that isn't available/defined.

[kito-cheng]

I will add the error message on the GCC side, but maybe next month because GCC 15 is not release yet

[amitch1999]

I agree with the error, and these are some good points, clang should not just ignore and user and do something else.
But I think it might limit building newlib with multilibs (32 and 64 together), maybe because of something similar to this issue gcc is doing if (riscv_cmodel == CM_LARGE && !TARGET_64BIT) riscv_cmodel = CM_MEDLOW;
But maybe that is something to add to newlib, the ability to have different C_FLAGS for different multilibs (unless it already exists and I missed it)

[kito-cheng]

Multilib is another story, which is not really belong newlib, more like GCC multlib issue...and I found we forgot to upstream one of our downstream change on the multilib generator on the gcc site, just commit that into GCC 14 and trunk.

You can configure gcc with --with-multilib-generator="rv64ima-lp64--;--cmodel=large"