Re: [musl] [PATCH] riscv: Fix setjmp assembly when compiling for ilp32f/lp64f.

["Alex Rønne Petersen" <alex@alexrp.com>]

On Wed, Jul 24, 2024 at 1:22 AM Rich Felker <dalias@libc.org> wrote:
>
> On Wed, Jul 24, 2024 at 01:12:33AM +0200, Alex Rønne Petersen wrote:
> > On Wed, Jul 24, 2024 at 12:58 AM Rich Felker <dalias@libc.org> wrote:
> > >
> > > On Wed, Jul 24, 2024 at 12:47:14AM +0200, Alex Rønne Petersen wrote:
> > > > On Tue, Jul 23, 2024 at 11:22 PM Szabolcs Nagy <nsz@port70.net> wrote:
> > > > >
> > > > > * Alex Rønne Petersen <alex@alexrp.com> [2024-06-29 04:04:34 +0200]:
> > > > > > To keep things simple, I just changed the instruction mnemonics appropriately,
> > > > > > rather than adding complexity by changing the buffer size/offsets based on ABI.
> > > > > >
> > > > > > Signed-off-by: Alex Rønne Petersen <alex@alexrp.com>
> > > > >
> > > > > fwiw this looks good to me.
> > > > >
> > > > > the only weirdness is that the math code uses __riscv_flen
> > > > > and this code __riscv_float_abi*. i don't know if there
> > > > > is semantic difference.
> > > >
> > > > `__riscv_flen` tells you the width of the FP registers on the target
> > > > CPU. This is semantically distinct from `__riscv_float_abi`. For
> > > > example, while it would probably be a bit silly, there's no particular
> > > > reason why I couldn't target the LP64F ABI on an RV64IMAFDC machine.
> > > > In that case, no code needs to concern itself with the upper bits of
> > > > the FP registers.
> > > >
> > > > I took a quick peek at some of the `__riscv_flen` checks in musl. They
> > > > look ok. They're checking the capabilities of the machine for the
> > > > purposes of performing a computation; they're not making ABI
> > > > decisions. In my silly example above, if I tell the compiler to do so
> > > > with `-march=rv64...d`, it would theoretically be fine for the
> > > > compiler to generate double-precision float instructions for
> > > > computations as long as values are passed/returned according to LP64F
> > > > rules.
> > >
> > > If you're building code for -sf or -sp ABI, but could run on a machine
> > > with larger floating point register file, it's possible that the user
> > > could have libc built not to use fp registers at all or only 32-bit
> > > registers (respectively), but the calling application could be built
> > > for and running on a machine with 64-bit registers. In this case we
> > > need to understand what the ABI says. Are the 64-bit register, if
> > > present, call-saved in lower ABI levels where they don't participate
> > > in the calling convention? If so, no #ifdef is sufficient and there
> > > must be a runtime hwcap check here to determine which form of
> > > save/restore to do, like on arm and powerpc.
> >
> > I don't think this is a scenario that the ABI considers to be
> > supported. If you try to link code of different ABIs, you will get
> > linker errors such as:
> >
> >     /tmp/ccz2Y86f.o: can't link soft-float modules with double-float modules
> >
> > or
> >
> >     /tmp/cc7rTh7R.o: can't link single-float modules with double-float modules
>
> Those are different ABIs. You can't link modules with mismatched ABI,
> but you should be able to link modules that are both using -sf ABI (or
> both using -sp ABI), where one is not using the fpu and the other is
> using the full double fpu but only passing args in GPRs to conform
> with the ABI. If that's not allowed, I would consider it a tooling
> bug; there's no compatibility-constraint reason it can't be allowed.

Oh, of course. I misread your question, sorry.

Here's, I think, the relevant section of the calling convention:
https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc#floating-point-register-convention

This part is a bit awkward (to me, at least):

    Floating-point values in callee-saved registers are only preserved
across calls if they are no larger than the width of a floating-point
register in the targeted ABI. Therefore, these registers can always be
considered temporaries if targeting the base integer calling
convention.

I'm not really sure why they're talking about "values" there; I would
think the register width (in the machine vs the ABI) is the only thing
we're concerned about in this context. I'm assuming that what they
mean is:

    Floating-point registers in the callee-saved set are only
preserved across calls if they are no larger than the width of a
floating-point register in the targeted ABI.