From: Aleksandar Rakic <aleksandar.rakic@htecgroup.com>

Skip NaN mode check for soft-float since NaN mode is irrelevant if an ELF
binary's FPU mode is soft-float, i.e. it doesn't utilize a FPU.

Cherry-picked 63492a56485f6b755fccf7ad623f7a189bfc79b6
from https://github.com/MIPS/gnutools-qemu

Signed-off-by: Faraz Shahbazker <fshahbazker@wavecomp.com>
Signed-off-by: Aleksandar Rakic <aleksandar.rakic@htecgroup.com>
---
 linux-user/mips/cpu_loop.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/linux-user/mips/cpu_loop.c b/linux-user/mips/cpu_loop.c
index 462387a073..07c1ebe287 100644
--- a/linux-user/mips/cpu_loop.c
+++ b/linux-user/mips/cpu_loop.c
@@ -304,8 +304,10 @@ void target_cpu_copy_regs(CPUArchState *env, struct target_pt_regs *regs)
     if (env->insn_flags & ISA_NANOMIPS32) {
         return;
     }
-    if (((info->elf_flags & EF_MIPS_NAN2008) != 0) !=
-        ((env->active_fpu.fcr31 & (1 << FCR31_NAN2008)) != 0)) {
+    if (info->fp_abi != MIPS_ABI_FP_SOFT
+        && ((info->elf_flags & EF_MIPS_NAN2008) != 0) !=
+           ((env->active_fpu.fcr31 & (1 << FCR31_NAN2008)) != 0))
+      {
         if ((env->active_fpu.fcr31_rw_bitmask &
               (1 << FCR31_NAN2008)) == 0) {
             fprintf(stderr, "ELF binary's NaN mode not supported by CPU\n");
-- 
2.34.1

On Wed, 26 Feb 2025 at 17:03, Aleksandar Rakic
<aleksandar.rakic@htecgroup.com> wrote:
>
> From: Aleksandar Rakic <aleksandar.rakic@htecgroup.com>
>
> Skip NaN mode check for soft-float since NaN mode is irrelevant if an ELF
> binary's FPU mode is soft-float, i.e. it doesn't utilize a FPU.
>
> Cherry-picked 63492a56485f6b755fccf7ad623f7a189bfc79b6
> from https://github.com/MIPS/gnutools-qemu
>
> Signed-off-by: Faraz Shahbazker <fshahbazker@wavecomp.com>
> Signed-off-by: Aleksandar Rakic <aleksandar.rakic@htecgroup.com>
> ---
>  linux-user/mips/cpu_loop.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/linux-user/mips/cpu_loop.c b/linux-user/mips/cpu_loop.c
> index 462387a073..07c1ebe287 100644
> --- a/linux-user/mips/cpu_loop.c
> +++ b/linux-user/mips/cpu_loop.c
> @@ -304,8 +304,10 @@ void target_cpu_copy_regs(CPUArchState *env, struct target_pt_regs *regs)
>      if (env->insn_flags & ISA_NANOMIPS32) {
>          return;
>      }
> -    if (((info->elf_flags & EF_MIPS_NAN2008) != 0) !=
> -        ((env->active_fpu.fcr31 & (1 << FCR31_NAN2008)) != 0)) {
> +    if (info->fp_abi != MIPS_ABI_FP_SOFT
> +        && ((info->elf_flags & EF_MIPS_NAN2008) != 0) !=
> +           ((env->active_fpu.fcr31 & (1 << FCR31_NAN2008)) != 0))
> +      {
>          if ((env->active_fpu.fcr31_rw_bitmask &
>                (1 << FCR31_NAN2008)) == 0) {
>              fprintf(stderr, "ELF binary's NaN mode not supported by CPU\n");

Unless I'm misreading the code, the kernel's MIPS ELF loader
does not look at the fp_abi to decide how to handle the
NAN2008 bit:

https://elixir.bootlin.com/linux/v6.13.6/source/arch/mips/kernel/elf.c#L154

I think QEMU should handle this bit the same way as the kernel's
loader.

thanks
-- PMM

Hi,

The soft(-float) requirement means that the program being loaded has no
FPU dependency at all (i.e. it has no FPU instructions).
https://elixir.bootlin.com/linux/v6.13.6/source/arch/mips/kernel/elf.c#L34

When -msoft-float is used, the processor does not use hardware
floating-point instructions but relies on software emulation for
floating-point arithmetic. In that case, a NaN mode (such as NaN2008)
in the hardware FPU has no effect because the hardware FPU is not
active and all floating-point expressions and operations are handled by
functions from software libraries, such as those within glibc or other
emulation libraries. How NaN values ​​are handled in software emulation
depends on the implementation of those libraries and is not related to
FPU registers.

If the processor uses -msingle-float, NaN2008 mode can affect the way
NaN values ​​are encoded and interpreted. When -msoft-float is active,
the emulation ignores the hardware FPU and the NaN mode is not relevant
because floating-point processing is left to the software
implementation, not the hardware.

Kind regards,
Aleksandar Rakic

On Tue, 18 Mar 2025 at 13:00, Aleksandar Rakic
<aleksandar.rakic@htecgroup.com> wrote:
> The soft(-float) requirement means that the program being loaded has no
> FPU dependency at all (i.e. it has no FPU instructions).
> https://elixir.bootlin.com/linux/v6.13.6/source/arch/mips/kernel/elf.c#L34

Yes, I know. But the kernel loader does not use this as part
of the NaN mode checking, so we should not either. We are
supposed to behave the same way the kernel does.

thanks
-- PMM

Thank you.

Kind regards,
Aleksandar Rakic

Hi,

If ieee754 equals EMULATED, then the variables mips_use_nan_2008 and
mips_use_nan_legacy are set to true
( https://elixir.bootlin.com/linux/v6.13.6/source/arch/mips/kernel/fpu-probe.c#L208 )
and any binaries are accepted regardless of whether supported by the FPU
( https://elixir.bootlin.com/linux/v6.13.6/source/arch/mips/kernel/elf.c#L154 ).

Kind regards,
Aleksandar Rakic

On Tue, 18 Mar 2025 at 11:31, Aleksandar Rakic
<aleksandar.rakic@htecgroup.com> wrote:
>
> Hi,
>
> If ieee754 equals EMULATED, then the variables mips_use_nan_2008 and
> mips_use_nan_legacy are set to true
> ( https://elixir.bootlin.com/linux/v6.13.6/source/arch/mips/kernel/fpu-probe.c#L208 )
> and any binaries are accepted regardless of whether supported by the FPU
> ( https://elixir.bootlin.com/linux/v6.13.6/source/arch/mips/kernel/elf.c#L154 )

I presume from the name of that setting that in that case a
real kernel is doing trap-and-emulate for binaries of the
wrong FPU type (which must perform appallingly badly...)  ?

Does QEMU set up the CPU for linux-user mode so that it
effectively does emulation as-if both NaN modes were supported?
I can't see that happening from a quick grep.

Also, this doesn't seem to be related to the ABI_FP_SOFT setting.

thanks
-- PMM

Hi,

Could you please let us know if you have any comments
on the latest version of this patch?

Kind regards,
Aleksandar Rakic