diff options
| author | Peter Maydell <peter.maydell@linaro.org> | 2024-11-05 10:09:53 +0000 |
|---|---|---|
| committer | Peter Maydell <peter.maydell@linaro.org> | 2024-11-05 10:09:53 +0000 |
| commit | d1ff996788a41280e2e0213b9571afeca4d6ca90 (patch) | |
| tree | d2253df5e538d9731c2056e967b536245b1e0dce /linux-user | |
| parent | d22c9949d73d714e1ea2caf033131d882ad1d66a (diff) | |
target/arm: Explicitly set 2-NaN propagation rule
Set the 2-NaN propagation rule explicitly in the float_status words
we use. We wrap this plus the pre-existing setting of the
tininess-before-rounding flag in a new function
arm_set_default_fp_behaviours() to avoid repetition, since we have a
lot of float_status words at this point.
The situation with FPA11 emulation in linux-user is a little odd, and
arguably "correct" behaviour there would be to exactly match a real
Linux kernel's FPA11 emulation. However FPA11 emulation is
essentially dead at this point and so it seems better to continue
with QEMU's current behaviour and leave a comment describing the
situation.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20241025141254.2141506-4-peter.maydell@linaro.org
Diffstat (limited to 'linux-user')
| -rw-r--r-- | linux-user/arm/nwfpe/fpa11.c | 18 |
1 files changed, 18 insertions, 0 deletions
diff --git a/linux-user/arm/nwfpe/fpa11.c b/linux-user/arm/nwfpe/fpa11.c index 9a93610d24..8356beb52c 100644 --- a/linux-user/arm/nwfpe/fpa11.c +++ b/linux-user/arm/nwfpe/fpa11.c @@ -51,6 +51,24 @@ void resetFPA11(void) #ifdef MAINTAIN_FPCR fpa11->fpcr = MASK_RESET; #endif + + /* + * Real FPA11 hardware does not handle NaNs, but always takes an + * exception for them to be software-emulated (ARM7500FE datasheet + * section 10.4). There is no documented architectural requirement + * for NaN propagation rules and it will depend on how the OS + * level software emulation opted to do it. We here use prop_s_ab + * which matches the later VFP hardware choice and how QEMU's + * fpa11 emulation has worked in the past. The real Linux kernel + * does something slightly different: arch/arm/nwfpe/softfloat-specialize + * propagateFloat64NaN() has the curious behaviour that it prefers + * the QNaN over the SNaN, but if both are QNaN it picks A and + * if both are SNaN it picks B. In theory we could add this as + * a NaN propagation rule, but in practice FPA11 emulation is so + * close to totally dead that it's not worth trying to match it at + * this late date. + */ + set_float_2nan_prop_rule(float_2nan_prop_s_ab, &fpa11->fp_status); } void SetRoundingMode(const unsigned int opcode) |