On Mon, 31 Jul 2017, Manuel Lauss wrote:
> I'm unsure whether the patch is really correct, due to the unexpected
> binary size reduction. I do not have a hardfloat mips32 userland at hand
> therefore I'd appreciate it if someone could test it!
What exactly are you unsure about? Double operations using odd register
indices in the 32-bit FPR mode are architecturally unpredictable. Is this
what concerns you?
> diff --git a/arch/mips/math-emu/cp1emu.c b/arch/mips/math-emu/cp1emu.c
> index f08a7b4facb9..f3a9bf5285e0 100644
> --- a/arch/mips/math-emu/cp1emu.c
> +++ b/arch/mips/math-emu/cp1emu.c
> @@ -830,12 +830,12 @@ do { \
> } while (0)
>
> #define DIFROMREG(di, x) \
> - ((di) = get_fpr64(&ctx->fpr[(x) & ~(cop1_64bit(xcp) == 0)], 0))
> + ((di) = get_fpr64(&ctx->fpr[(x) & ~(cop1_64bit(xcp) == 0 ? 1 : 0)], 0))
Umm, I find the expression somewhat hard to follow. How about:
((di) = get_fpr64(&ctx->fpr[(x) & ~(cop1_64bit(xcp) ^ 1))
?
> #define DITOREG(di, x) \
> do { \
> unsigned fpr, i; \
> - fpr = (x) & ~(cop1_64bit(xcp) == 0); \
> + fpr = (x) & ~(cop1_64bit(xcp) == 0 ? 1 : 0); \
Likewise:
fpr = (x) & ~(cop1_64bit(xcp) ^ 1); \
?
Maciej
