| Message ID | 1457898620-1867-2-git-send-email-apinski@cavium.com |
|---|---|
| State | New |
| Headers | show |
On Mon, Mar 14, 2016 at 07:55:38AM +0100, Ard Biesheuvel wrote: > On 13 March 2016 at 20:50, Andrew Pinski <apinski@cavium.com> wrote: > > + movk x13, 0xe353, lsl 16 > > + lsr x11, x11, 3 > > + movk x13, 0x9ba5, lsl 32 > > + movk x13, 0x20c4, lsl 48 > > + /* x13 = 0x20c49ba5e353f7cf */ > > Could we clean this up a bit? Something along the lines of > > .set m, 0x20c49ba5e353f7cf > movz x13,#:abs_g3:m > movk x13, #:abs:g2_nc:m > movk x13, #:abs_g1_nc:m > movk x13, #:abs_g0_nc:m > > Actually, the movz/movk sequence should probably be implemented as a > macro in asm/assembler.h, with parameters for the register and the > symbol name. Agreed. > I think Mark proposed such a patch at some point That would be [1], which needs the relocations fixed up [2,3] to match the above. I didn't respin that as it turned out to be unnecessary at the time, but I'm more than happy for someone to pick it up. Mark. [1] http://lists.infradead.org/pipermail/linux-arm-kernel/2016-January/397563.html [2] http://lists.infradead.org/pipermail/linux-arm-kernel/2016-January/397572.html [3] http://lists.infradead.org/pipermail/linux-arm-kernel/2016-January/397573.html
On 13 March 2016 at 20:50, Andrew Pinski <apinski@cavium.com> wrote: > On many cores, udiv with a large value is slow, expand instead > the division out to be what GCC would have generated for the > divide by 1000. > > On ThunderX, the speeds up gettimeofday by 5%. > > Signed-off-by: Andrew Pinski <apinski@cavium.com> > --- > arch/arm64/kernel/vdso/gettimeofday.S | 20 ++++++++++++++++---- > 1 files changed, 16 insertions(+), 4 deletions(-) > > diff --git a/arch/arm64/kernel/vdso/gettimeofday.S b/arch/arm64/kernel/vdso/gettimeofday.S > index efa79e8..e5caef9 100644 > --- a/arch/arm64/kernel/vdso/gettimeofday.S > +++ b/arch/arm64/kernel/vdso/gettimeofday.S > @@ -64,10 +64,22 @@ ENTRY(__kernel_gettimeofday) > bl __do_get_tspec > seqcnt_check w9, 1b > > - /* Convert ns to us. */ > - mov x13, #1000 > - lsl x13, x13, x12 > - udiv x11, x11, x13 > + /* Undo the shift. */ > + lsr x11, x11, x12 > + > + /* Convert ns to us (division by 1000 by using multiply high). > + * This is how GCC converts the division by 1000 into. > + * This is faster than divide on most cores. > + */ > + mov x13, 63439 Please don't mix hex and decimal constants > + movk x13, 0xe353, lsl 16 > + lsr x11, x11, 3 > + movk x13, 0x9ba5, lsl 32 > + movk x13, 0x20c4, lsl 48 > + /* x13 = 0x20c49ba5e353f7cf */ Could we clean this up a bit? Something along the lines of .set m, 0x20c49ba5e353f7cf movz x13,#:abs_g3:m movk x13, #:abs:g2_nc:m movk x13, #:abs_g1_nc:m movk x13, #:abs_g0_nc:m Actually, the movz/movk sequence should probably be implemented as a macro in asm/assembler.h, with parameters for the register and the symbol name. I think Mark proposed such a patch at some point > + umulh x11, x11, x13 > + lsr x11, x11, 4 > + > stp x10, x11, [x0, #TVAL_TV_SEC] > 2: > /* If tz is NULL, return 0. */ > -- > 1.7.2.5 > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
diff --git a/arch/arm64/kernel/vdso/gettimeofday.S b/arch/arm64/kernel/vdso/gettimeofday.S index efa79e8..e5caef9 100644 --- a/arch/arm64/kernel/vdso/gettimeofday.S +++ b/arch/arm64/kernel/vdso/gettimeofday.S @@ -64,10 +64,22 @@ ENTRY(__kernel_gettimeofday) bl __do_get_tspec seqcnt_check w9, 1b - /* Convert ns to us. */ - mov x13, #1000 - lsl x13, x13, x12 - udiv x11, x11, x13 + /* Undo the shift. */ + lsr x11, x11, x12 + + /* Convert ns to us (division by 1000 by using multiply high). + * This is how GCC converts the division by 1000 into. + * This is faster than divide on most cores. + */ + mov x13, 63439 + movk x13, 0xe353, lsl 16 + lsr x11, x11, 3 + movk x13, 0x9ba5, lsl 32 + movk x13, 0x20c4, lsl 48 + /* x13 = 0x20c49ba5e353f7cf */ + umulh x11, x11, x13 + lsr x11, x11, 4 + stp x10, x11, [x0, #TVAL_TV_SEC] 2: /* If tz is NULL, return 0. */
On many cores, udiv with a large value is slow, expand instead the division out to be what GCC would have generated for the divide by 1000. On ThunderX, the speeds up gettimeofday by 5%. Signed-off-by: Andrew Pinski <apinski@cavium.com> --- arch/arm64/kernel/vdso/gettimeofday.S | 20 ++++++++++++++++---- 1 files changed, 16 insertions(+), 4 deletions(-) -- 1.7.2.5