Re: [PATCH v3] powerpc: Implement csum_ipv6_magic in assembly
On Thu, May 24, 2018 at 10:18:44AM +, Christophe Leroy wrote: > On 05/24/2018 06:20 AM, Christophe LEROY wrote: > >Le 23/05/2018 à 20:34, Segher Boessenkool a écrit : > >>On Tue, May 22, 2018 at 08:57:01AM +0200, Christophe Leroy wrote: > >>>The generic csum_ipv6_magic() generates a pretty bad result > >> > >> > >> > >>Please try with a more recent compiler, what you used is pretty ancient. > >>It's not like recent compilers do great on this either, but it's not > >>*that* bad anymore ;-) > > Here is what I get with GCC 8.1 > It doesn't look much better, does it ? There are no more mfocrf, which is a big speedup. Other than that it is pretty lousy still, I totally agree. This improvement happened quite a while ago, it's fixed in GCC 6. Segher
Re: [PATCH v3] powerpc: Implement csum_ipv6_magic in assembly
On Thu, May 24, 2018 at 08:20:16AM +0200, Christophe LEROY wrote: > Le 23/05/2018 à 20:34, Segher Boessenkool a écrit : > >On Tue, May 22, 2018 at 08:57:01AM +0200, Christophe Leroy wrote: > >>+_GLOBAL(csum_ipv6_magic) > >>+ lwz r8, 0(r3) > >>+ lwz r9, 4(r3) > >>+ lwz r10, 8(r3) > >>+ lwz r11, 12(r3) > >>+ addcr0, r5, r6 > >>+ adder0, r0, r7 > >>+ adder0, r0, r8 > >>+ adder0, r0, r9 > >>+ adder0, r0, r10 > >>+ adder0, r0, r11 > >>+ lwz r8, 0(r4) > >>+ lwz r9, 4(r4) > >>+ lwz r10, 8(r4) > >>+ lwz r11, 12(r4) > >>+ adder0, r0, r8 > >>+ adder0, r0, r9 > >>+ adder0, r0, r10 > >>+ adder0, r0, r11 > >>+ addze r0, r0 > >>+ rotlwi r3, r0, 16 > >>+ add r3, r0, r3 > >>+ not r3, r3 > >>+ rlwinm r3, r3, 16, 16, 31 > >>+ blr > >>+EXPORT_SYMBOL(csum_ipv6_magic) > > > >Clustering the loads and carry insns together is pretty much the worst you > >can do on most 32-bit CPUs. > > Oh, really ? __csum_partial is written that way too. I thought I told you about this before? Maybe not. > Right, now I tried interleaving the lwz and adde. I get no improvment at > all on a 885, but I get a 15% improvment on a 8321. It won't likely help on single-issue cores (like the one 885 has), yes. Segher
Re: [PATCH v3] powerpc: Implement csum_ipv6_magic in assembly
On 05/24/2018 06:20 AM, Christophe LEROY wrote: Le 23/05/2018 à 20:34, Segher Boessenkool a écrit : On Tue, May 22, 2018 at 08:57:01AM +0200, Christophe Leroy wrote: The generic csum_ipv6_magic() generates a pretty bad result Please try with a more recent compiler, what you used is pretty ancient. It's not like recent compilers do great on this either, but it's not *that* bad anymore ;-) Here is what I get with GCC 8.1 It doesn't look much better, does it ? net/ipv6/ip6_checksum.o: file format elf32-powerpc Disassembly of section .text: : 0: 94 21 ff f0 stwur1,-16(r1) 4: 80 04 00 00 lwz r0,0(r4) 8: 81 64 00 04 lwz r11,4(r4) c: 81 04 00 08 lwz r8,8(r4) 10: 93 e1 00 0c stw r31,12(r1) 14: 81 43 00 00 lwz r10,0(r3) 18: 83 e3 00 04 lwz r31,4(r3) 1c: 81 23 00 08 lwz r9,8(r3) 20: 81 83 00 0c lwz r12,12(r3) 24: 7c ea 3a 14 add r7,r10,r7 28: 7d 4a 38 10 subfc r10,r10,r7 2c: 7c ff 3a 14 add r7,r31,r7 30: 81 44 00 0c lwz r10,12(r4) 34: 7c 63 19 10 subfe r3,r3,r3 38: 7c 63 38 50 subfr3,r3,r7 3c: 7f ff 18 10 subfc r31,r31,r3 40: 7c e9 1a 14 add r7,r9,r3 44: 83 e1 00 0c lwz r31,12(r1) 48: 7c 63 19 10 subfe r3,r3,r3 4c: 38 21 00 10 addir1,r1,16 50: 7c 63 38 50 subfr3,r3,r7 54: 7d 29 18 10 subfc r9,r9,r3 58: 7d 2c 1a 14 add r9,r12,r3 5c: 7c 63 19 10 subfe r3,r3,r3 60: 7c 63 48 50 subfr3,r3,r9 64: 7d 8c 18 10 subfc r12,r12,r3 68: 7d 20 1a 14 add r9,r0,r3 6c: 7c 63 19 10 subfe r3,r3,r3 70: 7c 63 48 50 subfr3,r3,r9 74: 7c 00 18 10 subfc r0,r0,r3 78: 7d 2b 1a 14 add r9,r11,r3 7c: 7c 63 19 10 subfe r3,r3,r3 80: 7c 63 48 50 subfr3,r3,r9 84: 7d 6b 18 10 subfc r11,r11,r3 88: 7d 28 1a 14 add r9,r8,r3 8c: 7c 63 19 10 subfe r3,r3,r3 90: 7c 63 48 50 subfr3,r3,r9 94: 7d 08 18 10 subfc r8,r8,r3 98: 7d 2a 1a 14 add r9,r10,r3 9c: 7c 63 19 10 subfe r3,r3,r3 a0: 7c 63 48 50 subfr3,r3,r9 a4: 7d 4a 18 10 subfc r10,r10,r3 a8: 7d 23 2a 14 add r9,r3,r5 ac: 7c 63 19 10 subfe r3,r3,r3 b0: 7c 63 48 50 subfr3,r3,r9 b4: 7c a5 18 10 subfc r5,r5,r3 b8: 7c 63 32 14 add r3,r3,r6 bc: 7d 29 49 10 subfe r9,r9,r9 c0: 7d 29 18 50 subfr9,r9,r3 c4: 7c c6 48 10 subfc r6,r6,r9 c8: 7c 63 19 10 subfe r3,r3,r3 cc: 7c 63 48 50 subfr3,r3,r9 d0: 54 69 80 3e rotlwi r9,r3,16 d4: 7c 63 4a 14 add r3,r3,r9 d8: 7c 63 18 f8 not r3,r3 dc: 54 63 84 3e rlwinm r3,r3,16,16,31 e0: 4e 80 00 20 blr net/ipv6/ip6_checksum.o: file format elf64-powerpc Disassembly of section .text: <.csum_ipv6_magic>: 0: fb e1 ff f8 std r31,-8(r1) 4: 81 43 00 00 lwz r10,0(r3) 8: 81 83 00 04 lwz r12,4(r3) c: 81 23 00 08 lwz r9,8(r3) 10: 80 03 00 0c lwz r0,12(r3) 14: 7c e7 52 14 add r7,r7,r10 18: 80 64 00 08 lwz r3,8(r4) 1c: 81 04 00 00 lwz r8,0(r4) 20: 78 ff 00 20 clrldi r31,r7,32 24: 7c ec 3a 14 add r7,r12,r7 28: 81 64 00 04 lwz r11,4(r4) 2c: 7f ea f8 50 subfr31,r10,r31 30: 81 44 00 0c lwz r10,12(r4) 34: 7b ff 0f e0 rldicl r31,r31,1,63 38: 7c ff 3a 14 add r7,r31,r7 3c: eb e1 ff f8 ld r31,-8(r1) 40: 78 e4 00 20 clrldi r4,r7,32 44: 7c e9 3a 14 add r7,r9,r7 48: 7d 8c 20 50 subfr12,r12,r4 4c: 79 8c 0f e0 rldicl r12,r12,1,63 50: 7d 8c 3a 14 add r12,r12,r7 54: 79 87 00 20 clrldi r7,r12,32 58: 7d 80 62 14 add r12,r0,r12 5c: 7d 29 38 50 subfr9,r9,r7 60: 79 29 0f e0 rldicl r9,r9,1,63 64: 7d 29 62 14 add r9,r9,r12 68: 79 27 00 20 clrldi r7,r9,32 6c: 7d 28 4a 14 add r9,r8,r9 70: 7c 00 38 50 subfr0,r0,r7 74: 78 00 0f e0 rldicl r0,r0,1,63 78: 7c 00 4a 14 add r0,r0,r9 7c: 78 09 00 20 clrldi r9,r0,32 80: 7c 0b 02 14 add r0,r11,r0 84: 7d 08 48 50 subfr8,r8,r9 88: 79 08 0f e0 rldicl r8,r8,1,63 8c: 7d 08 02 14 add r8,r8,r0 90: 79 09 00 20 clrldi r9,r8,32 94: 7d 03 42 14 add r8,r3,r8 98: 7d 2b 48 50 subfr9,r11,r9 9c: 79 29 0f e0 rldicl r9,r9,1,63 a0: 7d 29 42 14 add r9,r9,r8 a4: 79 28 00 20 clrldi r8,r9,32 a8: 7d 2a 4a 14 add r9,r10,r9 ac: 7d 03 40 50 subfr8,r3,r8 b0: 79 08 0f e0 rldicl r8,r8,1,63 b4: 7d 08 4a 14 add r8,r8,r9 b8:
Re: [PATCH v3] powerpc: Implement csum_ipv6_magic in assembly
Le 23/05/2018 à 20:34, Segher Boessenkool a écrit : On Tue, May 22, 2018 at 08:57:01AM +0200, Christophe Leroy wrote: The generic csum_ipv6_magic() generates a pretty bad result Please try with a more recent compiler, what you used is pretty ancient. It's not like recent compilers do great on this either, but it's not *that* bad anymore ;-) --- a/arch/powerpc/lib/checksum_32.S +++ b/arch/powerpc/lib/checksum_32.S @@ -293,3 +293,36 @@ dst_error: EX_TABLE(51b, dst_error); EXPORT_SYMBOL(csum_partial_copy_generic) + +/* + * static inline __sum16 csum_ipv6_magic(const struct in6_addr *saddr, + * const struct in6_addr *daddr, + * __u32 len, __u8 proto, __wsum sum) + */ + +_GLOBAL(csum_ipv6_magic) + lwz r8, 0(r3) + lwz r9, 4(r3) + lwz r10, 8(r3) + lwz r11, 12(r3) + addcr0, r5, r6 + adder0, r0, r7 + adder0, r0, r8 + adder0, r0, r9 + adder0, r0, r10 + adder0, r0, r11 + lwz r8, 0(r4) + lwz r9, 4(r4) + lwz r10, 8(r4) + lwz r11, 12(r4) + adder0, r0, r8 + adder0, r0, r9 + adder0, r0, r10 + adder0, r0, r11 + addze r0, r0 + rotlwi r3, r0, 16 + add r3, r0, r3 + not r3, r3 + rlwinm r3, r3, 16, 16, 31 + blr +EXPORT_SYMBOL(csum_ipv6_magic) Clustering the loads and carry insns together is pretty much the worst you can do on most 32-bit CPUs. Oh, really ? __csum_partial is written that way too. Right, now I tried interleaving the lwz and adde. I get no improvment at all on a 885, but I get a 15% improvment on a 8321. Christophe Segher
Re: [PATCH v3] powerpc: Implement csum_ipv6_magic in assembly
On Tue, May 22, 2018 at 08:57:01AM +0200, Christophe Leroy wrote: > The generic csum_ipv6_magic() generates a pretty bad result Please try with a more recent compiler, what you used is pretty ancient. It's not like recent compilers do great on this either, but it's not *that* bad anymore ;-) > --- a/arch/powerpc/lib/checksum_32.S > +++ b/arch/powerpc/lib/checksum_32.S > @@ -293,3 +293,36 @@ dst_error: > EX_TABLE(51b, dst_error); > > EXPORT_SYMBOL(csum_partial_copy_generic) > + > +/* > + * static inline __sum16 csum_ipv6_magic(const struct in6_addr *saddr, > + * const struct in6_addr *daddr, > + * __u32 len, __u8 proto, __wsum sum) > + */ > + > +_GLOBAL(csum_ipv6_magic) > + lwz r8, 0(r3) > + lwz r9, 4(r3) > + lwz r10, 8(r3) > + lwz r11, 12(r3) > + addcr0, r5, r6 > + adder0, r0, r7 > + adder0, r0, r8 > + adder0, r0, r9 > + adder0, r0, r10 > + adder0, r0, r11 > + lwz r8, 0(r4) > + lwz r9, 4(r4) > + lwz r10, 8(r4) > + lwz r11, 12(r4) > + adder0, r0, r8 > + adder0, r0, r9 > + adder0, r0, r10 > + adder0, r0, r11 > + addze r0, r0 > + rotlwi r3, r0, 16 > + add r3, r0, r3 > + not r3, r3 > + rlwinm r3, r3, 16, 16, 31 > + blr > +EXPORT_SYMBOL(csum_ipv6_magic) Clustering the loads and carry insns together is pretty much the worst you can do on most 32-bit CPUs. Segher
