| Message ID | 20171205233135.1763-1-james.hogan@mips.com (mailing list archive) |
|---|---|
| State | Accepted |
| Delegated to: | Herbert Xu |
| Headers | show |
On Tue, Dec 05, 2017 at 11:31:35PM +0000, James Hogan wrote: > From: James Hogan <jhogan@kernel.org> > > Current MIPS64r6 toolchains aren't able to generate efficient > DMULU/DMUHU based code for the C implementation of umul_ppmm(), which > performs an unsigned 64 x 64 bit multiply and returns the upper and > lower 64-bit halves of the 128-bit result. Instead it widens the 64-bit > inputs to 128-bits and emits a __multi3 intrinsic call to perform a 128 > x 128 multiply. This is both inefficient, and it results in a link error > since we don't include __multi3 in MIPS linux. > > For example commit 90a53e4432b1 ("cfg80211: implement regdb signature > checking") merged in v4.15-rc1 recently broke the 64r6_defconfig and > 64r6el_defconfig builds by indirectly selecting MPILIB. The same build > errors can be reproduced on older kernels by enabling e.g. CRYPTO_RSA: > > lib/mpi/generic_mpih-mul1.o: In function `mpihelp_mul_1': > lib/mpi/generic_mpih-mul1.c:50: undefined reference to `__multi3' > lib/mpi/generic_mpih-mul2.o: In function `mpihelp_addmul_1': > lib/mpi/generic_mpih-mul2.c:49: undefined reference to `__multi3' > lib/mpi/generic_mpih-mul3.o: In function `mpihelp_submul_1': > lib/mpi/generic_mpih-mul3.c:49: undefined reference to `__multi3' > lib/mpi/mpih-div.o In function `mpihelp_divrem': > lib/mpi/mpih-div.c:205: undefined reference to `__multi3' > lib/mpi/mpih-div.c:142: undefined reference to `__multi3' > > Therefore add an efficient MIPS64r6 implementation of umul_ppmm() using > inline assembly and the DMULU/DMUHU instructions, to prevent __multi3 > calls being emitted. > > Fixes: 7fd08ca58ae6 ("MIPS: Add build support for the MIPS R6 ISA") > Signed-off-by: James Hogan <jhogan@kernel.org> > Cc: Ralf Baechle <ralf@linux-mips.org> > Cc: Herbert Xu <herbert@gondor.apana.org.au> > Cc: "David S. Miller" <davem@davemloft.net> > Cc: linux-mips@linux-mips.org > Cc: linux-crypto@vger.kernel.org > --- > Please can somebody apply this fix for v4.15, as the MIPS 64r6_defconfig > and 64r6el_defconfig builds are broken without it. I can take this but I'd like to see an ack from someone on the MIPS side. Thanks,
On Fri, Dec 22, 2017 at 06:08:08PM +1100, Herbert Xu wrote: > I can take this but I'd like to see an ack from someone on the > MIPS side. Sure: Acked-by: Ralf Baechle <ralf@linux-mips.org> Ralf
On Fri, Dec 22, 2017 at 08:59:21AM +0100, Ralf Baechle wrote: > On Fri, Dec 22, 2017 at 06:08:08PM +1100, Herbert Xu wrote: > > > I can take this but I'd like to see an ack from someone on the > > MIPS side. > > Sure: > > Acked-by: Ralf Baechle <ralf@linux-mips.org> Thanks Ralf!
On Tue, Dec 05, 2017 at 11:31:35PM +0000, James Hogan wrote: > From: James Hogan <jhogan@kernel.org> > > Current MIPS64r6 toolchains aren't able to generate efficient > DMULU/DMUHU based code for the C implementation of umul_ppmm(), which > performs an unsigned 64 x 64 bit multiply and returns the upper and > lower 64-bit halves of the 128-bit result. Instead it widens the 64-bit > inputs to 128-bits and emits a __multi3 intrinsic call to perform a 128 > x 128 multiply. This is both inefficient, and it results in a link error > since we don't include __multi3 in MIPS linux. > > For example commit 90a53e4432b1 ("cfg80211: implement regdb signature > checking") merged in v4.15-rc1 recently broke the 64r6_defconfig and > 64r6el_defconfig builds by indirectly selecting MPILIB. The same build > errors can be reproduced on older kernels by enabling e.g. CRYPTO_RSA: > > lib/mpi/generic_mpih-mul1.o: In function `mpihelp_mul_1': > lib/mpi/generic_mpih-mul1.c:50: undefined reference to `__multi3' > lib/mpi/generic_mpih-mul2.o: In function `mpihelp_addmul_1': > lib/mpi/generic_mpih-mul2.c:49: undefined reference to `__multi3' > lib/mpi/generic_mpih-mul3.o: In function `mpihelp_submul_1': > lib/mpi/generic_mpih-mul3.c:49: undefined reference to `__multi3' > lib/mpi/mpih-div.o In function `mpihelp_divrem': > lib/mpi/mpih-div.c:205: undefined reference to `__multi3' > lib/mpi/mpih-div.c:142: undefined reference to `__multi3' > > Therefore add an efficient MIPS64r6 implementation of umul_ppmm() using > inline assembly and the DMULU/DMUHU instructions, to prevent __multi3 > calls being emitted. > > Fixes: 7fd08ca58ae6 ("MIPS: Add build support for the MIPS R6 ISA") > Signed-off-by: James Hogan <jhogan@kernel.org> > Cc: Ralf Baechle <ralf@linux-mips.org> > Cc: Herbert Xu <herbert@gondor.apana.org.au> > Cc: "David S. Miller" <davem@davemloft.net> > Cc: linux-mips@linux-mips.org > Cc: linux-crypto@vger.kernel.org Patch applied. Thanks.
diff --git a/lib/mpi/longlong.h b/lib/mpi/longlong.h index 57fd45ab7af1..08c60d10747f 100644 --- a/lib/mpi/longlong.h +++ b/lib/mpi/longlong.h @@ -671,7 +671,23 @@ do { \ ************** MIPS/64 ************** ***************************************/ #if (defined(__mips) && __mips >= 3) && W_TYPE_SIZE == 64 -#if (__GNUC__ >= 5) || (__GNUC__ >= 4 && __GNUC_MINOR__ >= 4) +#if defined(__mips_isa_rev) && __mips_isa_rev >= 6 +/* + * GCC ends up emitting a __multi3 intrinsic call for MIPS64r6 with the plain C + * code below, so we special case MIPS64r6 until the compiler can do better. + */ +#define umul_ppmm(w1, w0, u, v) \ +do { \ + __asm__ ("dmulu %0,%1,%2" \ + : "=d" ((UDItype)(w0)) \ + : "d" ((UDItype)(u)), \ + "d" ((UDItype)(v))); \ + __asm__ ("dmuhu %0,%1,%2" \ + : "=d" ((UDItype)(w1)) \ + : "d" ((UDItype)(u)), \ + "d" ((UDItype)(v))); \ +} while (0) +#elif (__GNUC__ >= 5) || (__GNUC__ >= 4 && __GNUC_MINOR__ >= 4) #define umul_ppmm(w1, w0, u, v) \ do { \ typedef unsigned int __ll_UTItype __attribute__((mode(TI))); \