math/big: improve performance of mulAddVWW on ppc64x
This changes the assembly implementation on ppc64x
to improve performance by reordering some instructions.
It also eliminates an unnecessary move by changing an
ADDZE to use the correct target register.
Improvement on power9:
MulAddVWW/1 6.89ns ± 0% 7.30ns ± 0% +5.95% (p=1.000 n=1+1)
MulAddVWW/2 8.04ns ± 0% 8.06ns ± 0% +0.25% (p=1.000 n=1+1)
MulAddVWW/3 9.39ns ± 0% 9.39ns ± 0% ~ (all equal)
MulAddVWW/4 9.76ns ± 0% 9.48ns ± 0% -2.87% (p=1.000 n=1+1)
MulAddVWW/5 10.5ns ± 0% 10.3ns ± 0% -1.90% (p=1.000 n=1+1)
MulAddVWW/10 15.4ns ± 0% 14.9ns ± 0% -3.25% (p=1.000 n=1+1)
MulAddVWW/100 149ns ± 0% 125ns ± 0% -16.11% (p=1.000 n=1+1)
MulAddVWW/1000 1.42µs ± 0% 1.28µs ± 0% -9.74% (p=1.000 n=1+1)
MulAddVWW/10000 14.2µs ± 0% 12.8µs ± 0% -9.73% (p=1.000 n=1+1)
MulAddVWW/100000 144µs ± 0% 129µs ± 0% -10.10% (p=1.000 n=1+1)
Change-Id: I0ae7002a69783ca19d7a4e3e42042ae75dc60069
Reviewed-on: https://go-review.googlesource.com/c/go/+/248721
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Carlos Eduardo Seo <cseo@linux.vnet.ibm.com>
Reviewed-by: Paul Murphy <murp@ibm.com>
diff --git a/src/math/big/arith_ppc64x.s b/src/math/big/arith_ppc64x.s
index dbb168a..409e10a 100644
--- a/src/math/big/arith_ppc64x.s
+++ b/src/math/big/arith_ppc64x.s
@@ -394,17 +394,16 @@
ADDZE R21
MULLD R9, R22, R26
MULHDU R9, R22, R22
- ADDC R21, R26
- ADDZE R22
MULLD R9, R23, R27
MULHDU R9, R23, R23
- ADDC R22, R27
- ADDZE R23
+ ADDC R21, R26
+ ADDZE R22
MOVD R24, 8(R10) // z[i]
MOVD R25, 16(R10) // z[i+1]
+ ADDC R22, R27
+ ADDZE R23,R4 // update carry
MOVD R26, 24(R10) // z[i+2]
MOVDU R27, 32(R10) // z[i+3]
- MOVD R23, R4 // R4 = c
ADD $-4, R11 // R11 = z_len - 4
BC 16, 0, loop // bdnz
