math/big: improve use of addze in mulAddVWW on ppc64x

Improve the use of addze to avoid unnecessary register
moves on ppc64x.

goos: linux
goarch: ppc64le
pkg: math/big
cpu: POWER10
                 │   old.out    │               new.out               │
                 │    sec/op    │    sec/op     vs base               │
MulAddVWW/1         4.524n ± 3%   4.248n ±  0%   -6.10% (p=0.002 n=6)
MulAddVWW/2         5.634n ± 0%   5.283n ±  0%   -6.24% (p=0.002 n=6)
MulAddVWW/3         6.406n ± 0%   5.918n ±  0%   -7.63% (p=0.002 n=6)
MulAddVWW/4         6.484n ± 0%   5.859n ±  0%   -9.64% (p=0.002 n=6)
MulAddVWW/5         7.363n ± 0%   6.766n ±  0%   -8.11% (p=0.002 n=6)
MulAddVWW/10       10.920n ± 0%   9.856n ±  0%   -9.75% (p=0.002 n=6)
MulAddVWW/100       83.46n ± 0%   66.95n ±  0%  -19.78% (p=0.002 n=6)
MulAddVWW/1000      856.0n ± 0%   681.6n ±  0%  -20.38% (p=0.002 n=6)
MulAddVWW/10000     8.589µ ± 1%   6.774µ ±  0%  -21.14% (p=0.002 n=6)
MulAddVWW/100000    86.22µ ± 0%   67.71µ ± 43%  -21.48% (p=0.065 n=6)
geomean             73.34n        63.62n        -13.26%

Change-Id: I95d6ac49ff6b64aa678e6896f57af9d85c923aad
Reviewed-on: https://go-review.googlesource.com/c/go/+/579235
Reviewed-by: Carlos Amedee <carlos@golang.org>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Paul Murphy <murp@ibm.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

diff --git a/src/math/big/arith_ppc64x.s b/src/math/big/arith_ppc64x.s
index c483e252ab5ceb87d3475bb932ef50f6ac1f1e12..330bc7e46cbe177e40e7b378935de51d1b128607 100644 (file)
--- a/src/math/big/arith_ppc64x.s
+++ b/src/math/big/arith_ppc64x.s
@@ -514,9 +514,8 @@ TEXT ·mulAddVWW(SB), NOSPLIT, $0
        MULLD   R9, R20, R6       // R6 = z0 = Low-order(x[i]*y)
        MULHDU  R9, R20, R7       // R7 = z1 = High-order(x[i]*y)
        ADDC    R4, R6            // R6 = z0 + r
-       ADDZE   R7                // R7 = z1 + CA
+       ADDZE   R7, R4            // R4 = z1 + CA
        CMP     R0, R11
-       MOVD    R7, R4            // R4 = c
        MOVD    R6, 0(R10)        // z[i]
        BEQ     done
 
@@ -536,20 +535,17 @@ loop:
        MULLD   R9, R20, R24      // R24 = z0[i]
        MULHDU  R9, R20, R20      // R20 = z1[i]
        ADDC    R4, R24           // R24 = z0[i] + c
-       ADDZE   R20               // R7 = z1[i] + CA
        MULLD   R9, R21, R25
        MULHDU  R9, R21, R21
-       ADDC    R20, R25
-       ADDZE   R21
+       ADDE    R20, R25
        MULLD   R9, R22, R26
        MULHDU  R9, R22, R22
        MULLD   R9, R23, R27
        MULHDU  R9, R23, R23
-       ADDC    R21, R26
-       ADDZE   R22
+       ADDE    R21, R26
        MOVD    R24, 8(R10)       // z[i]
        MOVD    R25, 16(R10)      // z[i+1]
-       ADDC    R22, R27
+       ADDE    R22, R27
        ADDZE   R23,R4            // update carry
        MOVD    R26, 24(R10)      // z[i+2]
        MOVDU   R27, 32(R10)      // z[i+3]
@@ -567,10 +563,9 @@ tail:
        MULHDU  R9, R20, R25      // R25 = z1[i]
        ADD     $-1, R11          // R11 = z_len - 1
        ADDC    R4, R24
-       ADDZE   R25
+       ADDZE   R25, R4
        MOVDU   R24, 8(R10)       // z[i]
        CMP     R0, R11
-       MOVD    R25, R4           // R4 = c
        BEQ     done              // If R11 = 0, we are done
 
        MOVDU   8(R8), R20
@@ -578,10 +573,9 @@ tail:
        MULHDU  R9, R20, R25
        ADD     $-1, R11
        ADDC    R4, R24
-       ADDZE   R25
+       ADDZE   R25, R4
        MOVDU   R24, 8(R10)
        CMP     R0, R11
-       MOVD    R25, R4
        BEQ     done
 
        MOVD    8(R8), R20
@@ -589,9 +583,8 @@ tail:
        MULHDU  R9, R20, R25
        ADD     $-1, R11
        ADDC    R4, R24
-       ADDZE   R25
+       ADDZE   R25,R4
        MOVD    R24, 8(R10)
-       MOVD    R25, R4
 
 done:
        MOVD    R4, c+64(FP)