math/big: faster assembly kernels for AddVx/SubVx for 386.

(analog to Change-Id: Ia473e9ab9c63a955c252426684176bca566645ae)

Fixes #9243.

benchmark              old ns/op     new ns/op     delta
BenchmarkAddVV_1       5.76          5.60          -2.78%
BenchmarkAddVV_2       7.17          6.98          -2.65%
BenchmarkAddVV_3       8.69          8.57          -1.38%
BenchmarkAddVV_4       10.5          10.5          +0.00%
BenchmarkAddVV_5       13.3          11.6          -12.78%
BenchmarkAddVV_1e1     20.4          19.3          -5.39%
BenchmarkAddVV_1e2     166           140           -15.66%
BenchmarkAddVV_1e3     1588          1278          -19.52%
BenchmarkAddVV_1e4     16138         12657         -21.57%
BenchmarkAddVV_1e5     167608        127836        -23.73%
BenchmarkAddVW_1       4.87          4.76          -2.26%
BenchmarkAddVW_2       6.10          6.07          -0.49%
BenchmarkAddVW_3       7.75          7.65          -1.29%
BenchmarkAddVW_4       9.30          9.39          +0.97%
BenchmarkAddVW_5       10.8          10.9          +0.93%
BenchmarkAddVW_1e1     18.8          18.8          +0.00%
BenchmarkAddVW_1e2     143           134           -6.29%
BenchmarkAddVW_1e3     1390          1266          -8.92%
BenchmarkAddVW_1e4     13877         12545         -9.60%
BenchmarkAddVW_1e5     155330        125432        -19.25%

benchmark              old MB/s     new MB/s     speedup
BenchmarkAddVV_1       5556.09      5715.12      1.03x
BenchmarkAddVV_2       8926.55      9170.64      1.03x
BenchmarkAddVV_3       11042.15     11201.77     1.01x
BenchmarkAddVV_4       12168.21     12245.50     1.01x
BenchmarkAddVV_5       12041.39     13805.73     1.15x
BenchmarkAddVV_1e1     15659.65     16548.18     1.06x
BenchmarkAddVV_1e2     19268.57     22728.64     1.18x
BenchmarkAddVV_1e3     20141.45     25033.36     1.24x
BenchmarkAddVV_1e4     19827.86     25281.92     1.28x
BenchmarkAddVV_1e5     19092.06     25031.92     1.31x
BenchmarkAddVW_1       822.12       840.92       1.02x
BenchmarkAddVW_2       1310.89      1317.89      1.01x
BenchmarkAddVW_3       1549.31      1568.26      1.01x
BenchmarkAddVW_4       1720.45      1703.77      0.99x
BenchmarkAddVW_5       1857.12      1828.66      0.98x
BenchmarkAddVW_1e1     2126.39      2132.38      1.00x
BenchmarkAddVW_1e2     2784.49      2969.21      1.07x
BenchmarkAddVW_1e3     2876.89      3157.35      1.10x
BenchmarkAddVW_1e4     2882.32      3188.51      1.11x
BenchmarkAddVW_1e5     2575.16      3188.96      1.24x

(measured on OS X 10.9.5, 2.3 GHz Intel Core i7, 8GB 1333 MHz DDR3)

Change-Id: I46698729d5e0bc3e277aa0146a9d7a086c0c26f1
Reviewed-on: https://go-review.googlesource.com/2560
Reviewed-by: Keith Randall <khr@golang.org>

diff --git a/src/math/big/arith_386.s b/src/math/big/arith_386.s
index 649bc4dc88f336d150c63dc533af3a0428a6bc93..eb17bc1459eb737ff107b1677e05f470aa39a88e 100644 (file)
--- a/src/math/big/arith_386.s
+++ b/src/math/big/arith_386.s
@@ -7,8 +7,6 @@
 // This file provides fast assembly versions for the elementary
 // arithmetic operations on vectors implemented in arith.go.
 
-// TODO(gri) Replace uses of RCRL/RCLL with ADDL/SBBL respectively.
-
 // func mulWW(x, y Word) (z1, z0 Word)
 TEXT ·mulWW(SB),NOSPLIT,$0
        MOVL x+0(FP), AX
@@ -39,15 +37,16 @@ TEXT ·addVV(SB),NOSPLIT,$0
        JMP E1
 
 L1:    MOVL (SI)(BX*4), AX
-       RCRL $1, DX
+       ADDL DX, DX             // restore CF
        ADCL (CX)(BX*4), AX
-       RCLL $1, DX
+       SBBL DX, DX             // save CF
        MOVL AX, (DI)(BX*4)
        ADDL $1, BX             // i++
 
 E1:    CMPL BX, BP             // i < n
        JL L1
 
+       NEGL DX
        MOVL DX, c+36(FP)
        RET
 
@@ -64,15 +63,16 @@ TEXT ·subVV(SB),NOSPLIT,$0
        JMP E2
 
 L2:    MOVL (SI)(BX*4), AX
-       RCRL $1, DX
+       ADDL DX, DX             // restore CF
        SBBL (CX)(BX*4), AX
-       RCLL $1, DX
+       SBBL DX, DX             // save CF
        MOVL AX, (DI)(BX*4)
        ADDL $1, BX             // i++
 
 E2:    CMPL BX, BP             // i < n
        JL L2
 
+       NEGL DX
        MOVL DX, c+36(FP)
        RET
 
@@ -88,8 +88,8 @@ TEXT ·addVW(SB),NOSPLIT,$0
 
 L3:    ADDL (SI)(BX*4), AX
        MOVL AX, (DI)(BX*4)
-       RCLL $1, AX
-       ANDL $1, AX
+       SBBL AX, AX             // save CF
+       NEGL AX
        ADDL $1, BX             // i++
 
 E3:    CMPL BX, BP             // i < n
@@ -108,11 +108,11 @@ TEXT ·subVW(SB),NOSPLIT,$0
        MOVL $0, BX             // i = 0
        JMP E4
 
-L4:    MOVL (SI)(BX*4), DX     // TODO(gri) is there a reverse SUBL?
+L4:    MOVL (SI)(BX*4), DX
        SUBL AX, DX
        MOVL DX, (DI)(BX*4)
-       RCLL $1, AX
-       ANDL $1, AX
+       SBBL AX, AX             // save CF
+       NEGL AX
        ADDL $1, BX             // i++
 
 E4:    CMPL BX, BP             // i < n