math/big: implement mulAddVWW in riscv64 assembly

This provides an assembly implementation of mulAddVWW for riscv64,
processing up to four words per loop, resulting in a significant
performance gain.

On a StarFive VisionFive 2:

                   │ muladdvww.1  │             muladdvww.2             │
                   │    sec/op    │   sec/op     vs base                │
MulAddVWW/1-4         68.18n ± 0%   65.49n ± 0%   -3.95% (p=0.000 n=10)
MulAddVWW/2-4         82.81n ± 0%   78.85n ± 0%   -4.78% (p=0.000 n=10)
MulAddVWW/3-4         97.49n ± 0%   72.18n ± 0%  -25.96% (p=0.000 n=10)
MulAddVWW/4-4        112.20n ± 0%   85.54n ± 0%  -23.76% (p=0.000 n=10)
MulAddVWW/5-4        126.90n ± 0%   98.90n ± 0%  -22.06% (p=0.000 n=10)
MulAddVWW/10-4        200.3n ± 0%   144.3n ± 0%  -27.96% (p=0.000 n=10)
MulAddVWW/100-4      1532.0n ± 0%   860.0n ± 0%  -43.86% (p=0.000 n=10)
MulAddVWW/1000-4     14.757µ ± 0%   8.076µ ± 0%  -45.27% (p=0.000 n=10)
MulAddVWW/10000-4     204.0µ ± 0%   137.1µ ± 0%  -32.77% (p=0.000 n=10)
MulAddVWW/100000-4    2.066m ± 0%   1.382m ± 0%  -33.12% (p=0.000 n=10)
geomean               1.311µ        950.0n       -27.51%

                   │ muladdvww.1  │             muladdvww.2              │
                   │     B/s      │     B/s       vs base                │
MulAddVWW/1-4        895.1Mi ± 0%   932.0Mi ± 0%   +4.11% (p=0.000 n=10)
MulAddVWW/2-4        1.440Gi ± 0%   1.512Gi ± 0%   +5.02% (p=0.000 n=10)
MulAddVWW/3-4        1.834Gi ± 0%   2.477Gi ± 0%  +35.07% (p=0.000 n=10)
MulAddVWW/4-4        2.125Gi ± 0%   2.787Gi ± 0%  +31.15% (p=0.000 n=10)
MulAddVWW/5-4        2.349Gi ± 0%   3.013Gi ± 0%  +28.28% (p=0.000 n=10)
MulAddVWW/10-4       2.975Gi ± 0%   4.130Gi ± 0%  +38.79% (p=0.000 n=10)
MulAddVWW/100-4      3.891Gi ± 0%   6.930Gi ± 0%  +78.11% (p=0.000 n=10)
MulAddVWW/1000-4     4.039Gi ± 0%   7.380Gi ± 0%  +82.72% (p=0.000 n=10)
MulAddVWW/10000-4    2.922Gi ± 0%   4.346Gi ± 0%  +48.74% (p=0.000 n=10)
MulAddVWW/100000-4   2.884Gi ± 0%   4.313Gi ± 0%  +49.52% (p=0.000 n=10)
geomean              2.321Gi        3.202Gi       +37.95%

Change-Id: If08191607913ce5c7641f34bae8fa5c9dfb44777
Reviewed-on: https://go-review.googlesource.com/c/go/+/595399


Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
Reviewed-by: Mark Ryan <markdryan@rivosinc.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Meng Zhuo <mengzhuo1203@gmail.com>

src/math/big/arith_riscv64.s

@@ -300,7 +300,79 @@ TEXT ·shrVU(SB),NOSPLIT,$0
 	JMP ·shrVU_g(SB)
 
 TEXT ·mulAddVWW(SB),NOSPLIT,$0
-	JMP ·mulAddVWW_g(SB)
+	MOV	x+24(FP), X5
+	MOV	y+48(FP), X6
+	MOV	z+0(FP), X7
+	MOV	z_len+8(FP), X30
+	MOV	r+56(FP), X29
+
+	MOV	$4, X28
+
+	BEQ	ZERO, X30, done
+	BLTU	X30, X28, loop1
+
+loop4:
+	MOV	0(X5), X8	// x[0]
+	MOV	8(X5), X11	// x[1]
+	MOV	16(X5), X14	// x[2]
+	MOV	24(X5), X17	// x[3]
+
+	MULHU	X8, X6, X9	// z_hi[0] = x[0] * y
+	MUL	X8, X6, X8	// z_lo[0] = x[0] * y
+	ADD	X8, X29, X10	// z[0] = z_lo[0] + c
+	SLTU	X8, X10, X23
+	ADD	X23, X9, X29	// next c
+
+	MULHU	X11, X6, X12	// z_hi[1] = x[1] * y
+	MUL	X11, X6, X11	// z_lo[1] = x[1] * y
+	ADD	X11, X29, X13	// z[1] = z_lo[1] + c
+	SLTU	X11, X13, X23
+	ADD	X23, X12, X29	// next c
+
+	MULHU	X14, X6, X15	// z_hi[2] = x[2] * y
+	MUL	X14, X6, X14	// z_lo[2] = x[2] * y
+	ADD	X14, X29, X16	// z[2] = z_lo[2] + c
+	SLTU	X14, X16, X23
+	ADD	X23, X15, X29	// next c
+
+	MULHU	X17, X6, X18	// z_hi[3] = x[3] * y
+	MUL	X17, X6, X17	// z_lo[3] = x[3] * y
+	ADD	X17, X29, X19	// z[3] = z_lo[3] + c
+	SLTU	X17, X19, X23
+	ADD	X23, X18, X29	// next c
+
+	MOV	X10, 0(X7)	// z[0]
+	MOV	X13, 8(X7)	// z[1]
+	MOV	X16, 16(X7)	// z[2]
+	MOV	X19, 24(X7)	// z[3]
+
+	ADD	$32, X5
+	ADD	$32, X7
+	SUB	$4, X30
+
+	BGEU	X30, X28, loop4
+	BEQZ	X30, done
+
+loop1:
+	MOV	0(X5), X10	// x
+
+	MULHU	X10, X6, X12	// z_hi = x * y
+	MUL	X10, X6, X10	// z_lo = x * y
+	ADD	X10, X29, X13	// z_lo + c
+	SLTU	X10, X13, X15
+	ADD	X12, X15, X29	// next c
+
+	MOV	X13, 0(X7)	// z
+
+	ADD	$8, X5
+	ADD	$8, X7
+	SUB	$1, X30
+
+	BNEZ	X30, loop1
+
+done:
+	MOV	X29, c+64(FP)	// return c
+	RET
 
 TEXT ·addMulVVW(SB),NOSPLIT,$0
 	JMP ·addMulVVW_g(SB)