ROUND4(X0, X5, X15, X10, X1, X6, X12, X11, X2, X7, X13, X8,  X3, X4, X14, X9)

	ADD $-1, R1
	BNE loop

	// decrement length
	ADD $-256, R4

	// rearrange vectors
	SHUFFLE(X0, X1, X2, X3, M0, M1, M2, M3)
	ADDV(J0, X0, X1, X2, X3)
	SHUFFLE(X4, X5, X6, X7, M0, M1, M2, M3)
	ADDV(KEY0, X4, X5, X6, X7)
	SHUFFLE(X8, X9, X10, X11, M0, M1, M2, M3)
	ADDV(KEY1, X8, X9, X10, X11)

// license that can be found in the LICENSE file.

//go:build !math_big_pure_go && (mips || mipsle)

#include "textflag.h"

// This file provides fast assembly versions for the elementary
// arithmetic operations on vectors implemented in arith.go.

TEXT ·addVV(SB),NOSPLIT,$0
	JMP	·addVV_g(SB)

TEXT ·subVV(SB),NOSPLIT,$0
	JMP	·subVV_g(SB)

TEXT ·addVW(SB),NOSPLIT,$0
	JMP	·addVW_g(SB)

TEXT ·subVW(SB),NOSPLIT,$0

	BLT	tail		// Jump to the small string case
	SRD	$5, R4, R20
	MOVD	R20, CTR
	MOVD	$16, R21
	XXLXOR	V4, V4, V4
	XXLXOR	V5, V5, V5

	PCALIGN	$16
cmploop:
	LXVD2X	(R0)(R3), V0	// Count 32B per loop with two vector accumulators.
	LXVD2X	(R21)(R3), V2
	VCMPEQUB V2, V1, V2
	VCMPEQUB V0, V1, V0
	VPOPCNTD V2, V2		// A match is 0xFF or 0. Count the bits into doubleword buckets.
	VPOPCNTD V0, V0

	MOVL (AX)(X0*1), AX             // ERROR "invalid instruction"
	MOVL (AX)(Y0*1), AX             // ERROR "invalid instruction"
	// VSIB/VM is invalid without vector index.
	// TODO(quasilyte): improve error message (#21860).
	// "invalid VSIB address (missing vector index)"
	VPGATHERQQ Y2, (BP), Y1         // ERROR "invalid instruction"
	// AVX2GATHER mask/index/dest #UD cases.

// license that can be found in the LICENSE file.

//go:build !math_big_pure_go

#include "textflag.h"

// This file provides fast assembly versions for the elementary
// arithmetic operations on vectors implemented in arith.go.

// DI = R3, CX = R4, SI = r10, r8 = r8, r9=r9, r10 = r2, r11 = r5, r12 = r6, r13 = r7, r14 = r1 (R0 set to 0) + use R11
// func addVV(z, x, y []Word) (c Word)

TEXT ·addVV(SB), NOSPLIT, $0

	// Create a struct callbackArgs on our stack.
	SUBL	$(12+callbackArgs__size), SP
	MOVL	AX, (12+callbackArgs_index)(SP)		// callback index
	MOVL	CX, (12+callbackArgs_args)(SP)		// address of args vector
	MOVL	$0, (12+callbackArgs_result)(SP)	// result
	LEAL	12(SP), AX	// AX = &callbackArgs{...}

	// Call cgocallback, which will call callbackWrap(frame).
	MOVL	$0, 8(SP)	// context

#ifdef GOARCH_ppc64le
#define VPERMLE(va,vb,vc,vt) VPERM va, vb, vc, vt
#else
#define VPERMLE(va,vb,vc,vt)
#endif

// 2 copies of each Kt, to fill both doublewords of a vector register
DATA  ·kcon+0x000(SB)/8, $0x428a2f98d728ae22
DATA  ·kcon+0x008(SB)/8, $0x428a2f98d728ae22
DATA  ·kcon+0x010(SB)/8, $0x7137449123ef65cd
DATA  ·kcon+0x018(SB)/8, $0x7137449123ef65cd

	MOVQ	R9, 24(SP)
	CALL	runtime·sehhandler(SB)
	MOVL	32(SP), AX

	ADJSP	$-40

	POP_REGS_HOST_TO_ABI0()
	RET

TEXT runtime·callbackasm1(SB),NOSPLIT|NOFRAME,$0
	// Construct args vector for cgocallback().
	// By windows/amd64 calling convention first 4 args are in CX, DX, R8, R9
	// args from the 5th on are on the stack.
	// In any case, even if function has 0,1,2,3,4 args, there is reserved

DATA  ·kcon+0x418(SB)/8, $0x1011121300010203
DATA  ·kcon+0x420(SB)/8, $0x1011121310111213
DATA  ·kcon+0x428(SB)/8, $0x0405060700010203
DATA  ·kcon+0x430(SB)/8, $0x1011121308090a0b
DATA  ·kcon+0x438(SB)/8, $0x0405060700010203
#else
DATA  ·kcon+0x410(SB)/8, $0x1011121300010203
DATA  ·kcon+0x418(SB)/8, $0x1011121310111213 // permutation control vectors
DATA  ·kcon+0x420(SB)/8, $0x0405060700010203

// license that can be found in the LICENSE file.

//go:build !math_big_pure_go

#include "textflag.h"

// This file provides fast assembly versions for the elementary
// arithmetic operations on vectors implemented in arith.go.

// func addVV(z, x, y []Word) (c Word)
TEXT ·addVV(SB),NOSPLIT,$0
	MOVL z+0(FP), DI
	MOVL x+12(FP), SI
	MOVL y+24(FP), CX
	MOVL z_len+4(FP), BP
	MOVL $0, BX		// i = 0