BR	notfound

index5plus:
	CMPBGT	R4, $15, index17plus
index2to16:
	CMPBGT	R7, R2, notfound
	MOVD	$1(R7), R8
	CMPBGT	R8, R2, index2to16tail
index2to16loop:
	// unrolled 2x
	VLL	R4, (R7), V1
	VLL	R4, 1(R7), V2
	VCEQGS	V0, V1, V3
	BEQ	found
	MOVD	$1(R7), R7
	VCEQGS	V0, V2, V4
	BEQ	found
	MOVD	$1(R7), R7
	CMPBLT	R7, R2, index2to16loop
	CMPBGT	R7, R2, notfound
index2to16tail:

	BYTE    $0xFF
	BYTE    $0xFF
	SRAD    $32, R1
	WORD    $0xB9170021 //llgtr %r2,%r1
	MOVW    R2, R6
	MOVW    R3, R7
	CMPBGT  R6, R7, L2
	WORD    $0xC0392FFF //iilf  %r3,805306367
	BYTE    $0xFF
	BYTE    $0xFF
	MOVW    R2, R6
	MOVW    R3, R7
	CMPBGT  R6, R7, L9
L3:
	FMOVD   144(R5), F2
	FMADD   F2, F0, F0
L1:
	FMOVD   F0, ret+8(FP)
	RET

L2:

	MOVD	R7, 0(R6)
	MOVD	R8, 8(R6)
	RET

movemt16:
	CMPBGT	R4, R6, forwards
	ADD	R5, R4, R7
	CMPBLE	R7, R6, forwards
	ADD	R5, R6, R8
backwards:
	MOVD	-8(R7), R3
	MOVD	R3, -8(R8)
	MOVD	-16(R7), R3
	MOVD	R3, -16(R8)
	ADD	$-16, R5
	ADD	$-16, R7
	ADD	$-16, R8
	CMP	R5, $16
	BGE	backwards
	BR	start

forwards:
	CMPBGT	R5, $64, forwards_fast
	MOVD	0(R4), R3
	MOVD	R3, 0(R6)
	MOVD	8(R4), R3

	MOVW	R1, R6
	MOVW	R3, R7
	CMPBGT	R6, R7, L2
	MOVH	$0x3FEF, R3
	MOVW	R3, R7
	CMPBGT	R6, R7, L3
	MOVH	$0x2FFF, R2
	MOVW	R2, R7
	CMPBGT	R6, R7, L4
	FMOVD	792(R9), F0
	WFSDB	V2, V0, V2
	FMOVD	F2, ret+8(FP)
	RET

L2:
	LTDBR	F0, F0
	MOVH	$0x0, R4
	BLTU	L3
	FMOVD	F0, F1
L9:
	MOVH	$0x400F, R3
	MOVW	R1, R6
	MOVW	R3, R7
	CMPBGT	R6, R7, L10
	FMOVD	784(R9), F3

	LGDR	F0, R1
	FMOVD	F0, F6
	SRAD	$48, R1
	MOVH	$16383, R3
	RISBGZ	$49, $63, $0, R1, R2
	MOVW	R2, R6
	MOVW	R3, R7
	CMPBGT	R6, R7, L2
	MOVH	$12287, R1
	MOVW	R1, R7
	CMPBLE	R6, R7 ,L12
	MOVH	$16367, R1
	MOVW	R1, R7
	CMPBGT	R6, R7, L5
	FMOVD	448(R5), F4
	FMADD	F0, F0, F4
	FMOVD	440(R5), F3
	WFMDB	V4, V4, V2
	FMOVD	432(R5), F0
	FMOVD	424(R5), F1
	WFMADB	V2, V0, V3, V0

		"BC",
		"BCL",
		"BEQ",
		"BGE",
		"BGT",
		"BL",
		"BLE",
		"BLEU",
		"BLT",
		"BLTU",
		"BNE",
		"BR",
		"BVC",
		"BVS",
		"BRCT",
		"BRCTG",
		"CMPBEQ",
		"CMPBGE",
		"CMPBGT",
		"CMPBLE",
		"CMPBLT",
		"CMPBNE",
		"CMPUBEQ",
		"CMPUBGE",
		"CMPUBGT",
		"CMPUBLE",
		"CMPUBLT",
		"CMPUBNE",
		"CRJ",
		"CGRJ",
		"CLRJ",
		"CLGRJ",
		"CIJ",
		"CGIJ",

	FMOVD	F0, F8
	SRAD	$32, R7
	WORD	$0xC0193FE6 //iilf  %r1,1072079005
	BYTE	$0xA0
	BYTE	$0x9D
	WORD	$0xB91700C7 //llgtr %r12,%r7
	MOVW	R12, R8
	MOVW	R1, R6
	CMPBGT	R8, R6, L2
	WORD	$0xC0193BFF //iilf  %r1,1006632959
	BYTE	$0xFF
	BYTE	$0xFF
	MOVW	R1, R6
	CMPBGT	R8, R6, L13
L3:
	FMOVD	216(R9), F0
	FMADD	F0, F8, F8
L1:
	FMOVD	F8, ret+8(FP)
	RET
L2:
	WORD	$0xC0193FEF	//iilf	%r1,1072693247
	BYTE	$0xFF

	MOVD	R8, 0(R2)
	RET
notfoundr0:
	XOR	R0, R0          // reset R0
	MOVD	$-1, 0(R2)
	RET

vectorimpl:
	//if the address is not 16byte aligned, use loop for the header
	MOVD	R3, R8
	AND	$15, R8
	CMPBGT	R8, $0, notaligned

aligned:
	ADD	R6, R4, R8
	MOVD	R8, R7
	AND	$-16, R7
	// replicate c across V17
	VLVGB	$0, R5, V19
	VREPB	$0, V19, V17

vectorloop:
	CMPBGE	R3, R7, residual

	FMOVD   F6, F0
	WORD    $0xED005010     //adb %f0,.L28-.L18(%r5)
	BYTE    $0x00
	BYTE    $0x1A
	WORD    $0xA7184330     //lhi %r1,17200
	FADD    F2, F0
	MOVW    R2, R10
	MOVW    R1, R11
	CMPBGT  R10, R11, L17
	WORD    $0xED605010     //sdb %f6,.L28-.L18(%r5)
	BYTE    $0x00
	BYTE    $0x1B
	FADD    F6, F2
	WFDDB   V0, V2, V0
	FMOVD   F0, ret+8(FP)
	RET

L9:

	BGT	L10
L3:
	WFCEDBS	V0, V0, V2
	BVS	L1
	FMOVD	112(R9), F0
L1:
	FMOVD	F0, ret+8(FP)
	RET
L2:
	WORD	$0xC0297FEF	//iilf	%r2,2146435071
	BYTE	$0xFF
	BYTE	$0xFF
	MOVW	R1, R6
	MOVW	R2, R7
	CMPBGT	R6, R7, L1
	FMOVD	F0, F8
	FMOVD	$0, F0
	WFADB	V0, V8, V0
	WORD	$0xC0398006	//iilf	%r3,2147909631
	BYTE	$0x7F
	BYTE	$0xFF
	LGDR	F0, R5
	SRAD	$32, R5
	MOVH	$0x0, R1
	SUBW	R5, R3
	FMOVD	$0, F10