VCMPEQUW V1, V9, V11        // compare index 2, 6, ... with sep
	VCMPEQUW V1, V10, V12       // compare index 3, 7, ... with sep
	VSEL     V6, V5, V29, V13   // merge index 0, 1, 4, 5, using mask
	VSEL     V12, V11, V30, V14 // merge index 2, 3, 6, 7, using mask
	VSEL     V14, V13, V31, V7  // final merge
	VCLZD    V7, V18            // Find first index for each half
	MFVSRD   V18, R25           // Isolate value

	LXVD2X (P2ptr+R16), X2L
	LXVD2X (P2ptr+R17), Y2H
	LXVD2X (P2ptr+R18), Y2L
	LXVD2X (P2ptr+R19), Z2H
	LXVD2X (P2ptr+R20), Z2L

	VSEL X1H, X2H, SEL, X1H
	VSEL X1L, X2L, SEL, X1L
	VSEL Y1H, Y2H, SEL, Y1H
	VSEL Y1L, Y2L, SEL, Y1L
	VSEL Z1H, Z2H, SEL, Z1H
	VSEL Z1L, Z2L, SEL, Z1L

	STXVD2X X1H, (P3ptr+R0)
	STXVD2X X1L, (P3ptr+R16)
	STXVD2X Y1H, (P3ptr+R17)
	STXVD2X Y1L, (P3ptr+R18)

	VL 0(P2ptr), X2H
	VL 16(P2ptr), X2L
	VL 32(P2ptr), Y2H
	VL 48(P2ptr), Y2L
	VL 64(P2ptr), Z2H
	VL 80(P2ptr), Z2L

	VSEL X2L, X1L, SEL1, X1L
	VSEL X2H, X1H, SEL1, X1H
	VSEL Y2L, Y1L, SEL1, Y1L
	VSEL Y2H, Y1H, SEL1, Y1H
	VSEL Z2L, Z1L, SEL1, Z1L
	VSEL Z2H, Z1H, SEL1, Z1H

	VST X1H, 0(P3ptr)
	VST X1L, 16(P3ptr)
	VST Y1H, 32(P3ptr)
	VST Y1L, 48(P3ptr)
	VST Z1H, 64(P3ptr)

	VCEQF	V1, V4, V6         // compare index 1, 5, ...
	VCEQF	V1, V9, V11        // compare index 2, 6, ...
	VCEQF	V1, V10, V12       // compare index 3, 7, ...
	VSEL	V5, V6, V29, V13   // merge index 0, 1, 4, 5, ...
	VSEL	V11, V12, V30, V14 // merge index 2, 3, 6, 7, ...
	VSEL	V13, V14, V31, V7  // final merge
	VFEEBS	V16, V7, V17       // find leftmost index, set condition to 1 if found
	BLT	foundV17
	MOVD	$16(R7), R7        // R7+=16

GLOBL ·kcon(SB), RODATA, $1312

#define SHA512ROUND0(a, b, c, d, e, f, g, h, xi, idx) \
	VSEL		g, f, e, FUNC; \
	VSHASIGMAD	$15, e, $1, S1; \
	VADDUDM		xi, h, h; \
	VSHASIGMAD	$0, a, $1, S0; \
	VADDUDM		FUNC, h, h; \
	VXOR		b, a, FUNC; \
	VADDUDM		S1, h, h; \
	VSEL		b, c, FUNC, FUNC; \
	VADDUDM		KI, g, g; \
	VADDUDM		h, d, d; \
	VADDUDM		FUNC, S0, S0; \
	LVX		(TBL)(idx), KI; \

#endif

GLOBL ·kcon(SB), RODATA, $1088

#define SHA256ROUND0(a, b, c, d, e, f, g, h, xi, idx) \
	VSEL		g, f, e, FUNC; \
	VSHASIGMAW	$15, e, $1, S1; \
	VADDUWM		xi, h, h; \
	VSHASIGMAW	$0, a, $1, S0; \
	VADDUWM		FUNC, h, h; \
	VXOR		b, a, FUNC; \
	VADDUWM		S1, h, h; \
	VSEL		b, c, FUNC, FUNC; \
	VADDUWM		KI, g, g; \
	VADDUWM		h, d, d; \
	VADDUWM		FUNC, S0, S0; \
	LVX		(TBL)(idx), KI; \

#endif

GLOBL ·kcon(SB), RODATA, $1088

#define SHA256ROUND0(a, b, c, d, e, f, g, h, xi, idx) \
	VSEL		g, f, e, FUNC; \
	VSHASIGMAW	$15, e, $1, S1; \
	VADDUWM		xi, h, h; \
	VSHASIGMAW	$0, a, $1, S0; \
	VADDUWM		FUNC, h, h; \
	VXOR		b, a, FUNC; \
	VADDUWM		S1, h, h; \
	VSEL		b, c, FUNC, FUNC; \
	VADDUWM		KI, g, g; \
	VADDUWM		h, d, d; \
	VADDUWM		FUNC, S0, S0; \
	LVX		(TBL)(idx), KI; \

	"VCMPNEZB",
	"VCMPNEZBCC",
	"VCMPNEB",
	"VCMPNEBCC",
	"VCMPNEH",
	"VCMPNEHCC",
	"VCMPNEW",
	"VCMPNEWCC",
	"VPERM",
	"VPERMXOR",
	"VPERMR",
	"VBPERMQ",
	"VBPERMD",
	"VSEL",
	"VSPLTB",
	"VSPLTH",
	"VSPLTW",
	"VSPLTISB",
	"VSPLTISH",
	"VSPLTISW",
	"VCIPH",
	"VCIPHER",
	"VCIPHERLAST",
	"VNCIPH",
	"VNCIPHER",
	"VNCIPHERLAST",
	"VSBOX",
	"VSHASIGMA",

	"VPKSGS",
	"VPERM",
	"VPDI",
	"VPOPCT",
	"VREP",
	"VREPB",
	"VREPH",
	"VREPF",
	"VREPG",
	"VREPI",
	"VREPIB",
	"VREPIH",
	"VREPIF",
	"VREPIG",
	"VSCEF",
	"VSCEG",
	"VSEL",
	"VSL",
	"VSLB",
	"VSLDB",
	"VSRA",
	"VSRAB",
	"VSRL",
	"VSRLB",
	"VSEG",
	"VSEGB",
	"VSEGH",
	"VSEGF",
	"VST",
	"VSTEH",
	"VSTEF",
	"VSTEG",
	"VSTEB",

	VLEIF	$2, $-43, V16           // e700ffd52843
	VLEIG	$1, $32767, V31         // e7f07fff1842
	VSLDB	$3, V1, V16, V18        // e72100030a77
	VERIMB	$2, V31, V1, V2         // e72f10020472
	VSEL	V1, V2, V3, V4          // e7412000308d
	VGFMAH	V21, V31, V24, V0       // e705f10087bc
	VFMADB	V16, V8, V9, V10        // e7a08300948f
	WFMADB	V17, V18, V19, V20      // e74123083f8f