PPPIOCUNBRIDGECHAN               = 0x7434
	PPPIOCXFERUNIT                   = 0x744e
	PROT_BTI                         = 0x10
	PROT_MTE                         = 0x20
	PR_SET_PTRACER_ANY               = 0xffffffffffffffff
	PTRACE_PEEKMTETAGS               = 0x21
	PTRACE_POKEMTETAGS               = 0x22
	PTRACE_SYSEMU                    = 0x1f
	PTRACE_SYSEMU_SINGLESTEP         = 0x20
	RLIMIT_AS                        = 0x9

GLOBL PSHUFFLE_BYTE_FLIP_MASK<>(SB), (NOPTR+RODATA), $32

DATA MASK_YMM_LO<>+0x00(SB)/8, $0x0000000000000000
DATA MASK_YMM_LO<>+0x08(SB)/8, $0x0000000000000000
DATA MASK_YMM_LO<>+0x10(SB)/8, $0xFFFFFFFFFFFFFFFF
DATA MASK_YMM_LO<>+0x18(SB)/8, $0xFFFFFFFFFFFFFFFF

GLOBL MASK_YMM_LO<>(SB), (NOPTR+RODATA), $32

TEXT ·blockAVX2(SB), NOSPLIT, $56-32
	MOVQ dig+0(FP), SI
	MOVQ p_base+8(FP), DI
	MOVQ p_len+16(FP), DX

	SHRQ $7, DX

	PPPIOCSRASYNCMAP                 = 0x40047454
	PPPIOCSXASYNCMAP                 = 0x4020744f
	PPPIOCUNBRIDGECHAN               = 0x7434
	PPPIOCXFERUNIT                   = 0x744e
	PR_SET_PTRACER_ANY               = 0xffffffffffffffff
	PTRACE_ARCH_PRCTL                = 0x1e
	PTRACE_GETFPREGS                 = 0xe
	PTRACE_GETFPXREGS                = 0x12
	PTRACE_GET_THREAD_AREA           = 0x19
	PTRACE_OLDSETOPTIONS             = 0x15

	PPPIOCSRASYNCMAP                 = 0x40047454
	PPPIOCSXASYNCMAP                 = 0x4020744f
	PPPIOCUNBRIDGECHAN               = 0x7434
	PPPIOCXFERUNIT                   = 0x744e
	PR_SET_PTRACER_ANY               = 0xffffffffffffffff
	PTRACE_SYSEMU                    = 0x1f
	PTRACE_SYSEMU_SINGLESTEP         = 0x20
	RLIMIT_AS                        = 0x9
	RLIMIT_MEMLOCK                   = 0x8
	RLIMIT_NOFILE                    = 0x7

	PPPIOCSXASYNCMAP                 = 0x8020744f
	PPPIOCUNBRIDGECHAN               = 0x20007434
	PPPIOCXFERUNIT                   = 0x2000744e
	PR_SET_PTRACER_ANY               = 0xffffffffffffffff
	PTRACE_GETFPREGS                 = 0xe
	PTRACE_GET_THREAD_AREA           = 0x19
	PTRACE_GET_THREAD_AREA_3264      = 0xc4
	PTRACE_GET_WATCH_REGS            = 0xd0
	PTRACE_OLDSETOPTIONS             = 0x15

// on little endian.
DATA p256<>+0x00(SB)/8, $0xffffffff00000001 // P256
DATA p256<>+0x08(SB)/8, $0x0000000000000000 // P256
DATA p256<>+0x10(SB)/8, $0x00000000ffffffff // P256
DATA p256<>+0x18(SB)/8, $0xffffffffffffffff // P256
DATA p256<>+0x20(SB)/8, $0x0c0d0e0f1c1d1e1f // SEL d1 d0 d1 d0
DATA p256<>+0x28(SB)/8, $0x0c0d0e0f1c1d1e1f // SEL d1 d0 d1 d0
DATA p256<>+0x30(SB)/8, $0x0000000010111213 // SEL 0  d1 d0  0

	PPPIOCSXASYNCMAP                 = 0x8020744f
	PPPIOCUNBRIDGECHAN               = 0x20007434
	PPPIOCXFERUNIT                   = 0x2000744e
	PR_SET_PTRACER_ANY               = 0xffffffffffffffff
	PTRACE_GETFPAREGS                = 0x14
	PTRACE_GETFPREGS                 = 0xe
	PTRACE_GETFPREGS64               = 0x19
	PTRACE_GETREGS64                 = 0x16
	PTRACE_READDATA                  = 0x10

	PPPIOCSXASYNCMAP                 = 0x8020744f
	PPPIOCUNBRIDGECHAN               = 0x20007434
	PPPIOCXFERUNIT                   = 0x2000744e
	PROT_SAO                         = 0x10
	PR_SET_PTRACER_ANY               = 0xffffffffffffffff
	PTRACE_GETEVRREGS                = 0x14
	PTRACE_GETFPREGS                 = 0xe
	PTRACE_GETREGS64                 = 0x16
	PTRACE_GETVRREGS                 = 0x12
	PTRACE_GETVSRREGS                = 0x1b

	}
	for m := uint(1); m <= 64; m *= 2 {
		tests := []uint64{
			0x0000000000000000,
			0x00000000ffffffff,
			0xffffffff00000000,
			0x8000000000000001,
			0xf00000000000000f,
			0xf00000010050000f,
			0xffffffffffffffff,
			0x0000000000000001,
			0x0000000000000002,
			0x0000000000000008,
			uint64(1) << (m - 1),
			uint64(1) << m,
			// Try a few fixed arbitrary examples.
			0xb02b9effcf137016,
			0x3975a076a9fbff18,

(SGTUconst [c] (ANDconst [m] _)) && uint64(m) < uint64(c) => (MOVVconst [1])
(SGTconst [c] (SRLVconst _ [d])) && 0 <= c && 0 < d && d <= 63 && 0xffffffffffffffff>>uint64(d) < uint64(c) => (MOVVconst [1])
(SGTUconst [c] (SRLVconst _ [d])) && 0 < d && d <= 63 && 0xffffffffffffffff>>uint64(d) < uint64(c) => (MOVVconst [1])

// absorb constants into branches
(EQ  (MOVVconst [0]) yes no) => (First yes no)