for len(p) >= chunk {
		// Can interlace the computation of w with the
		// rounds below if needed for speed.
		for i := 0; i < 16; i++ {
			j := i * 4
			w[i] = uint32(p[j])<<24 | uint32(p[j+1])<<16 | uint32(p[j+2])<<8 | uint32(p[j+3])
		}

		a, b, c, d, e := h0, h1, h2, h3, h4

		// Each of the four 20-iteration rounds
		// differs only in the computation of f and
		// the choice of K (_K0, _K1, etc).
		i := 0

	MOVD	key+24(FP), KEYP
	MOVD	iv+32(FP), IVP
	MOVD	enc+40(FP), ENC
	MOVD	nr+48(FP), ROUNDS

#ifdef NEEDS_ESPERM
	MOVD	$·rcon(SB), R11
	LVX	(R11), ESPERM   // Permute value for P8_ macros.
#endif

	// Assume len > 0 && len % blockSize == 0.
	CMPW	ENC, $0
	P8_LXVB16X(IVP, R0, IVEC)
	CMPU	ROUNDS, $10, CR1
	CMPU	ROUNDS, $12, CR2 // Only sizes 10/12/14 are supported.

	JNC done

	CRC32B (SI), AX

done:
	MOVL AX, ret+32(FP)
	RET

// castagnoliSSE42Triple updates three (non-inverted) crcs with (24*rounds)
// bytes from each buffer.
//
// func castagnoliSSE42Triple(
//     crc1, crc2, crc3 uint32,
//     a, b, c []byte,
//     rounds uint32,
// ) (retA uint32, retB uint32, retC uint32)
TEXT ·castagnoliSSE42Triple(SB),NOSPLIT,$0
	MOVL crcA+0(FP), AX
	MOVL crcB+4(FP), CX

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

package strconv

// FormatComplex converts the complex number c to a string of the
// form (a+bi) where a and b are the real and imaginary parts,
// formatted according to the format fmt and precision prec.
//
// The format fmt and precision prec have the same meaning as in [FormatFloat].
// It rounds the result assuming that the original was obtained from a complex

#define nop(m,a)		// nop instead of final SHA256MSG1 for first and last few rounds

#define sha256msg1(m,a) \	// final SHA256MSG1 for middle rounds that require it
	SHA256MSG1		m, a

#define vmov(a,b) \		// msg copy for all but rounds 12-15
	VMOVDQA		a, b

#define vmovrev(a,b) \		// reverse copy for rounds 12-15
	VMOVDQA		b, a

// sha rounds 0 to 11
// identical with the exception of the final msg op

		b9 := b[9][i]
		b10 := b[10][i]
		b11 := b[11][i]
		b12 := b[12][i]
		b13 := b[13][i]
		b14 := b[14][i]
		b15 := b[15][i]

		// 4 iterations of eight quarter-rounds each is 8 rounds
		for round := 0; round < 4; round++ {
			b0, b4, b8, b12 = qr(b0, b4, b8, b12)
			b1, b5, b9, b13 = qr(b1, b5, b9, b13)
			b2, b6, b10, b14 = qr(b2, b6, b10, b14)
			b3, b7, b11, b15 = qr(b3, b7, b11, b15)

  queue: TaskQueue,
  message: String,
) {
  fine("${queue.name} ${String.format("%-22s", message)}: ${task.name}")
}

/**
 * Returns a duration in the nearest whole-number units like "999 µs" or "  1 s ". This rounds 0.5
 * units away from 0 and 0.499 towards 0. The smallest unit this returns is "µs"; the largest unit
 * it returns is "s". For values in [-499..499] this returns "  0 µs".
 *

)

func init() {
	intPool.New = func() interface{} {
		return &big.Int{}
	}
}

// scaledValue scales given unscaled value from scale to new Scale and returns
// an int64. It ALWAYS rounds up the result when scale down. The final result might
// overflow.
//
// scale, newScale represents the scale of the unscaled decimal.
// The mathematical value of the decimal is unscaled * 10**(-scale).

// run

// Copyright 2016 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

// This test checks if the compiler's internal constant
// arithmetic correctly rounds up floating-point values
// that become the smallest denormal value.
//
// See also related issue 14553 and test issue14553.go.

package main

import (
	"fmt"
	"math"
)

// implementation outline, the notation now() is used to mean reading
// the virtual clock. In the original paper’s terms, "R(t)" is the
// number of "rounds" that have been completed at real time t ---
// where a round consists of virtually transmitting one bit from every
// non-empty queue in the router (regardless of which queue holds the
// packet that is really being transmitted at the moment); in this