all[i] = rand.NormFloat64()
	}

	test(t, "+", add, fop(Fadd64), all)
	test(t, "-", sub, fop(Fsub64), all)
	if GOARCH != "386" { // 386 is not precise!
		test(t, "*", mul, fop(Fmul64), all)
		test(t, "/", div, fop(Fdiv64), all)
	}
}

// 64 -hw-> 32 -hw-> 64
func trunc32(f float64) float64 {
	return float64(float32(f))
}

// 64 -sw->32 -hw-> 64
func to32sw(f float64) float64 {

	return f64to32(fadd64(f32to64(x), f32to64(y)))
}

func fmul32(x, y uint32) uint32 {
	return f64to32(fmul64(f32to64(x), f32to64(y)))
}

func fdiv32(x, y uint32) uint32 {
	// TODO: are there double-rounding problems here? See issue 48807.
	return f64to32(fdiv64(f32to64(x), f32to64(y)))
}

func feq32(x, y uint32) bool {
	cmp, nan := fcmp64(f32to64(x), f32to64(y))
	return cmp == 0 && !nan
}

import (
	"internal/abi"
	"internal/goarch"
	"internal/goos"
	"internal/runtime/atomic"
	"runtime/internal/sys"
	"unsafe"
)

var Fadd64 = fadd64
var Fsub64 = fsub64
var Fmul64 = fmul64
var Fdiv64 = fdiv64
var F64to32 = f64to32
var F32to64 = f32to64
var Fcmp64 = fcmp64
var Fintto64 = fintto64
var F64toint = f64toint

var Entersyscall = entersyscall
var Exitsyscall = exitsyscall

	"RDINSTRET",
	"RDINSTRETH",
	"FRCSR",
	"FSCSR",
	"FRRM",
	"FSRM",
	"FRFLAGS",
	"FSFLAGS",
	"FSRMI",
	"FSFLAGSI",
	"FLW",
	"FSW",
	"FADDS",
	"FSUBS",
	"FMULS",
	"FDIVS",
	"FMINS",
	"FMAXS",
	"FSQRTS",
	"FMADDS",
	"FMSUBS",
	"FNMADDS",
	"FNMSUBS",
	"FCVTWS",
	"FCVTLS",
	"FCVTSW",
	"FCVTSL",
	"FCVTWUS",
	"FCVTLUS",
	"FCVTSWU",
	"FCVTSLU",

	"FCOMDPP",
	"FCOMF",
	"FCOMFP",
	"FCOMI",
	"FCOMIP",
	"FCOML",
	"FCOMLP",
	"FCOMW",
	"FCOMWP",
	"FCOS",
	"FDECSTP",
	"FDIVD",
	"FDIVDP",
	"FDIVF",
	"FDIVL",
	"FDIVRD",
	"FDIVRDP",
	"FDIVRF",
	"FDIVRL",
	"FDIVRW",
	"FDIVW",
	"FFREE",
	"FINCSTP",
	"FINIT",
	"FLD1",
	"FLDCW",
	"FLDENV",
	"FLDL2E",
	"FLDL2T",
	"FLDLG2",
	"FLDLN2",
	"FLDPI",

	"EXTSBCC",
	"EXTSH",
	"EXTSHCC",
	"FABS",
	"FABSCC",
	"FADD",
	"FADDCC",
	"FADDS",
	"FADDSCC",
	"FCMPO",
	"FCMPU",
	"FCTIW",
	"FCTIWCC",
	"FCTIWZ",
	"FCTIWZCC",
	"FDIV",
	"FDIVCC",
	"FDIVS",
	"FDIVSCC",
	"FMADD",
	"FMADDCC",
	"FMADDS",
	"FMADDSCC",
	"FMOVD",
	"FMOVDCC",
	"FMOVDU",
	"FMOVS",
	"FMOVSU",
	"FMOVSX",
	"FMOVSZ",
	"FMSUB",
	"FMSUBCC",

		{_M64, _M64, _M64 - 1, 1, 42},
	} {
		testMul("Mul64", Mul64, a.x, a.y, a.hi, a.lo)
		testMul("Mul64 symmetric", Mul64, a.y, a.x, a.hi, a.lo)
		testDiv("Div64", Div64, a.hi, a.lo+a.r, a.y, a.x, a.r)
		testDiv("Div64 symmetric", Div64, a.hi, a.lo+a.r, a.x, a.y, a.r)
		// The above code can't test intrinsic implementation, because the passed function is not called directly.

	quo, rem = uint32(z/uint64(y)), uint32(z%uint64(y))
	return
}

// Div64 returns the quotient and remainder of (hi, lo) divided by y:
// quo = (hi, lo)/y, rem = (hi, lo)%y with the dividend bits' upper
// half in parameter hi and the lower half in parameter lo.
// Div64 panics for y == 0 (division by zero) or y <= hi (quotient overflow).
func Div64(hi, lo, y uint64) (quo, rem uint64) {
	if y == 0 {
		panic(divideError)
	}

	// amd64:"MULSD",-"DIVSD"
	// arm/7:"MULD",-"DIVD"
	// arm64:"FMULD",-"FDIVD"
	// ppc64x:"FMUL",-"FDIV"
	// riscv64:"FMULD",-"FDIVD"
	x := f1 / 16.0

	// 386/sse2:"MULSD",-"DIVSD"
	// amd64:"MULSD",-"DIVSD"
	// arm/7:"MULD",-"DIVD"
	// arm64:"FMULD",-"FDIVD"
	// ppc64x:"FMUL",-"FDIVD"
	// riscv64:"FMULD",-"FDIVD"
	y := f2 / 0.125

	// 386/sse2:"ADDSD",-"DIVSD",-"MULSD"

	if queryPerformanceFrequency == 0 {
		queryPerformanceFrequency = windows.QueryPerformanceFrequency()
	}
	hi, lo := bits.Mul64(uint64(delta), uint64(time.Second)/uint64(time.Nanosecond))
	quo, _ := bits.Div64(hi, lo, uint64(queryPerformanceFrequency))
	return time.Duration(quo)
}

var queryPerformanceFrequency int64

// highPrecisionTimeSince returns duration since a.