math/big: Int.Lsh gives wrong results on s390x for n>=128

[mdempsky]

The dev.regabi branch has switched cmd/compile to using go/constant to represent compile-time constants, and it appears to be causing one of package math's unit tests to fail only on s390x: https://build.golang.org/log/0af0cbb9e4d7754d6f16f93a8ea452e265037a40

--- FAIL: TestPow10 (0.00s)
    all_test.go:2840: Pow10(200) = 3.027621884171472e+200, want 1e+200
FAIL
FAIL	math	0.004s

/cc @mundaym @griesemer

[mdempsky]

At least when cross-compiling from amd64 with GOARCH=s390x, math/pow10.go seems to compile identically before and after.

[gopherbot]

Change https://golang.org/cl/273271 mentions this issue: go/constant: add test for #42838

[mdempsky]

It looks like math/big.Rat.SetString is misbehaving on s390x. For SetString("1e192"), it's producing "3027621884171471837698311284263877940814570137795579187554254752777931419931368381990738248285955635616430044818426562647473833981586768834462721163780572037404681705524261520844063170669576192/1".

[mdempsky]

In hex, 1e192 should be (big-endian, 64 bits per line):

381c3de34e49d55a
a16ef894fd1ec505
c3c46289d6388cec
73add001e6a2cf4c
fea73c80f1b8a046
769dbb7e6412e125
a9e17e1fac815d01
0000000000000000
0000000000000000
0000000000000000

But the incorrect value that s390x is producing is:

a9e17e1fac815d01
fea73c80f1b8a046
769dbb7e6412e125
a9e17e1fac815d01
fea73c80f1b8a046
769dbb7e6412e125
a9e17e1fac815d01
0000000000000000
0000000000000000
0000000000000000

Notably, the last 6 words are correct, but then the first four words are words 5--7 repeating in cycle. It reminds me of the kind of data corruption you get when you use memcpy() instead of memmove() for overlapping data.

[mdempsky]

More reduced test case. Fails on s390x, but passes with -tags=math_big_pure_go:

func TestIssue42838_2(t *testing.T) {
	i := new(big.Int)
	i.SetString("159309191113245227702888039776771180559110455519261878607388585338616290151305816094308987472018268594098344692611135542392730712890625", 0)
	i = i.Lsh(i, 192)
	got := i.String()
	want := "1" + strings.Repeat("0", 192)
	if got != want {
		t.Errorf("got %v, want %v", got, want)
	}
}

[mdempsky]

Notably, changing i = i.Lsh(i, 192) to i = i.Lsh(new(big.Int).Set(i), 192) also makes it pass. So s390x's shlVU implementation appears not to handle when the x and z arguments alias.