cmd/vet,x/tools/go/analysis/passes/bools: false positive on x != k1 || x != k2 and k1, k2 constants are coincidentally equal

[cznic]

What version of Go are you using (go version)?

go version go1.11.1 linux/amd64

Does this issue reproduce with the latest release?

yes

What operating system and processor architecture are you using (go env)?

GOARCH="amd64"
GOBIN=""
GOCACHE="/home/jnml/.cache/go-build"
GOEXE=""
GOFLAGS=""
GOHOSTARCH="amd64"
GOHOSTOS="linux"
GOOS="linux"
GOPATH="/home/jnml"
GOPROXY=""
GORACE=""
GOROOT="/home/jnml/go"
GOTMPDIR=""
GOTOOLDIR="/home/jnml/go/pkg/tool/linux_amd64"
GCCGO="gccgo"
CC="gcc"
CXX="g++"
CGO_ENABLED="1"
GOMOD=""
CGO_CFLAGS="-g -O2"
CGO_CPPFLAGS=""
CGO_CXXFLAGS="-g -O2"
CGO_FFLAGS="-g -O2"
CGO_LDFLAGS="-g -O2"
PKG_CONFIG="pkg-config"
GOGCCFLAGS="-fPIC -m64 -pthread -fmessage-length=0 -fdebug-prefix-map=/tmp/go-build777349074=/tmp/go-build -gno-record-gcc-switches"

What did you do?

Using github.com/cznic/ql, branch file2, checked out at 3497607bfaba518eab540d317a6c168396b8002f.

==== jnml@4670:~/src/github.com/cznic/ql> go test
# github.com/cznic/ql
./file2.go:37: suspect or: szKey != szVal || szKey != szBuf
FAIL	github.com/cznic/ql [build failed]
==== jnml@4670:~/src/github.com/cznic/ql> 

What did you expect to see?

Build succeeds, tests are executed and failures reported.

What did you see instead?

Build fails.

Additional info

The line go vet does not like is here

func init() {
	if al := cfile.AllocAllign; al != 16 || al <= binary.MaxVarintLen64 {
		panic("internal error")
	}

	if szKey != szVal || szKey != szBuf { // <-- line 37
		panic("internal error")
	}
}

const (
	magic2      = "\x61\xdbql"
	szBuf       = szKey
	szKey       = cfile.AllocAllign // BTree
	szVal       = szKey             // BTree
	wal2PageLog = 12                //TODO tune pagelog
)

Line 37 tests for equality of the three values szKey, szValue, szBuf which other code depends on as it would have to be more complicated for the general case and the equality allow simplifying it. I see nothing wrong about the line, I donť think vet should reject it.

The vet test is located here

// checkSuspect checks for expressions of the form
//   x != c1 || x != c2
//   x == c1 && x == c2
// where c1 and c2 are constant expressions.
// If c1 and c2 are the same then it's redundant;
// if c1 and c2 are different then it's always true or always false.
// Exprs must contain only side effect free expressions.
func (op boolOp) checkSuspect(f *File, exprs []ast.Expr) {
...
}

The comment if c1 and c2 are different then it's always true or always false. is correct, but that does not imply something is wrong.

There's nothing wrong in implementing

a == b == c

in Go as

a == b && a == c

Where the negation, rejected by vet, is

a != b || a != c

Workaround: vet is happy if the expression is rewritten as

szKey != szVal || szVal != szBuf

even though the two versions are logical identities.

[dominikh]

a != b || a != c is always true, a cannot equal both b and c, so one of the two branches has to be true.

The negation you're looking for is !(a == b && a == c) or a != b && a != c

Edit: never mind, tired eyes should not review code.

[mvdan]

I'm not sure I follow this bug report. All three names are constants, and they're equal, so I presume that vet is seeing the condition as a != b || a != b, or even as a != a. That is, the condition will always be false.

It's not that vet doesn't like conditions that are always true or false. I believe this check is to spot human error when writing repetitive boolean expressions.

You're right that a != b || a != c in general is a valid boolean, but I don't think vet would flag that condition if the names involved were variables with unknown values.

[cznic]

All three names are constants, and they're equal, so I presume that vet is seeing the condition as a != b || a != b, or even as a != a. That is, the condition will always be false.

... will be always false iff the three constants are equal. And that's the very point of the test.

If anyone later changes any of the {szKey,szVal,szBuf} constants, each of which tunes an independent thing, in a way such that they become not equal, the condition will become true and the init() will panic as desired because the code simplifications rely on that invariant. So the constants must remain the same or the code must be adjusted to handle them not being equal.

[josharian]

My quick 2c: This is a false positive. But it’s the first one like we have ever had reported (and it has been years), and I don’t see a clean way to fix it, and there’s an easy workaround available. So let’s just take the workaround for now.

[mvdan]

... will be always false iff the three constants are equal. And that's the very point of the test.

Yes - that's why I started the paragraph with "all three names are constants" :)

I don't think I've ever seen an init check to see if constants have been tampered with. Is that something that you expect your users to do? I think this is why this false positive has never been brought up before, like Josh says.

[cznic]

Is that something that you expect your users to do?

I'm the prime suspect any time in the future. Been there, done that ;-)

[mvdan]

My previous comments weren't completely correct. The check documentation reads:

checkSuspect checks for expressions of the form
  x != c1 || x != c2
  x == c1 && x == c2
where c1 and c2 are constant expressions.
If c1 and c2 are the same then it's redundant;
if c1 and c2 are different then it's always true or always false.
Exprs must contain only side effect free expressions.

So this kind of expression is exactly what the vet check is supposed to flag. I mostly agree with Josh, that it should be easy enough for now to avoid this pattern and see if there are any other false positives reported in the future.

I was going to pull in the boolean check author for some second thoughts, but he's ahead of me :)

commit 7da5f193b77edcfcf26ac513864432001693735a
Author: Josh Bleecher Snyder <josharian@gmail.com>
Date:   Wed Jul 2 10:39:57 2014 -0700

    go.tools/cmd/vet: detect stupid boolean conditions

[cznic]

I still don't get where from comes the idea that there's anything wrong with the above quoted expressions.

Please ELI5 why, for example, a test for violating invariant a == b && a == c, written as if a != b || a != c { ... } should be rejected by vet, given a, b and c are constants, considering constants may change during code development/debugging etc.

If some/all of a, b, c would be constant values (literals), ie. not constant names, then the rejection would be obviously correct. (Like in a == 42 && a == 278)

[mvdan]

I think it doesn't distinguish named constants from unnamed ones. Perhaps it should - that would cover this false positive, but it might introduce some false negatives. It looks like none of the current tests use user-defined constants, so that might work. nil should still trigger the check, of course.

@josharian thoughts?

[josharian]

I think it doesn't distinguish named constants from unnamed ones.

That's correct.

I think it'd be useful to gather some data from a corpus about the nature of the constant expressions when this check fires.

[cbeuw]

I think the ultimate solution is to make this a warning. Along with unused import/variables, short edits in code for debugging is painful in Go.

[dfang]

go-vet in vscode complains this as suspect or

if authType != "scan" || authType != "web" {
}

even though add (), still gets

if (authType != "scan") || (authType != "web") {
}

unless rewrite to this, no warnings

if !(authType == "scan" || authType == "web") {
}

but the first one is more plain and straight ....