Possible bitshifting footgun

[lawrence-laz]

Depending on whether a is modified or not (const or var) the bitshift operator produces different results.

test "fails" {
    const a: i16 = -32767;
    try std.testing.expectEqual(-2048, a >> 4); // ● expected -2048, found -2047
}

test "passes" {
    var a: i16 = -32766;
    a -= 1;
    try std.testing.expectEqual(-2048, a >> 4);
}

Not sure if this could be regarded as a bug, but it does feel a bit dangerous

[rohlem]

Not sure if this could be regarded as a bug

Afaik the arithmetic operators are defined in terms of values, regardless where they may come from,
so the same values in the same operation leading to different results is definitely a bug.

[ianprime0509]

Just to confirm, is the result of the shift in the first test -2047 or -2046 for you? When I run the test, I get "expected -2048, found -2046", which aligns with the analysis of the fix in #21320 (off by 2).

Currently, a release compiler reproduces this issue for me, while a debug compiler does not. With #21320 applied, the test also passes under a release compiler. (The difference between debug and release has to do with the undefined here:

zig/src/Value.zig

Line 2779 in 3929cac

.positive = undefined,

)

[weskoerber]

It looks like the behavior here is slightly different between 0.14 and 0.13: 0.14 gives -2046, but 0.13 gives -2047.

❯ zvm use 0.13.0

❯ zig version
0.13.0

❯ zig test shift.zig
expected -2048, found -2047
1/2 shift.test.fails...FAIL (TestExpectedEqual)
/home/wes/.local/share/zvm/0.13.0/lib/std/testing.zig:93:17: 0x103cfe2 in expectEqualInner__anon_1014 (test)
                return error.TestExpectedEqual;
                ^
/home/wes/shift.zig:3:5: 0x103d121 in test.fails (test)
    try std.testing.expectEqual(-2048, a >> 4);
    ^
1 passed; 0 skipped; 1 failed.
error: the following test command failed with exit code 1:
/home/wes/.cache/zig/o/04383ca2a7673a2ec41206dc968a03a1/test

❯ zvm use 0.14.0-dev.1472+3929cac15

❯ zig version
0.14.0-dev.1472+3929cac15

❯ zig test shift.zig
expected -2048, found -2046
1/2 shift.test.fails...FAIL (TestExpectedEqual)
/home/wes/.local/share/zvm/0.14.0-dev.1472+3929cac15/lib/std/testing.zig:97:17: 0x1040772 in expectEqualInner__anon_150 (test)
                return error.TestExpectedEqual;
                ^
/home/wes/shift.zig:3:5: 0x10408b1 in test.fails (test)
    try std.testing.expectEqual(-2048, a >> 4);
    ^
1 passed; 0 skipped; 1 failed.
error: the following test command failed with exit code 1:
/home/wes/.cache/zig/o/cb855978d9e6144d0f1b922565e8e9e9/test --seed=0x49c54965

[ianprime0509]

Ah, that makes sense, I only tested on master. The difference between 0.13 and master is most likely explained by the changes done in #20585 (which fixed an off-by-one error present before that change).