raspberriesareyummy comments on RTFM

[–] raspberriesareyummy@lemmy.world 5 points 1 year ago (2 children)

Been there, found undefined behavior where there should not be any. Imagine a function that takes a bool param with the following code, but neither branch gets executed:

if (b)
   doStuffForTrue();
if (!b)
   doStuffForFalse();

In a function that is passed an uninitialized bool parameter, in gcc compiler, both branches can get executed even when b is const. Reason: uninitialized bool in gcc can have values of a random integer, and while if(b) {} else ({} is guaranteed to execute only one branch, bool evaluations of a bool value take a "shortcut" that only has defined behavior with an initialized bool.

Same code with an uninitialized integer works as expected, btw.

permalink

fedilink

source

parent

hideshow 2 child comments

[–] zerofk@lemmy.zip 2 points 1 year ago (1 child)

Don’t blame this on gcc or the library/function author - it is 100% user (i.e. programmer) error. Uninitialised memory of any type is undefined behaviour in the C and C++ abstract machine. That means optimising compilers can assume it does not exist.

For example, the compiler could see that your ‘b’ is never initialised. Therefore, using it would be undefined behaviour. So, the optimiser can assume it is never used, and it is as if that code simply does not exist: the behaviour you saw.

I’m not saying that is what happened, nor that it will always happen, but it is a possibility.

permalink

fedilink

source

parent

hideshow 1 child comment

[–] raspberriesareyummy@lemmy.world 1 point 1 year ago*

Don’t blame this on gcc or the library/function author - it is 100% user (i.e. programmer) error. Uninitialised memory of any type is undefined behaviour in the C and C++ abstract machine. That means optimising compilers can assume it does not exist.

I absolutely do blame this on the C++ standard being not specific enough, specifically for the way in how I learned about this: When writing a trivial function, you would never expect that - for a bool parameter - an "if (b)" branch can be executed as well as an "if (!b)" branch.

So basically, this mechanic sabotages input data validation in functions that test whether plausible parameters were provided. The problem is that a function you write that is bug-free and "perfect code" - despite input data validation - can exhibit undefined behavior due to an uninitialized bool type parameter. Something that can not happen with other uninitialized trivial (numeric) data types (int, float). Simply due to the way boolean checks are translated to x86 assembly:

Here's an example: https://godbolt.org/z/T3f9csohd

Note the assembly lines 176-182: The only difference for the "if (!b)" check is that the lowest bit of the boolean is flipped with an xor - which assumes about the implementation that a boolean can never hold values other than 0 or 1. Which I - as a naive user - also assumed until this happened. Correction: I assumed that negating a bool would result in the inverse boolean value.

So the problem boils down to: The value range of any given (built-in) numerical data type fully encloses the value range that an uninitialized variable of that type can have. This is not necessarily true for boolean: In g++, the value range is [0;1] and the range of an uninitialized bool is [0;255].

Accordingly, I would expect the C++ standard to fix this by stating that an uninitialized bool must have a value for which only one of two conditions evluates to true: b or !b, but not both.

permalink

fedilink

source

parent