Indexing out-of-bounds not always handled properly

[syclik]

$ ./bin/stanc --version
stanc3 9df5448 (Unix)

When indexing a Stan array or vector with an index that's out-of-bounds, the Stan program terminates with different error conditions:

• I would expect the program to catch the error, print a message, and terminate. This happens under some conditions.
• Eigen assertion is triggered.
• segmentation fault
• no error -- looks like it's reading off the end of memory.

Examples

This triggers an Eigen assertion:

parameters {
  real theta;
}
transformed parameters {
  vector[0] x;
}
model {
  theta ~ normal(x[1], 1);
}

This seg faults (may not consistently seg fault):

transformed data {
  vector[0] x[2];
}
parameters {
  real theta;
}
model {
  theta ~ normal(x[0][1], 1);
}

This looks like it's reading off the end of memory:

transformed data {
  vector[0] x[2];
}
parameters {
  real theta;
}
model {
  theta ~ normal(x[3], 1);
}

I believe @bbbales2 has looked at the generated C++ and confirmed the issue.

[syclik]

edit: @SteveBronder has looked at the generated code.

[bbbales2]

I think this is the same as the issue here: #489, that got fixed here: #521, which got reverted here cause it caused a performance regression: #656

Given that it causes segfaults, I kinda like the check back even if it causes noticeable slowdowns this time.

[seantalts]

Ah, interesting! Is it the case that this only happens on index literals (eg y[0], not y[i]) like in the example here and in the first linked issue? Those we could check at compile time with no runtime performance cost. That would be neat.

[rok-cesnovar]

Is it the case that this only happens on index literals (eg y[0], not y[i]) like in the example here and in the first linked issue?

Unfortunately not.

data {
    int N;
}
transformed data {
  vector[N] x[2];
}
parameters {
  real theta;
}
model {
  theta ~ normal(x[N][1], 1);
}

suffers from the same issue.

[seantalts]

Only if N = 0?

[rok-cesnovar]

Indeed.

[bbbales2]

I ran into someone having this problem today. They were getting the error:

Chain 1 Assertion failed: (index >= 0 && index < size()), function operator[], file stan/lib/stan_math/lib/eigen_3.3.7/Eigen/src/Core/DenseCoeffsBase.h, line 408.

I'm into putting @nhuurre 's patch back in and then in the future someone can fix the performance thing with rvalue.

[bbbales2]

@rybern I vaguely remember reading in an issue somewhere about rvalues and when to use them for assignments or not. Are you in r-value land now?

I forgot about this bug, but apparently it was fixed by forcing everything to go through rvalue (#521). We reverted that fix for performance reasons, but never got a fix for this in.

[rybern]

@bbbales2 Not currently in r-value land, I'd have to spend some time catching up on this

[bbbales2]

@rybern ah ah, I thought I remembered something about accessing arrays directly and that reminded me of this.

Anyway I guess it's something different so don't context switch into this just yet. Someone else might have a fix for this laying around. Will check in with @rok-cesnovar on it tomorrow.