NaN result in a simple test

[monadius]

Test code:

var:  
[7.84, 8] x1;

cost:
sqrt(x1) * 1;

Result:

bin/dop_gelpia benchmarks/tests/sqrt.dop
[[2.82842712474619,NaN], {'x1' : [7.839999999999999, 8],}]

[monadius]

Everything works well when the cost function is sqrt(x1)

[keram88]

This one will take a while to track down. If I print the result of the function call before returning, then everything works correctly. This suggests some undefined behavior.

I'll see if I can reproduce it in c++.

[keram88]

This occurs for all expressions of the form x * sqrt(y), x/sqrt(y), sqrt(y)*x, sqrt(y)/x

[keram88]

The newest version appears to fix this issue. In the past I've seen Rust compiler issues with SIMD values where the result of a function (in the xmm0 register) has to be observed forcefully for the return value to actually be read.

There was an issue from an old Rust compiler where the caller of a function returning a SIMD value would read the result from the eax register instead of xmm0; Gelpia used to have inline assembly code to force the return value to be read. This was fixed in some Rust version.

I'll leave this open since I'd like to do an assembly level analysis (😷) to see what the difference between the existing code and the previous code is. It may be another Rust bug.

[keram88]

The old build.sh passed the --disable-preserve-rounding flag to configure while building GAOL. The new script in https://github.com/soarlab/gelpia/commit/99b9735ae89ffc683206cd622930409bb3fc4311 builds GAOL in the untrusted rounding mode. Historically, gelpia checked that the FP rounding mode never changed. This remained true for a long time, and the checks were removed. Evidently, our new code or modifications to the Rust compiler has introduced rounding mode changes. The erroneous results in issue https://github.com/soarlab/gelpia/issues/17 and this issue are caused by GAOL returning empty intervals which occurs if the upper bound is less than the lower bound.

To take advantage of this, it is necessary to do make clean; make clean-requirements; make requirements; make

We will have to assess the performance impact of this change, but correctness comes first.