generator bug - matrix indexing after refactor is inefficient

[mitzimorris]

Summary:

models compiled with Stan 2.19.0 which do a lot of matrix element access are running 10x slower than same model compiled with Stan 2.18.1

see: https://discourse.mc-stan.org/t/cmdstan-slower-than-pystan/8332/11

Description:

The parser/generator code refactor went in with the 2.19 release, which is what introduced this bug. The culprit is in file src/stan/lang/generator/generate_indexed_expr.hpp.

Reproducible Steps:

See model and file referenced in discourse post.
Running compiled model with supplied data from command line using unix time command
under cmdstan 2.18.1 and cmdstan 2.19.0 shows consistent difference in time.
On a MacBook Pro w/ 2 cores, for 2.18.1 1 chain with all defaults runs in 24s, 2.19.0 takes just under 4 minutes - 240 seconds.

Current Output:

Running compiled model with supplied data from command line using unix time command
under cmdstan 2.18.1 and cmdstan 2.19.0 shows consistent difference in time.
On a MacBook Pro w/ 2 cores, for 2.18.1 1 chain with all defaults runs in 24s, 2.19.0 takes just under 4 minutes - 240 seconds.

Expected Output:

2.19.0 should run as fast or faster.

Additional Information:

Current Version:

v2.19.0

[mitzimorris]

multiple indexing semantics described here:
https://mc-stan.org/docs/2_19/reference-manual/language-multi-indexing-section.html

further examples in test model:
https://github.com/stan-dev/stan/blob/develop/src/test/test-models/good/assignments.stan

in particular, note semantics for indexing on arrays of matrices:

matrix[5,6] m2[2,3];
matrix[5,6] m3[2,3,4];
row_vector[6] tz;
tz <- m2[1,1,1];
tz <- m3[1,1,1,1];

[seantalts]

I think I added the model to example-models to a new "regressions" directory and updated the performance-cmdstan-tests to make sure we run it. That job will run here: http://d1m1s1b1.stat.columbia.edu:8080/job/CmdStan%20Performance%20Tests/job/master/590/