Error trap for reduction operations on vector arguments

[rperry-usgs]

Hi there! I ran into this problem, fixed it, and subsequently read that nimble functions don't handle negative indices, as in R. Indeed, in a function more complex than the reproduceable example below (a custom sampler), I received an error when indexing by x[-d] where d = length(x). Here's the specific error returned on compilation:

Error:
Error: NIMBLE compiler does not support reduction operations on scalar arguments.
This occurred for: sum(x_next[-(d)])
This was part of the call: totTime - sum(x_next[-(d)])

However, the function compiled successfully and happily returned a wrong result when using x[-c(i,d)] where i is a for loop index. I only caught this error because results did not match my sampler implemented in R, but I could see how it might go unnoticed or be difficult to debug in some cases. So I'm flagging this as a potential candidate for a stop-execution error trap that doesn't return a result, similar to the error message above.

Below is a simple case reproducing the error. I'm using nimble v1.3.0.

testNegIdx <- nimbleFunction(
run = function(x = double(1)) {
ans <- x[-c(1,4)]
return(ans)
returnType(double(1))
} )

Should return c(2, 3), and does.
x <- 1:4

testNegIdx(x)
[1] 2 3

Test compiled version
`CtestNegIdx <- compileNimble(testNegIdx)

CtestNegIdx(x)
[1] 7.410985e-323 8.332910e-312
CtestNegIdx(x)
[1] 0.000000e+00 8.332191e-312
CtestNegIdx(x)
[1] 1.022650e-312 1.447621e+166

Not c(2,3) and returns different values each time it's called.

[paciorek]

@rperry-usgs thanks for this report.

Indeed in both the x[-2] and x[-c(2,4)] cases we should error trap. Furthermore, it doesn't look to me that we mention this lack of functionality in the manual.

Let's consider putting a check for - in sizeIndexingBracket. It looks like we could simply check for - in code$args, e.g,

sapply(code$args, \(x) exists('name', x) && x$name =="-")