Segfault adding

[eddelbuettel]

Segfault adding

[mikepb]

Let me try again with gcc instead of clang and update this thread in a few hours.

[eddelbuettel]

There is nothing reproducible in your bug report so my ability to help you from here is limited.

Did you use 0.0.7 from CRAN as your base, or the current master from GitHub (which is marginally ahead and which I called 0.0.7.1) ? FWIW it is worth I can run R CMD check on both just fine (and this build and installs to run tests). But this stuff has been affected by R changes and/or g++ changes.

[erikbern]

i pushed some pretty extensive changes to annoy yesterday but don't think it's a part of rcppannoy yet right? just wanted to double check

[eddelbuettel]

Hi @erikbern and thanks for being so proactive. RcppAnnoy does not automagically copy code from Annoy; it is pretty much a manual copy (along with copious regression checks).

Now, I can't speak for @mikepb, but I dont't think he copied code either. At this point we need more details (as eg provided by R's sessionInfo() along with OS, compiler, ... versions. As I said, "works for me" and also works at CRAN or else they'd come after me..

[mikepb]

Sorry, this was my bad! I had NA for the item value. Here's reproducible code:

sessionInfo()

R version 3.2.3 (2015-12-10)
Platform: x86_64-apple-darwin15.2.0 (64-bit)
Running under: OS X 10.11.3 (El Capitan)

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] RcppAnnoy_0.0.7

loaded via a namespace (and not attached):
[1] Rcpp_0.12.3      codetools_0.2-14

f <- 1049
knn <- new(AnnoyEuclidean, f)
sv <- rnorm(f)
knn$addItem(NA, sv)

 *** caught segfault ***
address 0xfffff7c600000008, cause 'memory not mapped'

Traceback:
 1: .External(list(name = "CppMethod__invoke_void", address = <pointer: 0x7fb5e3431380>,     dll = list(name = "Rcpp", path = "/usr/local/lib/R/3.2/site-library/Rcpp/libs/Rcpp.so",         dynamicLookup = TRUE, handle = <pointer: 0x7fb5e3558f40>,         info = <pointer: 0x10a7af3d0>), numParameters = -1L),     <pointer: 0x7fb5e3544400>, <pointer: 0x7fb5e354d760>, .pointer,     ...)
 2: knn$addItem(NA, sv)

Possible actions:
1: abort (with core dump, if enabled)
2: normal R exit
3: exit R without saving workspace
4: exit R saving workspace

Should the API check for this error condition?

[mikepb]

Looks like the segfault likely comes from annoy passing the NA to realloc, which then crashes the program.

https://github.com/eddelbuettel/rcppannoy/blob/master/inst/include/annoylib.h#L422

[eddelbuettel]

Thanks for looking into this. We can probably catch that.

[eddelbuettel]

Sorry for taking so long. This basically blew up whenever x in knn$addItem(x, vector) was negative, and NA is a special negative value (actually, - INT_MAX). But because we passed from R's SEXP to an int32_t already, I no longer have R's NA tests (easily) so I just test for negative values.

Commit and fix coming up.