R devel - Sep 2015 - Optimization bug when byte compiling with gcc 5.2.0 on windows

When building R-devel with gcc 5.2.0 (mingw-w64 v4) on Windows, make
check fails reg-tests-1b.R at the following check:

  x <- c(1:2, NA)
  sx <- sd(x)
  !is.nan(sx)

Here 'sx' should be 'NA' but it is 'NaN'. It turns out
this problem
only appears when the function is byte compiled with optimization
level 3:

  mysd <- function (x, na.rm = FALSE)
    sqrt(var(if (is.vector(x)) x else as.double(x), na.rm = na.rm))

  mysd(x)
  # [1] NA
  compiler::cmpfun(mysd, list(optimize = 1L))(x)
  # [1] NA
  compiler::cmpfun(mysd, list(optimize = 2L))(x)
  # [1] NA
  compiler::cmpfun(mysd, list(optimize = 3L))(x)
  # [1] NaN

The problem does not appear with gcc 5.2.0 on Debian, and also not
with gcc 4.9.3 on Windows. Where do I start debugging this? The
disassembled output from the compiled functions is here:
https://gist.github.com/jeroenooms/3206945a6db6680a9c5c

Jeroen,

the difference is that level 3 is using the internal implementation of sqrt in
the compiler instead of calling the sqrt function. The internal path goes to
R_sqrt which is defined as

# define R_sqrt(x) (ISNAN(x) ? x : sqrt(x))

so you could check if that is where the problem happens. It theory NA should
yield itself, but perhaps the compiler thinks all NANs are equal and will return
a constant one. If that's not it, then something else in the direct
FastMath1(R_sqrt, R_SqrtSym) path...

Cheers,
Simon

> On Sep 14, 2015, at 8:49 AM, Jeroen Ooms <jeroen.ooms at
stat.ucla.edu> wrote:
> 
> When building R-devel with gcc 5.2.0 (mingw-w64 v4) on Windows, make
> check fails reg-tests-1b.R at the following check:
> 
>  x <- c(1:2, NA)
>  sx <- sd(x)
>  !is.nan(sx)
> 
> Here 'sx' should be 'NA' but it is 'NaN'. It turns
out this problem
> only appears when the function is byte compiled with optimization
> level 3:
> 
>  mysd <- function (x, na.rm = FALSE)
>    sqrt(var(if (is.vector(x)) x else as.double(x), na.rm = na.rm))
> 
>  mysd(x)
>  # [1] NA
>  compiler::cmpfun(mysd, list(optimize = 1L))(x)
>  # [1] NA
>  compiler::cmpfun(mysd, list(optimize = 2L))(x)
>  # [1] NA
>  compiler::cmpfun(mysd, list(optimize = 3L))(x)
>  # [1] NaN
> 
> The problem does not appear with gcc 5.2.0 on Debian, and also not
> with gcc 4.9.3 on Windows. Where do I start debugging this? The
> disassembled output from the compiled functions is here:
> https://gist.github.com/jeroenooms/3206945a6db6680a9c5c
> 
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>

I believe the issue is that on Windows the sqrt function when called
with a NaN does not return the same NaN, as it does on other platforms.
We have

#if (defined(_WIN32) || defined(_WIN64)) && defined(__GNUC__) &&
\
     __GNUC__ <= 4
# define R_sqrt(x) (ISNAN(x) ? x : sqrt(x))
#else
# define R_sqrt sqrt
#endif

for implementing the SQRT opcode. I suspect this came from Duncan
Murdoch; I don't know the reason for restricting to __GNUC__ <= 4. I
seem to recall that there are other places in the code where we have
similar workarounds but I can't find them at the moment.

The builtin sqrt always pretests with ISNAN.

Best,

luke



On Mon, 14 Sep 2015, Jeroen Ooms wrote:
> When building R-devel with gcc 5.2.0 (mingw-w64 v4) on Windows, make
> check fails reg-tests-1b.R at the following check:
>
>  x <- c(1:2, NA)
>  sx <- sd(x)
>  !is.nan(sx)
>
> Here 'sx' should be 'NA' but it is 'NaN'. It turns
out this problem
> only appears when the function is byte compiled with optimization
> level 3:
>
>  mysd <- function (x, na.rm = FALSE)
>    sqrt(var(if (is.vector(x)) x else as.double(x), na.rm = na.rm))
>
>  mysd(x)
>  # [1] NA
>  compiler::cmpfun(mysd, list(optimize = 1L))(x)
>  # [1] NA
>  compiler::cmpfun(mysd, list(optimize = 2L))(x)
>  # [1] NA
>  compiler::cmpfun(mysd, list(optimize = 3L))(x)
>  # [1] NaN
>
> The problem does not appear with gcc 5.2.0 on Debian, and also not
> with gcc 4.9.3 on Windows. Where do I start debugging this? The
> disassembled output from the compiled functions is here:
> https://gist.github.com/jeroenooms/3206945a6db6680a9c5c
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>
-- 
Luke Tierney
Ralph E. Wareham Professor of Mathematical Sciences
University of Iowa                  Phone:             319-335-3386
Department of Statistics and        Fax:               319-335-3017
    Actuarial Science
241 Schaeffer Hall                  email:   luke-tierney at uiowa.edu
Iowa City, IA 52242                 WWW:  http://www.stat.uiowa.edu

On 14/09/2015 9:36 AM, luke-tierney at uiowa.edu wrote:
> I believe the issue is that on Windows the sqrt function when called
> with a NaN does not return the same NaN, as it does on other platforms.
> We have
>
> #if (defined(_WIN32) || defined(_WIN64)) && defined(__GNUC__)
&& \
>       __GNUC__ <= 4
> # define R_sqrt(x) (ISNAN(x) ? x : sqrt(x))
> #else
> # define R_sqrt sqrt
> #endif
>
> for implementing the SQRT opcode. I suspect this came from Duncan
> Murdoch; I don't know the reason for restricting to __GNUC__ <= 4.
That was an update to keep the workaround for gcc 4.9.2.  The previous 
code restricted the change to 4.6.x or earlier.

I think the reason for the earlier restriction is the assumption that 
eventually gcc will be fixed and this workaround won't be necessary, but 
apparently 5.2.0 still has the same problem.

Duncan Murdoch
>   I
> seem to recall that there are other places in the code where we have
> similar workarounds but I can't find them at the moment.
>
> The builtin sqrt always pretests with ISNAN.
>
> Best,
>
> luke
>
>
>
> On Mon, 14 Sep 2015, Jeroen Ooms wrote:
>
> > When building R-devel with gcc 5.2.0 (mingw-w64 v4) on Windows, make
> > check fails reg-tests-1b.R at the following check:
> >
> >  x <- c(1:2, NA)
> >  sx <- sd(x)
> >  !is.nan(sx)
> >
> > Here 'sx' should be 'NA' but it is 'NaN'. It
turns out this problem
> > only appears when the function is byte compiled with optimization
> > level 3:
> >
> >  mysd <- function (x, na.rm = FALSE)
> >    sqrt(var(if (is.vector(x)) x else as.double(x), na.rm = na.rm))
> >
> >  mysd(x)
> >  # [1] NA
> >  compiler::cmpfun(mysd, list(optimize = 1L))(x)
> >  # [1] NA
> >  compiler::cmpfun(mysd, list(optimize = 2L))(x)
> >  # [1] NA
> >  compiler::cmpfun(mysd, list(optimize = 3L))(x)
> >  # [1] NaN
> >
> > The problem does not appear with gcc 5.2.0 on Debian, and also not
> > with gcc 4.9.3 on Windows. Where do I start debugging this? The
> > disassembled output from the compiled functions is here:
> > https://gist.github.com/jeroenooms/3206945a6db6680a9c5c
> >
> > ______________________________________________
> > R-devel at r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-devel
> >
>