Dynamic VFP support and math/test-fpucw

Dynamic VFP support and math/test-fpucw

[Joseph S. Myers]

On ARM systems without VFP hardware, the test math/test-fpucw fails 
because it tries to use _FPU_GETCW, so using unsupported VFP instructions.

Normal user programs using _FPU_GETCW in a soft-float configuration would 
get the fallback soft-float version of that macro and so not have that 
issue.  But the condition for the soft-float version is

#if !defined(_LIBC) && defined(__SOFTFP__)

and _LIBC is defined when testcases are built, not just for building the 
installed libraries.

Really I think most testcases should be built using an environment more 
like that for ordinary user programs, not using internal headers 
(including include/ wrappers) or defining _LIBC unless there is a specific 
need to do so in a particular testcase.  But in the absence of such a 
testsuite overhaul, how might one best stop this testcase failing?  One 
obvious possibility is to use e.g. _LIBC_DYNAMIC_VFP rather than just 
_LIBC in fpu_control.h, and then to define _LIBC_DYNAMIC_VFP in another 
header (not used in the testcase), e.g. arm-features.h.  But that would 
introduce a header ordering dependency (arm-features.h needs to be 
included before fpu_control.h), and while one could make arm-features.h 
use #error if fpu_control.h was included first, this still seems rather 
fragile.

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: Dynamic VFP support and math/test-fpucw

[Roland McGrath]

You could override the test with a sysdeps/arm/test-fpucw.c.
You could override the installed fpu_control.h with a
sysdeps/arm/include/fpu_control.h that tests NOT_IN_libc or something.

Re: Dynamic VFP support and math/test-fpucw

[Joseph S. Myers]

On Mon, 17 Jun 2013, Roland McGrath wrote:

> You could override the test with a sysdeps/arm/test-fpucw.c.
> You could override the installed fpu_control.h with a
> sysdeps/arm/include/fpu_control.h that tests NOT_IN_libc or something.

The trouble is, those options seem rather fragile as well.  Maybe the 
least bad is overriding the test with a file that just defines e.g. 
_LIBC_TEST and includes the main test-fpucw.c (with fpu_control.h then 
checking _LIBC_TEST).  The difficulty with the second is that dynamic VFP 
detection is used in both libc and libm, so NOT_IN_libc isn't sufficient 
on its own.  Really I suppose what's wanted, in the absence of not 
defining _LIBC for tests, would be a macro such as IN_tests that gets 
automatically defined for testcase code; then at least fpu_control.h could 
test that after _LIBC (if it's only tested conditional on _LIBC, namespace 
issues wouldn't matter for IN_tests because no user application should 
ever define _LIBC).

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: Dynamic VFP support and math/test-fpucw

[Roland McGrath]

> The trouble is, those options seem rather fragile as well.

Sure.  I was just offering potential paths given the status quo.  We now
that the eventual solution will be to clean up how we build test programs.
That's not something you can rely on any time soon, but perhaps that being
the plan for the future is sufficient to worry less than you might
otherwise about fragility in a solution that is just a workaround pending
that future.

> Maybe the least bad is overriding the test with a file that just defines
> e.g.  _LIBC_TEST and includes the main test-fpucw.c (with fpu_control.h
> then checking _LIBC_TEST).

That is indeed tolerable.

> The difficulty with the second is that dynamic VFP detection is used in
> both libc and libm, so NOT_IN_libc isn't sufficient on its own.  

If this is saying something more subtle than that you'll want:
	#if !defined NOT_IN_libc && !defined IS_IN_libm
then I am not following.

> Really I suppose what's wanted, in the absence of not defining _LIBC for
> tests, would be a macro such as IN_tests that gets automatically defined
> for testcase code; then at least fpu_control.h could test that after
> _LIBC (if it's only tested conditional on _LIBC, namespace issues
> wouldn't matter for IN_tests because no user application should ever
> define _LIBC).

Defining IN_tests is easy enough.  Having an installed header refer to an
IN_* identifier (even if technically "safe") is beyond the pale, IMHO.


Thanks,
Roland