This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Testcase for fortran/PR3743 "Reference to intrinsic `ISHFT' invalid".


	I tracked down the underlying cause of the failure, but the cause
is a fundamental design flaw in g77.

	The symptom is intrin.c:ffeintrin_is_intrinsic() returns FALSE for
some Fortran intrinsics.  Intrinsics are checked using bsearch() on the
array of intrinsics in intrin.def.  The names are compared in the case
determined by ffe_case_intrin() which is initialized to 
FFETARGET_defaultCASE_INTRIN in top.c, using the default value
FFE_caseLOWER in target.h.

	bsearch() requires a sorted array.  The failing testcase,
BIT_SIZE, includes an underscore in the name.  Underscore is collated
between uppercase letter and lowercase letters.  intrin.def is sorted for
uppercase names.

DEFNAME ("BITEST",      "bitest",       "BITest",       genNONE, specBITEST)     /* VXT */
DEFNAME ("BIT_SIZE",    "bit_size",     "Bit_Size",     genNONE, specBIT_SIZE)   /* F90 */
DEFNAME ("BJTEST",      "bjtest",       "BJTest",       genNONE, specBJTEST)     /* VXT */

bsearch() looks through the array using a lowercase comparison and expects
to find bit_size collated before bitest, not BIT_SIZE collated after
BITEST.  This problem may be the cause of other errors with respect to
types of intrinsics.

	Because of g77's options for cases, re-sorting the array or
changing the default case would only cover up the problem.  libiberty
provides strcasecmp, so maybe that should be used for the comparison
instead of writing all case variants.

David


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]