GCC / gfortran 4.6.0 status

[Tobias Burnus <burnus at net-b dot de>]

GCC status: 156 serious regressions of which 17 are P1, 102 are P2 and 
37 are P3. (Last GCC status update 3 Nov [1].)


gfortran status: 11 regressions of which is 1 P1 and only one also 
affects older versions (4.5).

(1) 46842  P1, wrong results for MATHMUL(...,TRANSPOSE(function(...))

while taking the transpose of a function, this kind of wrong-code is 
pretty bad. Additionally, it affects the beloved Tonto chemistry program 
(which is also part of SPEC CPU 2006).

Decl (-fwhole-file) related:
(2) 44945 wrong-code with MODULE variables
(3) 46818  ICE on pointer assignment

OOP related (regressions related to the deep-copy patch)
(4) 46448 rejects valid (due to doubly defined symbol) - possibly a decl 
problem
(5) 46262 ICE on valid:

(6) 45743 That's a multi bug - including multi-decl issue for external 
functions, ICE (via tree checking), error vs. warning (incl. test suite) 
and maybe more ...

(7) 46607  Relocated install of libgfortran fails now (as libjava always 
did) [Might be somehow libquadmath related]

Quadmath-related PRs:
(8) 46520  Build problem on bare irons (if it still exists). If it does: 
symbol checks require link test. The same is true for libgfortran - but 
seemingly no one enables Fortran on those systems...
(9) 46520 Spurious build failure on Thomas' computer (not reproducible; 
status WAITING)

(10) ICE non-trivial conversion at assignment -- PR assigned to Richard. 
Seemingly there is a middle end issue; whether there is a real or latent 
front-end issue is a bit unclear.

(11) [4.5/4.6] CPP issue: missing target declarations since libcpp is 
directly called. Requires the modification of several targets and is 
thus probably a 4.7 issue. (Bit rotted patch of FX exists.)


I think we should try to start working on some of the regressions rather 
sooner than later as I can imagine that, e.g., the decl issues might 
requires some nontrivial changes. Additionally, there are also some 
other bugs which should be fixed (and haven't because of adding nifty 
features). I assume that in about 3 months there will be a release - 
given that certain changes should not be done during Stage 4: That's not 
that much time.

Otherwise, I think gfortran seems to be working quite nice. Also most 
integration issues of libquadmath are now solved.

Tobias

PS: There were already quite a few Fortran P1 regressions in the past; 
in 2003 (1 PR), 2004 (4), 2005 (11), 2006 (2), 2009 (1) and 2010 (3).

PPS: In a small tool program I wrote yesterday, gfortran was 23% faster 
than ifort. However, the formatted I/O took only half of the time with 
Intel (2.4 vs. 1.2 seconds for reading two 460,000 lines files).

[1] http://gcc.gnu.org/ml/gcc/2010-11/msg00087.html and 
http://gcc.gnu.org/ml/gcc/2010-10/msg00417.html