[forwarded from http://bugs.debian.org/290438] seen with 2.95.4, 3.3.6, 3.4.3. I'm unable to reproduce the ICE mentioned below. The libraries mentioned are from cernlib 2004.11.04. Matthias The bug submitter writes: I have tracked down a bug report I was sent about libmathlib1 to a compiler optimization problem in g77 on i386. Please see the attached file for a test case. (Gunzip and un-tar it, cd into the resulting directory, and run "make" on a system with g77, make, libmathlib1-dev, libpacklib1-dev, and libkernlib1-dev packages installed.) The test case calculates the value of the triple integral of (sin(x)*sin(y)*sin(z))^2 with all three limits going from 1.d-8 to pi, whose value is equal to a tiny amount less than (pi^3)/8. (The setup is done in test.f and the main work in radmul.f, which comes from Cernlib source code.) The program prints out the numerical value of the integral, percent estimated error, number of iterations, and returned error flag. When the file radmul.f is compiled with g77 -O0 -fno-automatic flags, or just with -O1 flag, the following correct output is produced: calling dadmul I= 3.87578459+- 7.11202394E-12% 495 evals error 0 When the flags -fno-automatic -O1 (or higher) are given, the output is quite wrong: calling dadmul I= 42336.3332+- 2.45015127% 999999 evals error 1 If lines 10 and 11 of test.f are changed to set A(1) and A(2) [the lower bounds of integration] to 0.d0 instead of 1.d-8, the output with -fno-automatic -O1 is no longer outrageously large, but it still has a lot of error: calling dadmul I= 3.87588078+- 0.000990792397% 18315 evals error 0 The above behavior is that of g77-3.3 and g77-2.95. g77-3.4 does even worse. It produces the same result as g77-3.3 for the -fno-automatic -O1 case. Furthermore, when g77-3.4 is called with only -O1 (no -fno-automatic flag), the resulting program segfaults. (Run the attached testcase with "make F77=g77-3.4" to see.) For the three arches I've tested, these problems occur only on i386, not on powerpc nor on sparc. You may wish to clone this report to g77-3.4 and g77-2.95. I owe thanks to Eduardo A. Menendez Proupin for finding this bug, tracking down which Cernlib source code file it came from, and composing the test case.
Created attachment 7966 [details] test case
Hmm, this really sounds like PR 323. (I would have them look at that PR and make sure they try out -ffloat-store.
Also since this depends on a library, we really need a self contained example. I still think -ffloat-store will "fix" their problem. If it does then it is a dup of PR 323.
g77 -O1 -ffloat-store work for 3.3.4, but not for 3.4.3.
Hi, I'm the original submitter of the bug to Debian's BTS. On my system, as you predicted, adding -ffloat-store to the options used in compiling radmul.f to object code solves all the problems with g77 3.3 and 3.4. (I haven't re-tested 2.95.) This includes the g77-3.4 segfault. That is, the executable that has radmul.f compiled with "g77-3.4 -c -O1" segfaults, but with "g77-3.4 -c -O1 -ffloat-store" it exhibits the expected behavior. I will try to have a self-contained test case within a week or so (time permitting) so you can check whether this really is the same thing as PR 323.
Since -ffloat-store makes it work, then this is a dup of bug 323. *** This bug has been marked as a duplicate of 323 ***