http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53265
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Attachment #29650|0 |1
is obsolete| |
--- Comment #21 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-03-12
20:09:58 UTC ---
Created attachment 29657
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29657
gcc48-pr53265.patch
Updated patch that actually passed bootstrap/regtest. Unfortunately, for fn4
and fn7 in the testcase to pass, I had to give up with only warning in
PROP_loops state, because the cunrolli pass optimizes those loops.
While it bootstrapped fine (with unwind-dw2.c fix I'll attach too), there were
some regressions, some -Waggressive-loop-optimizations warnings during
bootstrap/regtest and also with the additional debugging hack in various places
the number of iterations upper bound increased.
The regressions are:
../../gcc/ada/exp_ch7.adb:3540:11: warning: iteration 2147483646 invokes
undefined behavior [-Waggressive-loop-optimizations]
warning during x86_64 Ada build and:
+FAIL: gcc.dg/torture/pr49518.c -O3 -fomit-frame-pointer (test for excess
errors)
+FAIL: gcc.dg/torture/pr49518.c -O3 -fomit-frame-pointer -funroll-loops (test
for excess errors)
+FAIL: gcc.dg/torture/pr49518.c -O3 -fomit-frame-pointer -funroll-all-loops
-finline-functions (test for excess errors)
+FAIL: gcc.dg/torture/pr49518.c -O3 -g (test for excess errors)
+FAIL: g++.dg/opt/longbranch2.C -std=gnu++98 (test for excess errors)
+FAIL: g++.dg/opt/longbranch2.C -std=gnu++11 (test for excess errors)
+FAIL: libmudflap.c/fail37-frag.c (-O2) (test for excess errors)
+FAIL: libmudflap.c/fail37-frag.c (-O3) (test for excess errors)
(longbranch2.C only on i686, not on x86_64). For fail32-frag.c, I believe we
just want to add asm ("" : "+r" (i)); into the loop to hide stuff from the
optimizers.
With the debugging hack, the cases of increasing number of iterations at the
end of estimate_numbers_of_iterations_loop are (except for pr53265.c, which of
course has tons of them):
64 ../../gcc/ada/exp_ch7.adb:3540 7ffffffd 0 ffffffff 0
{32,64} /usr/src/gcc/gcc/testsuite/gcc.dg/torture/pr49518.c:10 2 0 fffffffd 0
32 /usr/src/gcc/gcc/testsuite/g++.dg/opt/longbranch2.C:36 100 0 1ff 0
32 /usr/src/gcc/gcc/testsuite/g++.dg/opt/longbranch2.C:36 ff 0 1fe 0
{32,64} /usr/src/gcc/gcc/testsuite/gfortran.dg/do_1.f90:10 9 0 a 0
{32,64} /usr/src/gcc/gcc/testsuite/gfortran.dg/do_1.f90:16 4 0 5 0
{32,64} /usr/src/gcc/gcc/testsuite/gfortran.dg/do_1.f90:21 3 0 4 0
{32,64} /usr/src/gcc/libmudflap/testsuite/libmudflap.c/fail37-frag.c:15 4 0 5 0
In exp_ch7.adb it looks like pass ordering, cunrolli is run on dead code there,
*.copyrename2 has:
[../../gcc/ada/exp_ch7.adb : 3540:11] ind.127_3 = 1;
[../../gcc/ada/exp_ch7.adb : 3540:11] if (ind.127_3 > 1)
goto <bb 3>;
else
goto <bb 6>;
<bb 3>:
# fent_39 = PHI <[../../gcc/ada/exp_ch7.adb : 3535:7] fent_2(2)>
# j_40 = PHI <[../../gcc/ada/exp_ch7.adb : 3540:11] 1(2)>
<bb 4>:
# fent_6 = PHI <fent_39(3), [../../gcc/ada/exp_ch7.adb : 3541:10] fent_7(5)>
# j_4 = PHI <j_40(3), [../../gcc/ada/exp_ch7.adb : 3540:11] j_5(5)>
# DEBUG j => j_4
# DEBUG fent => fent_6
[../../gcc/ada/exp_ch7.adb : 3540:11] j_5 = j_4 + 1;
[../../gcc/ada/exp_ch7.adb : 3540:11] # DEBUG j => j_5
[../../gcc/ada/exp_ch7.adb : 3541:10] fent_7 = sinfo.next_entity (fent_6);
[../../gcc/ada/exp_ch7.adb : 3541:10] # DEBUG fent => fent_7
[../../gcc/ada/exp_ch7.adb : 3540:25] if (ind.127_3 != j_5)
goto <bb 5>;
else
goto <bb 6>;
where both j and ind.127 are signed SImode vars. number_of_latch_iterations
estimates the bound as 0xffffffff, signed overflow decreases that guess, but it
is really in dead code (something visible only through inlining though).
ccp2 immediately after cunrolli would fix this up, but that is too late.
For pr49518.c, I guess it is reasonable to warn, if the loop invariant
conditions are all such that it keeps looping, then the loop indeed triggers
undefined behavior.
Thoughts?
