The following code compiles at O2, but not O3. ~/test > gfortran481 -O3 -c module_diffusion_em.f90 module_diffusion_em.f90: In function ‘cal_helicity’: module_diffusion_em.f90:1:0: internal compiler error: in eliminate_temp_copies, at tree-predcom.c:1913 SUBROUTINE cal_helicity (uh, ph, phb, ht, ims, ime, jms, jme, kms, kme, & ^ 0xc81b7a eliminate_temp_copies ../../gcc-4.8.1/gcc/tree-predcom.c:1913 0xc81b7a tree_predictive_commoning_loop ../../gcc-4.8.1/gcc/tree-predcom.c:2519 0xc82277 tree_predictive_commoning() ../../gcc-4.8.1/gcc/tree-predcom.c:2554 Please submit a full bug report, with preprocessed source if appropriate. Please include the complete backtrace with any bug report. See <http://gcc.gnu.org/bugs.html> for instructions. ~/test > cat module_diffusion_em.f90 SUBROUTINE cal_helicity (uh, ph, phb, ht, ims, ime, jms, jme, kms, kme, & its, ite, jts, jte, kts, kte) INTEGER, INTENT( IN ) :: ims, ime, jms, jme, kms, kme, & its, ite, jts, jte, kts, kte REAL, DIMENSION( ims:ime , jms:jme ), INTENT( IN ) :: ht REAL, DIMENSION( ims:ime, kms:kme, jms:jme ), INTENT( IN ) :: ph, phb REAL, DIMENSION( ims:ime, jms:jme ), INTENT( INOUT ) :: uh INTEGER :: i, j, k, ktf, i_start, i_end, j_start, j_end REAL :: zl, zu REAL, DIMENSION( its-2:ite+2, kts:kte, jts-2:jte+2 ) :: wavg, rvort real, parameter :: g = 9.81 DO j = j_start, j_end DO k = kts, ktf DO i = i_start, i_end zu = (( ph(i ,k+1,j ) + phb(i ,k+1,j ) ) / g - ht(i ,j ) ) + & (( ph(i-1,k+1,j ) + phb(i-1,k+1,j ) ) / g - ht(i-1,j ) ) + & (( ph(i ,k+1,j-1) + phb(i ,k+1,j-1) ) / g - ht(i ,j-1) ) IF ( zl .GE. 2000. .AND. zu .LE. 5000. ) THEN IF ( wavg(i,k,j) .GT. 0. .AND. wavg(i,k+1,j) .GT. 0. ) THEN uh(i,j) = uh(i,j) + ( wavg(i,k,j) * rvort(i,k,j) + & wavg(i,k+1,j) * rvort(i,k+1,j) ) * ( zu - zl ) ENDIF ENDIF END DO END DO END DO END SUBROUTINE cal_helicity ~/test > gfortran481 -v Using built-in specs. COLLECT_GCC=/home/decker/local/gcc481/bin/gfortran481 COLLECT_LTO_WRAPPER=/home/decker/local/gcc481/libexec/gcc/x86_64-unknown-linux-gnu/4.8.1/lto-wrapper Target: x86_64-unknown-linux-gnu Configured with: ../gcc-4.8.1/configure --prefix=/home/decker/local/gcc481 --program-suffix=481 --enable-languages=c,c++,fortran --with-gmp=/home/decker/local --with-mpfr=/home/decker/local --with-mpc=/home/decker/local --with-isl=/home/decker/local --with-cloog=/home/decker/local Thread model: posix gcc version 4.8.1 (GCC)
Confirmed. > ./f951 -quiet t.f90 -O3 t.f90: In function 'cal_helicity': t.f90:1:0: error: definition in block 9 does not dominate use in block 35 SUBROUTINE cal_helicity (uh, ph, phb, ht, ims, ime, jms, jme, kms, kme, & ^ for SSA_NAME: D_lsm0.70_121 in statement: D_lsm0.70_120 = PHI <_314(33), D_lsm0.70_121(35)> PHI argument D_lsm0.70_121 for PHI node D_lsm0.70_120 = PHI <_314(33), D_lsm0.70_121(35)> t.f90:1:0: internal compiler error: verify_ssa failed 0xe45652 verify_ssa(bool) /space/rguenther/src/svn/trunk/gcc/tree-ssa.c:1194 0xd9b5f8 verify_loop_closed_ssa(bool) /space/rguenther/src/svn/trunk/gcc/tree-ssa-loop-manip.c:590 0xd9bcd9 gimple_duplicate_loop_to_header_edge(loop*, edge_def*, unsigned int, simple_bitmap_def*, edge_def*, vec<edge_def*, va_heap, vl_ptr>*, int) /space/rguenther/src/svn/trunk/gcc/tree-ssa-loop-manip.c:761 0xd9cd38 tree_transform_and_unroll_loop(loop*, unsigned int, edge_def*, tree_niter_desc*, void (*)(loop*, void*), void*) /space/rguenther/src/svn/trunk/gcc/tree-ssa-loop-manip.c:1179 0xcf0bc6 tree_predictive_commoning_loop /space/rguenther/src/svn/trunk/gcc/tree-predcom.c:2520 gone latent on the trunk for me, still broken on the 4.8 branch. I will have a look.
Ok, so we are having a combined chain where the combination is in a conditional path. Reduced testcase: SUBROUTINE cal_helicity (uh, ph, phb, wavg, ims, ime, its, ite) INTEGER, INTENT( IN ) :: ims, ime, its, ite REAL, DIMENSION( ims:ime), INTENT( IN ) :: ph, phb, wavg REAL, DIMENSION( ims:ime), INTENT( INOUT ) :: uh INTEGER :: i REAL :: zu DO i = its, ite zu = (ph(i ) + phb(i)) + (ph(i-1) + phb(i-1)) IF (wavg(i) .GT. 0) THEN uh(i) = uh(i) + zu ENDIF END DO END SUBROUTINE cal_helicity which again fails on trunk as well.
We also fail to sink the loads (because appearantly we don't want to do that, eh), which in the end causes this havoc (not-always-executed combination). C testcase: int x[1024], y[1024], z[1024], w[1024]; void foo (void) { int i; for (i = 1; i < 1024; ++i) { int a = x[i]; int b = y[i]; int c = x[i-1]; int d = y[i-1]; if (w[i]) z[i] = (a + b) + (c + d); } } Fixed by for example the simple Index: gcc/tree-predcom.c =================================================================== --- gcc/tree-predcom.c (revision 204948) +++ gcc/tree-predcom.c (working copy) @@ -2068,7 +2068,11 @@ combinable_refs_p (dref r1, dref r2, stmt = find_common_use_stmt (&name1, &name2); - if (!stmt) + if (!stmt + /* A simple post-dominance check - make sure the combination + is executed under the same condition as the references. */ + || (gimple_bb (stmt) != gimple_bb (r1->stmt) + && gimple_bb (stmt) != gimple_bb (r2->stmt))) return false; acode = gimple_assign_rhs_code (stmt); Zdenek?
Author: rguenth Date: Tue Nov 19 09:21:07 2013 New Revision: 205010 URL: http://gcc.gnu.org/viewcvs?rev=205010&root=gcc&view=rev Log: 2013-11-19 Richard Biener <rguenther@suse.de> PR tree-optimization/57517 * tree-predcom.c (combinable_refs_p): Verify the combination is always executed when the refs are. * gfortran.fortran-torture/compile/pr57517.f90: New testcase. * gcc.dg/torture/pr57517.c: Likewise. Added: trunk/gcc/testsuite/gcc.dg/torture/pr57517.c trunk/gcc/testsuite/gfortran.fortran-torture/compile/pr57517.f90 Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-predcom.c
Author: rguenth Date: Tue Nov 19 09:23:52 2013 New Revision: 205012 URL: http://gcc.gnu.org/viewcvs?rev=205012&root=gcc&view=rev Log: 2013-11-19 Richard Biener <rguenther@suse.de> PR tree-optimization/57517 * tree-predcom.c (combinable_refs_p): Verify the combination is always executed when the refs are. * gfortran.fortran-torture/compile/pr57517.f90: New testcase. * gcc.dg/torture/pr57517.c: Likewise. Added: branches/gcc-4_8-branch/gcc/testsuite/gcc.dg/torture/pr57517.c branches/gcc-4_8-branch/gcc/testsuite/gfortran.fortran-torture/compile/pr57517.f90 Modified: branches/gcc-4_8-branch/gcc/ChangeLog branches/gcc-4_8-branch/gcc/testsuite/ChangeLog branches/gcc-4_8-branch/gcc/tree-predcom.c
Fixed for trunk and 4.8.3 for now.
Author: rguenth Date: Tue Mar 18 08:46:21 2014 New Revision: 208632 URL: http://gcc.gnu.org/viewcvs?rev=208632&root=gcc&view=rev Log: 2014-03-18 Richard Biener <rguenther@suse.de> Backport from mainline 2013-08-27 Richard Biener <rguenther@suse.de> PR tree-optimization/57521 * tree-if-conv.c (if_convertible_bb_p): Verify that at least one edge is non-critical. (find_phi_replacement_condition): Make sure to use a non-critical edge. Cleanup and remove old bug workarounds. (bb_postdominates_preds): Remove. (if_convertible_loop_p_1): Do not compute post-dominators. (combine_blocks): Do not free post-dominators. (main_tree_if_conversion): Likewise. * gcc.dg/torture/pr57521.c: New testcase. 2013-09-03 Richard Biener <rguenther@suse.de> PR middle-end/57656 * fold-const.c (negate_expr_p): Fix division case. (negate_expr): Likewise. * gcc.dg/torture/pr57656.c: New testcase. 2013-11-19 Richard Biener <rguenther@suse.de> PR tree-optimization/57517 * tree-predcom.c (combinable_refs_p): Verify the combination is always executed when the refs are. * gfortran.fortran-torture/compile/pr57517.f90: New testcase. * gcc.dg/torture/pr57517.c: Likewise. Added: branches/gcc-4_7-branch/gcc/testsuite/gcc.dg/torture/pr57517.c branches/gcc-4_7-branch/gcc/testsuite/gcc.dg/torture/pr57521.c branches/gcc-4_7-branch/gcc/testsuite/gcc.dg/torture/pr57656.c branches/gcc-4_7-branch/gcc/testsuite/gfortran.fortran-torture/compile/pr57517.f90 Modified: branches/gcc-4_7-branch/gcc/ChangeLog branches/gcc-4_7-branch/gcc/fold-const.c branches/gcc-4_7-branch/gcc/testsuite/ChangeLog branches/gcc-4_7-branch/gcc/tree-if-conv.c branches/gcc-4_7-branch/gcc/tree-predcom.c
Fixed.