https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115841
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
r15-2054-g1e3aa9c9278db6, when backported to the branch, avoids the failure,
it's still latent of course.
The fortran loop is the DO KR=1,NRM loop from
module_mp_fast_sbm.fppized.f90:6100 which is the JERTIMESC subroutine
SUBROUTINE JERTIMESC(FI1,X1,SFN11,SFN12 &
& ,B11_MY,B12_MY,RIEC,CF,ID,COL,NKR)
IMPLICIT NONE
INTEGER NRM,KR,ICE,ID,NKR
REAL B12,B11,FUN,DELM,FK,CF,SFN12S,SFN11S
REAL COL, &
& X1(NKR,ID),FI1(NKR,ID),B11_MY(NKR,ID),B12_MY(NKR,ID) &
&,RIEC(NKR,ID),SFN11,SFN12
NRM=NKR-1
DO 1 ICE=1,ID
SFN11S=0.
SFN12S=0.
SFN11=CF*SFN11S
SFN12=CF*SFN12S
DO KR=1,NRM
! VALUE OF DISTRIBUTION FUNCTION
FK=FI1(KR,ICE)
! DELTA-M
DELM=X1(KR,ICE)*3.*COL
! INTEGRAL'S EXPRESSION
FUN=FK*DELM
! VALUES OF INTEGRALS
B11=B11_MY(KR,ICE)
B12=B12_MY(KR,ICE)
SFN11S=SFN11S+FUN*B11
SFN12S=SFN12S+FUN*B12
ENDDO
! CORRECTION
SFN11=CF*SFN11S
SFN12=CF*SFN12S
1 CONTINUE
! END
RETURN
END SUBROUTINE JERTIMESC
It's an inlined copy in ONECOND1 (and that is a IPA CP clone).
The key to reproduce is the peeling for alignment. We have
module_mp_fast_sbm.fppized.f90:6100:19: note: vectorization_factor = 16,
niters = 32
module_mp_fast_sbm.fppized.f90:6100:19: note: ===
vect_analyze_data_refs_alignment ===
module_mp_fast_sbm.fppized.f90:6100:19: note: recording new base alignment
for &A.170
alignment: 64
misalignment: 0
based on: fk_206 = MEM <float[0:D.7150]> [(float[0:D.7065]
*)&A.170][_205];
module_mp_fast_sbm.fppized.f90:6100:19: note: recording new base alignment
for &xl
alignment: 32
misalignment: 0
based on: _171 = MEM <float[0:D.7174]> [(float[0:D.7069] *)&xl][_205];
module_mp_fast_sbm.fppized.f90:6100:19: note: recording new base alignment
for &A.166
alignment: 64
misalignment: 0
based on: b11_164 = MEM <float[0:D.7153]> [(float[0:D.7075]
*)&A.166][_205];
module_mp_fast_sbm.fppized.f90:6100:19: note: recording new base alignment
for &A.167
alignment: 64
misalignment: 0
based on: b12_161 = MEM <float[0:D.7156]> [(float[0:D.7079]
*)&A.167][_205];
module_mp_fast_sbm.fppized.f90:6100:19: note:
vect_compute_data_ref_alignment:
module_mp_fast_sbm.fppized.f90:6100:19: missed: misalign = 0 bytes of ref MEM
<float[0:D.7150]> [(float[0:D.7065] *)&A.170][_205]
module_mp_fast_sbm.fppized.f90:6100:19: note:
vect_compute_data_ref_alignment:
module_mp_fast_sbm.fppized.f90:6100:19: note: can't force alignment of ref:
MEM <float[0:D.7174]> [(float[0:D.7069] *)&xl][_205]
module_mp_fast_sbm.fppized.f90:6100:19: note:
vect_compute_data_ref_alignment:
module_mp_fast_sbm.fppized.f90:6100:19: missed: misalign = 0 bytes of ref MEM
<float[0:D.7153]> [(float[0:D.7075] *)&A.166][_205]
module_mp_fast_sbm.fppized.f90:6100:19: note:
vect_compute_data_ref_alignment:
module_mp_fast_sbm.fppized.f90:6100:19: missed: misalign = 0 bytes of ref MEM
<float[0:D.7156]> [(float[0:D.7079] *)&A.167][_205]
module_mp_fast_sbm.fppized.f90:6100:19: note: ===
vect_prune_runtime_alias_test_list ===
module_mp_fast_sbm.fppized.f90:6100:19: note: ===
vect_enhance_data_refs_alignment ===
module_mp_fast_sbm.fppized.f90:6100:19: missed: Unknown misalignment,
naturally aligned
module_mp_fast_sbm.fppized.f90:6100:19: note: vect_can_advance_ivs_p:
module_mp_fast_sbm.fppized.f90:6100:19: note: Analyze phi: sfn11s_17 = PHI
<sfn11s_156(69), 0.0(7)>
module_mp_fast_sbm.fppized.f90:6100:19: note: reduc or virtual phi. skip.
module_mp_fast_sbm.fppized.f90:6100:19: note: Analyze phi: sfn12s_2 = PHI
<sfn12s_133(69), 0.0(7)>
...
module_mp_fast_sbm.fppized.f90:6100:19: note: Alignment of access forced
using peeling.
module_mp_fast_sbm.fppized.f90:6100:19: note: Peeling for alignment will be
applied.
where we align the xl load and end up with misaligned others. The C testcase
has the As aligned to 16 and xl aligned to 32. We're doing runtime
alignment of xl with a scalar prologue.
