[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 Segher Boessenkool changed: What|Removed |Added Status|NEW |RESOLVED Resolution|--- |FIXED --- Comment #15 from Segher Boessenkool --- Fixed on trunk, doesn't seem to happen on 6, closing as fixed.
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #14 from Segher Boessenkool --- Author: segher Date: Sun Jan 15 17:06:00 2017 New Revision: 244477 URL: https://gcc.gnu.org/viewcvs?rev=244477=gcc=rev Log: Make rtl_split_edge work for jumps that fall through (PR72749) If a jump always falls through but that cannot be optimised away, like is the case with the PowerPC bdnz insn if its jump target is the same as the fallthrough, sched gets confused if it schedules some instructions from before that jump instruction to behind it: it splits the fallthrough branch, but the jump target isn't updated, and things fall apart as in the PR. This patch fixes it. The second patch fragment makes -fsched-verbose=N work for N>=4; the currently scheduled fragment can now contain a label. Everything else seems to work fine with that. PR target/72749 * cfgrtl.c (rtl_split_edge): Also patch jump insns that jump to the fallthrough. * haifa-sched.c (dump_insn_stream): Don't crash if there is a label in the currently scheduled RTL fragment. Modified: trunk/gcc/ChangeLog trunk/gcc/cfgrtl.c trunk/gcc/haifa-sched.c
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #13 from Segher Boessenkool --- I have a patch bootstrapping, let's not close this yet.
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 Alan Modra changed: What|Removed |Added CC|amodra at gcc dot gnu.org | --- Comment #12 from Alan Modra --- No sure whether we want to close this bug yet. I'll leave that to Segher or Jakub. The patch I committed could be seen a just a workaround, unless we're willing to say later rtl passes can count on earlier ones removing the degenerate conditional branches that cause comment #5 problems.
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #11 from Alan Modra --- Author: amodra Date: Sat Jan 14 13:29:29 2017 New Revision: 244466 URL: https://gcc.gnu.org/viewcvs?rev=244466=gcc=rev Log: Testcase from pr72749 PR target/72749 * gcc.c-torture/compile/pr72749.c: New test. Added: trunk/gcc/testsuite/gcc.c-torture/compile/pr72749.c Modified: trunk/gcc/testsuite/ChangeLog
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #10 from Alan Modra --- Author: amodra Date: Sat Jan 14 13:24:46 2017 New Revision: 244465 URL: https://gcc.gnu.org/viewcvs?rev=244465=gcc=rev Log: Avoid PR72749 by not using unspecs Rather than using unspecs in doloop insns to stop combine creating these insns, use legitimate_combined_insn. PR target/72749 * combine.c (recog_for_combine_1): Set INSN_CODE before calling target legitimate_combined_insn. * config/rs6000/rs6000.c (TARGET_LEGITIMATE_COMBINED_INSN): Define. (rs6000_legitimate_combined_insn): New function. * config/rs6000/rs6000.md (UNSPEC_DOLOOP): Delete, and remove all uses. (ctr_internal3): Rename from *ctr_internal5. (ctr_internal4): Rename from *ctr_internal6. (ctr_internal1, ctr_internal2): Remove '*' from name. Modified: trunk/gcc/ChangeLog trunk/gcc/combine.c trunk/gcc/config/rs6000/rs6000.c trunk/gcc/config/rs6000/rs6000.md
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #9 from Segher Boessenkool --- (In reply to Arseny Solokha from comment #7) > I wonder if the following ICE is somehow related to the one reported here. > I'll file a new PR if it's not. This is a different bug (it still happens with my candidate fix for the original issue). In the future, please open new bugs unless you are quite sure it is the same issue; it is much easier and less confusing to merge bugs than to split them. Thanks in advance!
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749
--- Comment #8 from Arseny Solokha ---
int
gd (char ff)
{
int mi;
for (;;)
{
ff += 59 /* 197 also works */;
if (ff != 0)
mi = 0;
else
return 0;
}
}
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #7 from Arseny Solokha --- I wonder if the following ICE is somehow related to the one reported here. I'll file a new PR if it's not. % powerpc-e500v2-linux-gnuspe-gcc-7.0.0-alpha20170108 -O1 -c dsmbs5u1.c dsmbs5u1.c: In function 'gd': dsmbs5u1.c:14:1: error: invalid rtl sharing found in the insn } ^ (insn 40 39 41 2 (set (subreg:QI (reg:SI 172) 3) (subreg:QI (reg:SI 174) 3)) -1 (nil)) dsmbs5u1.c:14:1: error: shared rtx (subreg:QI (reg:SI 172) 3) dsmbs5u1.c:14:1: internal compiler error: internal consistency failure 0x2e55cd6de9b verify_rtx_sharing /var/tmp/portage/cross-powerpc-e500v2-linux-gnuspe/gcc-7.0.0_alpha20170108/work/gcc-7-20170108/gcc/emit-rtl.c:2774 0x2e55cd6ddd7 verify_rtx_sharing /var/tmp/portage/cross-powerpc-e500v2-linux-gnuspe/gcc-7.0.0_alpha20170108/work/gcc-7-20170108/gcc/emit-rtl.c:2789 0x2e55cd6e30e verify_insn_sharing /var/tmp/portage/cross-powerpc-e500v2-linux-gnuspe/gcc-7.0.0_alpha20170108/work/gcc-7-20170108/gcc/emit-rtl.c:2860 0x2e55cd742f7 verify_rtl_sharing() /var/tmp/portage/cross-powerpc-e500v2-linux-gnuspe/gcc-7.0.0_alpha20170108/work/gcc-7-20170108/gcc/emit-rtl.c:2883 0x2e55d058785 execute_function_todo /var/tmp/portage/cross-powerpc-e500v2-linux-gnuspe/gcc-7.0.0_alpha20170108/work/gcc-7-20170108/gcc/passes.c:1982 0x2e55d0595ab execute_todo /var/tmp/portage/cross-powerpc-e500v2-linux-gnuspe/gcc-7.0.0_alpha20170108/work/gcc-7-20170108/gcc/passes.c:2015 % powerpc-e500v2-linux-gnuspe-gcc-7.0.0-alpha20170108 -O1 -c dsmbs5u1.c -mcpu=power7 No ICE % powerpc-e300c3-linux-gnu-gcc-7.0.0-alpha20170108 -O1 -c dsmbs5u1.c No ICE % powerpc-e300c3-linux-gnu-gcc-7.0.0-alpha20170108 -O1 -c dsmbs5u1.c -mcpu=8548 ICE
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 Richard Biener changed: What|Removed |Added Priority|P3 |P1
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #6 from Alan Modra --- OK, currently testing TARGET_LEGITIMATE_COMBINED_INSN solution. Thanks for the analysis!
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #5 from Jakub Jelinek --- I think it is rtl_split_edge and patch_jump_insn not being able to deal with such conditional jump insns. It has code to deal e.g. with degenerate asm goto and similar insns. Note, I'm not really sure what should splitting an edge mean in that case when there is just one edge. Shall it redirect the branch so that both when the condition is false and true the conditional jump points to the new basic block? Then there should be a new CODE_LABEL added to the new basic block and the conditional jump adjusted to branch to that new CODE_LABEL. Or something else?
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 --- Comment #4 from Jakub Jelinek --- Perhaps use the TARGET_LEGITIMATE_COMBINED_INSN target hook? Though if that hook rejects all doloop insns, that would mean you can't combine into the doloop insns created by the doloop pass, not sure if it wouldn't hurt though. Without the UNSPEC, the doloop insn has been removed during the fwprop2 pass, together with tons of other insns, while with the UNSPEC that doesn't happen. If we want to keep the UNSPEC, then we'd need to debug whatever in the sched2 pass decides to remove the CODE_LABEL from the fallthrough block after the degenerate doloop insn.
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 Alan Modra changed: What|Removed |Added CC||amodra at gmail dot com, ||segher at gcc dot gnu.org --- Comment #3 from Alan Modra --- Is there a better way to stop combine creating ctr type insns, to gain the benefits shown in https://gcc.gnu.org/ml/gcc-patches/2016-07/msg00356.html ?
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 Jakub Jelinek changed: What|Removed |Added CC||amodra at gcc dot gnu.org, ||jakub at gcc dot gnu.org --- Comment #2 from Jakub Jelinek --- Started with r238207. So, what I see is that the loop2_doloop pass creates a strange doloop JUMP_INSN, which falls thru into an empty basic block whose single successor is the other successor of the doloop insn. During various passes that doloop insn's bb turns into one that has just a single edge, then again 2 edges and back a single edge, and finally during sched2 some checking ICEs on it. Before r238207 the doloop insn didn't have UNSPEC inside of it and so has been optimized away.
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 Martin Liška changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2016-08-01 CC||marxin at gcc dot gnu.org Ever confirmed|0 |1 --- Comment #1 from Martin Liška --- Confirmed on powerpc64le.
[Bug target/72749] [7 Regression] ICE: verify_flow_info failed (error: wrong amount of branch edges after conditional jump in bb 5) w/ -O2 -fsched2-use-superblocks
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72749 Richard Biener changed: What|Removed |Added Target Milestone|--- |7.0
