[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 Andreas Krebbel changed: What|Removed |Added CC||stli at linux dot ibm.com --- Comment #15 from Andreas Krebbel --- *** Bug 98269 has been marked as a duplicate of this bug. ***
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 Jeffrey A. Law changed: What|Removed |Added Status|NEW |RESOLVED Resolution|--- |FIXED --- Comment #14 from Jeffrey A. Law --- Fixed by Bin's patch on the trunk.
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 --- Comment #13 from amker at gcc dot gnu.org --- Author: amker Date: Fri Jan 27 14:42:23 2017 New Revision: 244979 URL: https://gcc.gnu.org/viewcvs?rev=244979=gcc=rev Log: PR rtl-optimization/78559 * combine.c (try_combine): Discard REG_EQUAL and REG_EQUIV for other_insn in combine. gcc/testsuite PR rtl-optimization/78559 * gcc.c-torture/execute/pr78559.c: New test. Added: trunk/gcc/testsuite/gcc.c-torture/execute/pr78559.c Modified: trunk/gcc/ChangeLog trunk/gcc/combine.c trunk/gcc/testsuite/ChangeLog
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 --- Comment #12 from Bernd Schmidt --- Sorry, long pause while editing that comment made me leave out part of what I was trying to say - I meant only discard notes that reference the CC reg. But it seems an unnecessary complication.
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 --- Comment #11 from Bernd Schmidt --- Looks like other_insn is only used for cases where we rewrite cc sets in this way, so Bin's patch does look reasonably narrow. We could maybe record the CC reg being changed and only discard reg notes, but in my testing I've not been able to produce code generation differences except for the testcase. Let's wait for Segher's testing, but I think the patch is OK.
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 --- Comment #10 from Segher Boessenkool --- I am leaning toward accepting Bin's patch, but testing whether that hurts generated code too much still hasn't finished.
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 Bernd Schmidt changed: What|Removed |Added CC||bernds at gcc dot gnu.org Assignee|unassigned at gcc dot gnu.org |bernds at gcc dot gnu.org --- Comment #9 from Bernd Schmidt --- Ok, I'll look.
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 Jeffrey A. Law changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2017-01-24 Ever confirmed|0 |1
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 Jeffrey A. Law changed: What|Removed |Added CC||bernds at redhat dot com --- Comment #8 from Jeffrey A. Law --- Bernd. There's a belief this may be a combiner issue, so it might be something you'd want to poke at. I can still trigger this failure on the trunk using -ftree-loop-if-convert on an aarch64 native. If you need access to an aarch64 system, let me know, I've got one provisioned.
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 Jeffrey A. Law changed: What|Removed |Added Priority|P3 |P1 CC||law at redhat dot com
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 --- Comment #7 from amker at gcc dot gnu.org --- (In reply to Jakub Jelinek from comment #6) > So do you think combiner should throw away the REG_EQUAL note in that case, > or something different? It's invalid after transformation, so should be discarded in this case. What I am not sure is that how to tell invalid ones from all notes and keep others. I sent a proposed patch for discussion before @https://gcc.gnu.org/ml/gcc-patches/2016-12/msg00026.html Thanks.
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 Jakub Jelinek changed: What|Removed |Added CC||jakub at gcc dot gnu.org, ||segher at gcc dot gnu.org --- Comment #6 from Jakub Jelinek --- So do you think combiner should throw away the REG_EQUAL note in that case, or something different?
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559
--- Comment #5 from amker at gcc dot gnu.org ---
(insn 37 35 39 7 (set (reg:SI 96)
(sign_extend:SI (subreg:QI (reg:SI 95) 0))) 86 {*extendqisi2_aarch64}
(expr_list:REG_DEAD (reg:SI 95)
(nil)))
(insn 39 37 40 7 (set (reg:CC 66 cc)
(compare:CC (reg:SI 96)
(const_int 0 [0]))) 392 {cmpsi}
(expr_list:REG_DEAD (reg:SI 96)
(nil)))
(insn 40 39 41 7 (set (reg:HI 93)
(if_then_else:HI (ne (reg:CC 66 cc)
(const_int 0 [0]))
(subreg/s/u:HI (reg:SI 81 [ iftmp.0_17 ]) 0)
(reg:HI 97))) 443 {*cmovhi_insn}
(expr_list:REG_DEAD (reg:CC 66 cc)
(expr_list:REG_EQUAL (if_then_else:HI (ne (reg:CC 66 cc)
(const_int 0 [0]))
(subreg/s/u:HI (reg:SI 81 [ iftmp.0_17 ]) 0)
(const_int 1 [0x1]))
(nil
-->
Trying 37 -> 39:
Failed to match this instruction:
(set (reg:CC 66 cc)
(reg:CC 66 cc))
Successfully matched this instruction:
(set (reg:HI 93)
(if_then_else:HI (le (reg:CC 66 cc)
(const_int 0 [0]))
(subreg/s/u:HI (reg:SI 81 [ iftmp.0_17 ]) 0)
(reg:HI 97)))
allowing combination of insns 37 and 39
original costs 4 + 4 = 12
replacement cost 8
deferring deletion of insn with uid = 37.
modifying other_insn40: r93:HI={(cc:CC<=0)?r81:SI#0:r97:HI}
REG_DEAD cc:CC
REG_EQUAL {(cc:CC!=0)?r81:SI#0:0x1}
deferring rescan insn with uid = 40.
modifying insn i339: cc:CC=cc:CC
deferring rescan insn with uid = 39.
Note after combine:
40: r93:HI={(cc:CC<=0)?r81:SI#0:r97:HI}
REG_DEAD cc:CC
REG_EQUAL {(cc:CC!=0)?r81:SI#0:0x1}
cc register has been changed, and (cc<=0) != (cc!=0).
[Bug rtl-optimization/78559] [7 Regression] wrong code due to tree if-conversion?
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78559 amker at gcc dot gnu.org changed: What|Removed |Added Component|middle-end |rtl-optimization --- Comment #4 from amker at gcc dot gnu.org --- I think it's a combine issue reveal by that change. I am testing a patch for public discussion, guess Segher may have better solutions.
