[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 Jakub Jelinek changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |FIXED --- Comment #11 from Jakub Jelinek --- Fixed.
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 --- Comment #10 from Jakub Jelinek --- Author: jakub Date: Tue Mar 6 20:41:37 2018 New Revision: 258301 URL: https://gcc.gnu.org/viewcvs?rev=258301=gcc=rev Log: PR target/84710 * combine.c (try_combine): Use reg_or_subregno instead of handling just paradoxical SUBREGs and REGs. * gcc.dg/pr84710.c: New test. Added: trunk/gcc/testsuite/gcc.dg/pr84710.c Modified: trunk/gcc/ChangeLog trunk/gcc/combine.c trunk/gcc/testsuite/ChangeLog
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 --- Comment #9 from Jakub Jelinek --- (In reply to Segher Boessenkool from comment #8) > Ah right, I had checking disabled. Need more coffee I guess. You don't really need to build the checking compiler for this though, it is enough to put a breakpoint on that REGNO (x) line and debug_rtx (x) first. > The insns I see are different: > > Trying 19, 20 -> 21: >19: r106:SI#0=0 >20: r108:SI=zero_extend(r106:SI#0) >21: r109:SI=r108:SI 0>>0x8 > REG_DEAD r108:SI > > (see insn 20); maybe the i3 in #c1 is already modified? Likely, yes. I've just done up and debug_rtx on the 3 insns passed to try_combine. > The #c5 patch is okay for trunk, thanks! Thanks.
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 --- Comment #8 from Segher Boessenkool --- Ah right, I had checking disabled. Need more coffee I guess. The insns I see are different: Trying 19, 20 -> 21: 19: r106:SI#0=0 20: r108:SI=zero_extend(r106:SI#0) 21: r109:SI=r108:SI 0>>0x8 REG_DEAD r108:SI (see insn 20); maybe the i3 in #c1 is already modified? The #c5 patch is okay for trunk, thanks!
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 --- Comment #7 from Zdenek Sojka --- Created attachment 43573 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=43573=edit original testcase Unreduced testcase; needs additional -fkeep-inline-functions: $ aarch64-unknown-linux-gnu-gcc -O -fno-forward-propagate -fkeep-inline-functions xxx.c -w during RTL pass: combine xxx.c: In function 'bar2': xxx.c:81:1: internal compiler error: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 } ...
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 --- Comment #6 from Zdenek Sojka --- (In reply to Segher Boessenkool from comment #4) > The testcase has UB of course (d is undefined). The original testcase had defined behavior. The uninitialized use is the result of delta/creduce reduction. There was an ICE with the unreduced testcase as well.
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 --- Comment #5 from Jakub Jelinek --- (In reply to Segher Boessenkool from comment #4) > The testcase does not crash for me; does it need more non-default options? > > The testcase has UB of course (d is undefined). The compiler needs to be configured with --enable-checking=yes,rtl . --- gcc/combine.c.jj2018-03-05 23:13:26.478215559 +0100 +++ gcc/combine.c 2018-03-06 08:50:17.756288841 +0100 @@ -4283,12 +4283,8 @@ try_combine (rtx_insn *i3, rtx_insn *i2, if (GET_CODE (x) == PARALLEL) x = XVECEXP (newi2pat, 0, 0); - /* It can only be a SET of a REG or of a paradoxical SUBREG of a REG. */ - x = SET_DEST (x); - if (paradoxical_subreg_p (x)) - x = SUBREG_REG (x); - - unsigned int regno = REGNO (x); + /* It can only be a SET of a REG or of a SUBREG of a REG. */ + unsigned int regno = reg_or_subregno (SET_DEST (x)); bool done = false; for (rtx_insn *insn = NEXT_INSN (i3); is the combine.c part I'll be testing. And, the testcase is dg-do compile only, so it doesn't matter if there is UB in it or not, we shouldn't ICE on it (and yes, my attempts to get rid of the uninit use just resulted in the ICE no longer happening).
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 --- Comment #4 from Segher Boessenkool --- The testcase does not crash for me; does it need more non-default options? The testcase has UB of course (d is undefined).
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 --- Comment #3 from Segher Boessenkool --- Well, yuck, I missed that part of rtl.texi (or somehow I thought this RTL would be refused earlier). Please use reg_or_subregno in the patch though?
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 Jakub Jelinek changed: What|Removed |Added Status|NEW |ASSIGNED Assignee|unassigned at gcc dot gnu.org |jakub at gcc dot gnu.org --- Comment #2 from Jakub Jelinek --- Created attachment 43572 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=43572=edit gcc8-pr84710.patch Full untested patch.
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710
Jakub Jelinek changed:
What|Removed |Added
Status|UNCONFIRMED |NEW
Last reconfirmed||2018-03-06
CC||jakub at gcc dot gnu.org,
||segher at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #1 from Jakub Jelinek ---
This is on try_combine with
i3:
(insn 21 20 22 2 (set (reg:SI 109)
(const_int 0 [0])) "pr84710.c":8 679 {*aarch64_lshr_sisd_or_int_si3}
(expr_list:REG_DEAD (reg:SI 108)
(nil)))
i2:
(insn 20 19 21 2 (set (reg:SI 108)
(zero_extend:SI (subreg:HI (reg:SI 106) 0))) "pr84710.c":8 89
{*zero_extendhisi2_aarch64}
(nil))
i1:
(insn 19 18 20 2 (set (subreg:HI (reg:SI 106) 0)
(const_int 0 [0])) "pr84710.c":8 45 {*movhi_aarch64}
(nil))
and on:
4286 /* It can only be a SET of a REG or of a paradoxical SUBREG of a
REG. */
4287 x = SET_DEST (x);
4288 if (paradoxical_subreg_p (x))
4289x = SUBREG_REG (x);
4290
4291 unsigned int regno = REGNO (x);
x is:
(subreg:HI (reg:SI 106) 0)
rtl.texi documents the behavior of SUBREGs as lvalues not just for paradoxical
SUBREGs, but also for normal SUBREGs.
So, we have a proof that the paradoxical_subreg_p guard is insufficient.
I'd go with:
--- gcc/combine.c.jj2018-03-05 23:13:26.478215559 +0100
+++ gcc/combine.c 2018-03-06 08:50:17.756288841 +0100
@@ -4283,9 +4283,9 @@ try_combine (rtx_insn *i3, rtx_insn *i2,
if (GET_CODE (x) == PARALLEL)
x = XVECEXP (newi2pat, 0, 0);
- /* It can only be a SET of a REG or of a paradoxical SUBREG of a REG.
*/
+ /* It can only be a SET of a REG or of a SUBREG of a REG. */
x = SET_DEST (x);
- if (paradoxical_subreg_p (x))
+ if (SUBREG_P (x))
x = SUBREG_REG (x);
unsigned int regno = REGNO (x);
(until proven otherwise).
[Bug target/84710] [8 Regression] ICE: RTL check: expected code 'reg', have 'subreg' in rhs_regno, at rtl.h:1896 with -O -fno-forward-propagate
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84710 Richard Biener changed: What|Removed |Added Priority|P3 |P1 Target Milestone|--- |8.0
