[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 --- Comment #10 from Jakub Jelinek --- Author: jakub Date: Fri Aug 30 11:49:11 2019 New Revision: 275106 URL: https://gcc.gnu.org/viewcvs?rev=275106&root=gcc&view=rev Log: Backported from mainline 2019-01-16 David Malcolm PR target/88861 * combine.c (delete_noop_moves): Convert to "bool" return, returning true if any edges are eliminated. (combine_instructions): Also return true if delete_noop_moves returns true. * g++.dg/opt/pr89188.C: Include ../torture/pr88861.C. 2019-01-16 David Malcolm PR target/88861 * g++.dg/torture/pr88861.C: New test. Added: branches/gcc-7-branch/gcc/testsuite/g++.dg/torture/pr88861.C Modified: branches/gcc-7-branch/gcc/ChangeLog branches/gcc-7-branch/gcc/combine.c branches/gcc-7-branch/gcc/testsuite/ChangeLog branches/gcc-7-branch/gcc/testsuite/g++.dg/opt/pr89188.C
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 --- Comment #9 from Jakub Jelinek --- Author: jakub Date: Sat Feb 9 08:51:41 2019 New Revision: 268717 URL: https://gcc.gnu.org/viewcvs?rev=268717&root=gcc&view=rev Log: PR middle-end/89243 * g++.dg/opt/pr89188.C: Include ../torture/pr88861.C. Backported from mainline 2019-01-16 David Malcolm PR target/88861 * combine.c (delete_noop_moves): Convert to "bool" return, returning true if any edges are eliminated. (combine_instructions): Also return true if delete_noop_moves returns true. * g++.dg/torture/pr88861.C: New test. Added: branches/gcc-8-branch/gcc/testsuite/g++.dg/torture/pr88861.C Modified: branches/gcc-8-branch/gcc/ChangeLog branches/gcc-8-branch/gcc/combine.c branches/gcc-8-branch/gcc/testsuite/ChangeLog branches/gcc-8-branch/gcc/testsuite/g++.dg/opt/pr89188.C
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 David Malcolm changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |FIXED --- Comment #8 from David Malcolm --- Should be fixed by r267984.
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 --- Comment #7 from David Malcolm --- Author: dmalcolm Date: Wed Jan 16 20:13:23 2019 New Revision: 267984 URL: https://gcc.gnu.org/viewcvs?rev=267984&root=gcc&view=rev Log: Fix ICE due to "combine" creating unreachable EH blocks (PR target/88861) PR target/88861 reports an ICE in "ce2" due to an unreachable basic block. The block becomes unreachable in "combine" when delete_noop_moves deletes an insn with a REG_EH_REGION, deleting the EH edge, the only edge leading to the basic block. Normally, rest_of_handle_combine would call cleanup_cfg, deleting unreachable blocks, if combine_instructions returns true, and combine_instructions does return true for some cases of edge-removal, but it doesn't for this case, leading to the ICE. This patch updates delete_noop_moves so that it returns true if it deletes any edges, and passes that through to combine_instructions, so that it too will return true if any edges were deleted, ensuring that cleanup_cfg will be called by rest_of_handle_combine for this case, deleting the now-unreachable block, and fixing the ICE. gcc/ChangeLog: PR target/88861 * combine.c (delete_noop_moves): Convert to "bool" return, returning true if any edges are eliminated. (combine_instructions): Also return true if delete_noop_moves returns true. gcc/testsuite/ChangeLog: PR target/88861 * g++.dg/torture/pr88861.C: New test. Added: trunk/gcc/testsuite/g++.dg/torture/pr88861.C Modified: trunk/gcc/ChangeLog trunk/gcc/combine.c trunk/gcc/testsuite/ChangeLog
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 --- Comment #6 from David Malcolm --- (In reply to Segher Boessenkool from comment #5) > Cool, thanks! Is the plan to simply not allow something that can throw to be > recognised as noop move? Candidate patch: https://gcc.gnu.org/ml/gcc-patches/2019-01/msg00914.html (caveat: I'm relatively new to this code)
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 --- Comment #5 from Segher Boessenkool --- Cool, thanks! Is the plan to simply not allow something that can throw to be recognised as noop move?
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 David Malcolm changed: What|Removed |Added Status|NEW |ASSIGNED Assignee|unassigned at gcc dot gnu.org |dmalcolm at gcc dot gnu.org --- Comment #4 from David Malcolm --- Am testing a fix.
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861
--- Comment #3 from David Malcolm ---
Assertion fails in dom_info::calc_dfs_tree:
457 /* This aborts e.g. when there is _no_ path from ENTRY to EXIT at
all. */
458 gcc_assert (m_nodes == (unsigned int) m_n_basic_blocks - 1);
(gdb) p m_nodes
$8 = 9
(gdb) p m_n_basic_blocks
$9 = 11
Block 9 in .263r.ud_dce has one in-edge from block 6 (insn 29), and no out-edge
Block 9 in .264r.combine has no in-edges or out-edges
Within .263r.ud_dce:
(note 25 69 26 6 [bb 6] NOTE_INSN_BASIC_BLOCK)
(insn 26 25 28 6 (set (reg/f:DI 142)
(mem/u/c:DI (unspec:DI [
(symbol_ref/u:DI ("*.LC1") [flags 0x2])
(reg:DI 2 2)
] UNSPEC_TOCREL) [3 S8 A8])) "/tmp/test.c":7:21 608
{*movdi_internal64}
(expr_list:REG_EQUAL (symbol_ref:DI ("*.LANCHOR1") [flags 0x182])
(nil)))
(insn 28 26 29 6 (set (reg:SI 135 [ i ])
(mem/c:SI (plus:DI (reg/f:DI 142)
(const_int 4 [0x4])) [1 i+0 S4 A32])) "/tmp/test.c":7:21 494
{*movsi_internal1}
(expr_list:REG_EQUAL (mem/c:SI (const:DI (plus:DI (symbol_ref:DI
("*.LANCHOR1") [flags 0x182])
(const_int 4 [0x4]))) [1 i+0 S4 A32])
(nil)))
(insn 29 28 30 6 (set (mem:SI (plus:DI (reg/f:DI 142)
(const_int 4 [0x4])) [1 s+4 S4 A32])
(reg:SI 135 [ i ])) "/tmp/test.c":7:21 494 {*movsi_internal1}
(expr_list:REG_DEAD (reg:SI 135 [ i ])
(expr_list:REG_EH_REGION (const_int 1 [0x1])
(nil
...but within .264r.combine:
[...snip...]
allowing combination of insns 28 and 29
original costs 4 + 4 = 8
replacement cost 0
deferring deletion of insn with uid = 28.
modifying insn i329: [r142:DI+0x4]=[r142:DI+0x4]
REG_EH_REGION 0x1
deferring rescan insn with uid = 29.
Can't combine i2 into i3
Can't combine i2 into i3
Can't combine i2 into i3
Can't combine i2 into i3
Can't combine i2 into i3
Can't combine i2 into i3
Can't combine i2 into i3
Can't combine i2 into i3
deleting noop move 29
deferring deletion of insn with uid = 29.
starting the processing of deferred insns
rescanning insn with uid = 23.
ending the processing of deferred insns
[...snip...]
(note 25 69 26 6 [bb 6] NOTE_INSN_BASIC_BLOCK)
(insn 26 25 28 6 (set (reg/f:DI 142)
(mem/u/c:DI (unspec:DI [
(symbol_ref/u:DI ("*.LC1") [flags 0x2])
(reg:DI 2 2)
] UNSPEC_TOCREL) [3 S8 A8])) "/tmp/test.c":7:21 608
{*movdi_internal64}
(expr_list:REG_EQUAL (symbol_ref:DI ("*.LANCHOR1") [flags 0x182])
(nil)))
(note 28 26 30 6 NOTE_INSN_DELETED)
which, if I'm reading it right, seems to have deleted the usage of EH region 1,
and has left block 9 orphaned.
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 David Malcolm changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2019-01-15 CC||dmalcolm at gcc dot gnu.org Ever confirmed|0 |1 --- Comment #2 from David Malcolm --- Confirmed (with target==ppc64le-redhat-linux)
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 --- Comment #1 from Richard Biener --- That means we have unreachable blocks.
[Bug target/88861] [9 Regression] ICE in calc_dfs_tree, at dominance.c:458
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88861 Martin Liška changed: What|Removed |Added Known to work||8.2.0 Target Milestone|--- |9.0 Known to fail||9.0
