Bug 94238 - [10 Regression] ICE in ix86_cc_mode, at config/i386/i386.c:15285 since r10-7268-g529ea7d9596b26ba
Summary: [10 Regression] ICE in ix86_cc_mode, at config/i386/i386.c:15285 since r10-72...
Alias: None
Product: gcc
Classification: Unclassified
Component: target (show other bugs)
Version: 10.0
: P1 normal
Target Milestone: 10.0
Assignee: Jeffrey A. Law
Keywords: ice-on-valid-code
: 94144 (view as bug list)
Depends on:
Reported: 2020-03-20 16:21 UTC by Martin Liška
Modified: 2020-03-24 00:00 UTC (History)
5 users (show)

See Also:
Host: x86_64-linux-gnu
Target: x86_64-linux-gnu
Known to work: 9.3.0
Known to fail: 10.0
Last reconfirmed: 2020-03-20 00:00:00


Note You need to log in before you can comment on or make changes to this bug.
Description Martin Liška 2020-03-20 16:21:55 UTC
I see the following ICE (isolated from linux kernel):

$ cat kernel.i
enum { false, true } a;
int b, c, d, e, f;
int fn3();
void fn2();

void fn1() {
  _Bool g, h = false, i = false;
  int j;
  c = b && f || d;
  if (c) {
    if (d)
      i = true;
    _Bool k = b;
    int l = e, m = a;
    g = k && l < m || l > m;
  if (g)
    h = true;
  if (i)
  h &&j &&fn3();

$ gcc kernel.i -c -O2
during RTL pass: combine
kernel.i: In function ‘fn1’:
kernel.i:22:1: internal compiler error: in ix86_cc_mode, at config/i386/i386.c:15285
   22 | }
      | ^
0x783f93 ix86_cc_mode(rtx_code, rtx_def*, rtx_def*)
0x16d3bef simplify_set
0x16d5887 combine_simplify_rtx
0x16d74ca subst
0x16d7674 subst
0x16da961 try_combine
0x16e1532 combine_instructions
0x16e1532 rest_of_handle_combine
0x16e1532 execute
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.
Comment 1 Uroš Bizjak 2020-03-20 16:59:54 UTC
Middle end is asking:

(gdb) up
#3  0x0000000001621333 in simplify_set (x=0x7fffea968a68) at /home/uros/git/gcc/gcc/combine.c:6938
6938            compare_mode = SELECT_CC_MODE (new_code, op0, op1);
(gdb) p new_code
$5 = LTGT
(gdb) p debug_rtx (op0)
(mem/c:SI (symbol_ref:DI ("e") [flags 0x2] <var_decl 0x7ffff7ffbd80 e>) [1 e+0 S4 A32])
$6 = void
(gdb) p debug_rtx (op1)
(mem/c:SI (symbol_ref:DI ("a") [flags 0x2] <var_decl 0x7ffff7ffbb40 a>) [1 a+0 S4 A32])
$7 = void

LTGT is invalid for integer operands.
Comment 2 Jeffrey A. Law 2020-03-20 18:10:07 UTC
Just getting started here.  But at least CSE seems to be doing exactly what we want.

We have this going into cse1:
(insn 15 14 16 4 (set (mem/c:SI (symbol_ref:DI ("c") [flags 0x2]  <var_decl 0x7f642a835c60 c>) [1 c+0 S4 A32])
        (const_int 1 [0x1])) "j.c":9:5 67 {*movsi_internal}
(insn 16 15 17 4 (set (reg/v:SI 94 [ l ])
        (mem/c:SI (symbol_ref:DI ("e") [flags 0x2]  <var_decl 0x7f642a835d80 e>) [1 e+0 S4 A32])) "j.c":14:9 67 {*movsi_internal}
(insn 17 16 18 4 (set (reg/v:SI 95 [ m ])
        (mem/c:SI (symbol_ref:DI ("a") [flags 0x2]  <var_decl 0x7f642a835b40 a>) [1 a+0 S4 A32])) "j.c":14:16 67 {*movsi_internal}
(insn 18 17 19 4 (set (reg:CCGC 17 flags)
        (compare:CCGC (reg/v:SI 94 [ l ])
            (reg/v:SI 95 [ m ]))) "j.c":15:16 11 {*cmpsi_1}
(insn 19 18 20 4 (set (reg:QI 103)
        (lt:QI (reg:CCGC 17 flags)
            (const_int 0 [0]))) "j.c":15:16 732 {*setcc_qi}
(insn 20 19 21 4 (set (reg:CCGC 17 flags)
        (compare:CCGC (reg/v:SI 94 [ l ])
            (reg/v:SI 95 [ m ]))) "j.c":15:20 11 {*cmpsi_1}

When processing reg18 we record that (reg 17) has the same value as (compare (reg 94) (reg 95)).  When we start processing insn 20, we find that equivalence.  So if we conceptually replace the RHS of insn 20 with (reg 17) it becomes a nop-set which is now discovered.  This results in cse removing insn 20.  This is good :-)

Introduction of the LTGT happens somewhere in combine.  I'm still digging into that.
Comment 3 Jeffrey A. Law 2020-03-20 18:54:53 UTC
Got it.  simplify-rtx does not validate that a comparison opcode it simplifies to is valid for the mode.    Fixing looks trivial.  Unfortunately something happened overnight that is causing regressions all over the tester, so I want to fix that first, then throw my simpify-rtx fix into the tester.  Regardless, I'll own this.

Essentially it has

(ior (lt (...)) (gt (...)) and simplifies that to (ltgt (...)) which isn't valid for integer or vector modes.
Comment 4 Jeffrey A. Law 2020-03-23 19:02:18 UTC
*** Bug 94144 has been marked as a duplicate of this bug. ***
Comment 5 CVS Commits 2020-03-23 23:59:45 UTC
The master branch has been updated by Jeff Law <law@gcc.gnu.org>:


commit r10-7346-g75fb811dfaa29d60a897924b0d1629577b90eee7
Author: Jeff Law <law@redhat.com>
Date:   Mon Mar 23 17:55:20 2020 -0600

    Verify the code used for the optimized comparison is valid for the comparison's mode.
            PR rtl-optimization/90275
            PR target/94238
            PR target/94144
            * simplify-rtx.c (comparison_code_valid_for_mode): New function.
            (simplify_logical_relational_operation): Use it.
            PR target/94144
            PR target/94238
            * gcc.c-torture/compile/pr94144.c: New test.
            * gcc.c-torture/compile/pr94238.c: New test.
Comment 6 Jeffrey A. Law 2020-03-24 00:00:26 UTC
Resolved by patch on the trunk.