https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118362
Bug ID: 118362
Summary: [15 Regression] ICE in require, at machmode.h:313
since r15-2002
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: jakub at gcc dot gnu.org
Target Milestone: ---
int a, b, c[18];
void
foo (void)
{
for (int i = 0; i < 8; i++)
if (b)
{
c[i * 2 + 1] = a;
c[i * 2 + 2] = 200;
}
else
{
c[i * 2 + 1] = 1000000;
c[i * 2 + 2] = 200;
}
}
ICEs on s390x-linux with -O2 -march=z14 starting with
r15-2002-g61715e9340ab8941d40d62158fe437e9dbe3e068:
/usr/src/gcc-test/objz3/gcc/cc1.r15-6702 -quiet -nostdinc -march=z14 -O2
ccPWXpC5.i
during RTL pass: combine
ccPWXpC5.i: In function ‘foo’:
ccPWXpC5.i:17:1: internal compiler error: in require, at machmode.h:323
17 | }
| ^
0x2b96cd7 internal_error(char const*, ...)
../../gcc/diagnostic-global-context.cc:517
0x2b666db fancy_abort(char const*, int, char const*)
../../gcc/diagnostic.cc:1722
0xf4d545 opt_mode<machine_mode>::require() const
../../gcc/machmode.h:323
0x1c1d389 s390_constant_via_vgbm_p(rtx_def*, unsigned int*)
../../gcc/config/s390/s390.cc:2831
0x1c06ba2 satisfies_constraint_jyy
../../gcc/config/s390/constraints.md:449
0x1c21810 s390_legitimate_constant_p
../../gcc/config/s390/s390.cc:4448
0x15ff76f general_operand(rtx_def*, machine_mode)
../../gcc/recog.cc:1529
0x27a8782 combine_simplify_rtx
../../gcc/combine.cc:5822
0x27a7edf subst
../../gcc/combine.cc:5630
0x27a7c12 subst
../../gcc/combine.cc:5557
0x27a7c12 subst
../../gcc/combine.cc:5557
0x27a0b25 try_combine
../../gcc/combine.cc:3379
0x279b0eb combine_instructions
../../gcc/combine.cc:1357
0x27c3446 rest_of_handle_combine
../../gcc/combine.cc:15158
0x27c3502 execute
../../gcc/combine.cc:15202
Please submit a full bug report, with preprocessed source (by using
-freport-bug).
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.
This is on
2830 /* Handle floats by bitcasting them to ints. */
2831 op = gen_lowpart (related_int_vector_mode (GET_MODE
(op)).require (),
2832 op);
where op is:
(const_vector:V8SI [
(const_int 1000000 [0xf4240]) repeated x4
(const_int 200 [0xc8]) repeated x4
])
The comment suggests we shouldn't reach it, but
the previous if just handles cases with GET_MODE_INNER (GET_MODE (op)) being
T[IF]mode.
Maybe just s390_constant_via_vgbm_p doesn't expect any 32-byte vectors?
Though, combine just tries to combine and simplify random stuff and attempts to
match, so the predicates better should return false instead of ICEing.
The ICE is when trying to simplify
(vec_concat:V8SI (if_then_else:V4SI (eq (not:V4SI (reg:V4SI 84 [
mask_patt_30.7_35 ]))
(const_vector:V4SI [
(const_int 0 [0]) repeated x4
]))
(mem/u/c:V4SI (symbol_ref/u:DI ("*.LC0") [flags 0x2]) [0 S16 A64])
(reg:V4SI 85 [ _36 ]))
(const_vector:V4SI [
(const_int 200 [0xc8]) repeated x4
]))
true_rtx from that becomes
(const_vector:V8SI [
(const_int 1000000 [0xf4240]) repeated x4
(const_int 200 [0xc8]) repeated x4
])
and false_rtx from that becomes
(vec_concat:V8SI (reg:V4SI 85 [ _36 ])
(const_vector:V4SI [
(const_int 200 [0xc8]) repeated x4
]))
and combine_simplify_rtx tests if both are general_operand.
--- gcc/config/s390/s390.cc.jj 2025-01-02 11:23:55.822968123 +0100
+++ gcc/config/s390/s390.cc 2025-01-08 19:30:33.763998865 +0100
@@ -2820,16 +2820,18 @@ s390_constant_via_vgbm_p (rtx op, unsign
if (GET_CODE (op) == CONST_VECTOR)
{
+ machine_mode relmode;
if (GET_MODE_INNER (GET_MODE (op)) == TImode
|| GET_MODE_INNER (GET_MODE (op)) == TFmode)
/* For the sake of simplicity, bitcast 16-byte one-element vector
into two-element vector so we don't have to special case them in
the following. */
op = gen_lowpart (V2DImode, op);
- else
+ else if (related_int_vector_mode (GET_MODE (op)).exists (&relmode))
/* Handle floats by bitcasting them to ints. */
- op = gen_lowpart (related_int_vector_mode (GET_MODE (op)).require (),
- op);
+ op = gen_lowpart (relmode, op);
+ else
+ return false;
}
else if ((CONST_WIDE_INT_P (op) && CONST_WIDE_INT_NUNITS (op) == 2)
|| (CONST_DOUBLE_P (op) && GET_MODE (op) == TFmode))
fixes the ICE, but dunno whether that is the best fix.
