https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61749
Bug ID: 61749
Summary: arm_neon.h "_lane" and "_n" intrinsics can cause ICEs
Product: gcc
Version: 4.8.4
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: jgreenhalgh at gcc dot gnu.org
Given the following, all release branch and trunk AArch64 compilers will ICE
trying to expand the intrinsic.
#include "arm_neon.h"
int32x4_t foo (int32x4_t a, int16x4_t b, int16x4_t c, int d)
{
return vqdmlal_lane_s16 (a, b, c, 0);
}
As so:
/work/gcc-dev/build-aarch64-none-elf/install/lib/gcc/aarch64-none-elf/4.10.0/include/arm_neon.h:
In function 'foo':
/work/gcc-dev/build-aarch64-none-elf/install/lib/gcc/aarch64-none-elf/4.10.0/include/arm_neon.h:19294:10:
internal compiler error: in aarch64_simd_lane_bounds, at
config/aarch64/aarch64.c:7715
return __builtin_aarch64_sqdmlal_lanev4hi (__a, __b, __c, __d);
^
0xc608d0 aarch64_simd_lane_bounds(rtx_def*, long, long)
/work/gcc-dev/src/gcc/gcc/config/aarch64/aarch64.c:7715
0xcb0221 gen_aarch64_sqdmlal_lanev4hi(rtx_def*, rtx_def*, rtx_def*, rtx_def*,
rtx_def*)
/work/gcc-dev/src/gcc/gcc/config/aarch64/aarch64-simd.md:3015
0xc65b7f insn_gen_fn::operator()(rtx_def*, rtx_def*, rtx_def*, rtx_def*,
rtx_def*) const
/work/gcc-dev/src/gcc/gcc/recog.h:311
0xc65b7f aarch64_simd_expand_args
/work/gcc-dev/src/gcc/gcc/config/aarch64/aarch64-builtins.c:888
0xc66318 aarch64_simd_expand_builtin(int, tree_node*, rtx_def*)
/work/gcc-dev/src/gcc/gcc/config/aarch64/aarch64-builtins.c:990
0xc66968 aarch64_expand_builtin(tree_node*, rtx_def*, rtx_def*, machine_mode,
int)
/work/gcc-dev/src/gcc/gcc/config/aarch64/aarch64-builtins.c:1071
0x6062a2 expand_builtin(tree_node*, rtx_def*, rtx_def*, machine_mode, int)
/work/gcc-dev/src/gcc/gcc/builtins.c:5756
0x6f897b expand_expr_real_1(tree_node*, rtx_def*, machine_mode,
expand_modifier, rtx_def**, bool)
/work/gcc-dev/src/gcc/gcc/expr.c:10344
0x6f9e7c expand_expr_real(tree_node*, rtx_def*, machine_mode, expand_modifier,
rtx_def**, bool)
/work/gcc-dev/src/gcc/gcc/expr.c:7922
0x70078f store_expr(tree_node*, rtx_def*, int, bool)
/work/gcc-dev/src/gcc/gcc/expr.c:5312
0x7092b8 expand_assignment(tree_node*, tree_node*, bool)
/work/gcc-dev/src/gcc/gcc/expr.c:5098
0x626493 expand_call_stmt
/work/gcc-dev/src/gcc/gcc/cfgexpand.c:2305
0x626493 expand_gimple_stmt_1
/work/gcc-dev/src/gcc/gcc/cfgexpand.c:3202
0x626493 expand_gimple_stmt
/work/gcc-dev/src/gcc/gcc/cfgexpand.c:3354
0x628d75 expand_gimple_basic_block
/work/gcc-dev/src/gcc/gcc/cfgexpand.c:5192
0x62ad53 execute
/work/gcc-dev/src/gcc/gcc/cfgexpand.c:5799
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <http://gcc.gnu.org/bugs.html> for instructions.
The bug is in the treatment in aarch64-simd-builtins.def of these lane
intrinsics as "QUADOP"s. This allows their final parameter to *maybe* be an
immediate. But subsequent code on the expand path assumes it to *always* be an
immediate.
The fix would to special case the handling of these intrinsics and enforce that
the final parameter is *always* an immediate.
For 4.10 and 4.9 that means adding a new set of types to "qualifiers". For 4.8,
it means adding a new aarch64_simd_itype and handling it appropriately.
