Hi,
expanding AArch64 AdvSIMD builtins, aarch64_simd_expand_builtin puts
return type and arguments types in args[SIMD_MAX_BUILTIN_ARGS] array and
indicates the last argument with SIMD_ARG_STOP. When expanding
4-argument builtin with return value, aarch64_simd_expand_builtin stores
SIMD_ARG_STOP to args[5], that overflows args since
SIMD_MAX_BUILTIN_ARGS == 5. This causes ICE for aarch64 cross-compiler
on i686 host:
$
/home/m.ostapenko/jazz64/vd49_a5x/toolchain/aarch64-tizen/libexec/gcc/aarch64-tizen-linux-gnueabi/5.0.0/cc1
-quiet ref_vqdmlsl_lane.c -quiet -dumpbase ref_vqdmlsl_lane.c
-march=armv8-a+simd -mlittle-endian -mabi=lp64 -auxbase-strip
ref_vqdmlsl_lane.gccarm.o -g -Wall -Wno-unused-variable
-Wno-unused-function -Wno-error=unused-local-typedefs -fshort-wchar
-freport-bug
// In file included from ref_vqdmlal_lane.c:27:0,
// from ref_vqdmlsl_lane.c:29:
//
/home/m.ostapenko/jazz64/vd49_a5x/toolchain/aarch64-tizen/lib/gcc/aarch64-tizen-linux-gnueabi/5.0.0/include/arm_neon.h:
In function 'exec_vqdmlsl_lane':
//
/home/m.ostapenko/jazz64/vd49_a5x/toolchain/aarch64-tizen/lib/gcc/aarch64-tizen-linux-gnueabi/5.0.0/include/arm_neon.h:19165:10:
internal compiler error: tree check: accessed operand 8 of call_expr
with 7 operands in aarch64_simd_expand_args, at
config/aarch64/aarch64-builtins.c:915
// return __builtin_aarch64_sqdmlsl_lanev4hi (__a, __b, __c, __d);
// ^
// 0x89b4948 tree_operand_check_failed(int, tree_node const*, char
const*, int, char const*)
//
>----/home/m.ostapenko/jazz64/vd49_a5x/toolchain/build/sources/gcc_1/gcc/tree.c:9507
// 0x8a3c9f9 tree_operand_check
//
>----/home/m.ostapenko/jazz64/vd49_a5x/toolchain/build/sources/gcc_1/gcc/tree.h:3346
// 0x8a3c9f9 aarch64_simd_expand_args
//
>----/home/m.ostapenko/jazz64/vd49_a5x/toolchain/build/sources/gcc_1/gcc/config/aarch64/aarch64-builtins.c:915
// 0x8a3c9f9 aarch64_simd_expand_builtin(int, tree_node*, rtx_def*)
This trivial patch fixes the issue.
No regressions found for aarch64-linux-gnueabi (on x86_64 and i686
hosts), ok to commit?
-Maxim
gcc/ChangeLog:
2015-04-01 Max Ostapenko <m.ostape...@partner.samsung.com>
PR target/65624
* config/aarch64/aarch64-builtins.c (aarch64_simd_expand_builtin): Increase
args array size by one to avoid buffer overflow.
gcc/testsuite/ChangeLog:
2015-04-01 Max Ostapenko <m.ostape...@partner.samsung.com>
PR target/65624
* gcc.target/aarch64/pr65624.c: New test.
diff --git a/gcc/config/aarch64/aarch64-builtins.c b/gcc/config/aarch64/aarch64-builtins.c
index a9b3305..8a9fff8 100644
--- a/gcc/config/aarch64/aarch64-builtins.c
+++ b/gcc/config/aarch64/aarch64-builtins.c
@@ -1025,7 +1025,7 @@ aarch64_simd_expand_builtin (int fcode, tree exp, rtx target)
aarch64_simd_builtin_datum *d =
&aarch64_simd_builtin_data[fcode - AARCH64_SIMD_PATTERN_START];
enum insn_code icode = d->code;
- builtin_simd_arg args[SIMD_MAX_BUILTIN_ARGS];
+ builtin_simd_arg args[SIMD_MAX_BUILTIN_ARGS + 1];
int num_args = insn_data[d->code].n_operands;
int is_void = 0;
int k;
diff --git a/gcc/testsuite/gcc.target/aarch64/pr65624.c b/gcc/testsuite/gcc.target/aarch64/pr65624.c
new file mode 100644
index 0000000..8eb3786
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/pr65624.c
@@ -0,0 +1,15 @@
+/* { dg-do compile } */
+
+#include <arm_neon.h>
+
+int32x4_t
+foo (void)
+{
+ int32x4_t vector_int32x4;
+ int16x4_t vector3_int16x4;
+ int16x4_t vector4_int16x4;
+ static int32_t buffer_int32x4[32];
+
+ vector_int32x4 = vld1q_s32(buffer_int32x4);
+ return vqdmlsl_lane_s16(vector_int32x4, vector3_int16x4, vector4_int16x4, 0);
+}
