https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106763
Bug ID: 106763
Summary: Armv8.2 vmov.f16 instruction sometimes causes SIGILL
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: georgepee at gmail dot com
Target Milestone: ---
First noticed when going from gcc-10 to gcc-12 and a complex function being
compiled with -ftree-vectorize started to emit a vmov.f16 instruction and
sometimes SIGILL on it.
It appears that this commit is when it started to emit that instruction:
https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=6390c5047adb75960f86d56582e6322aaa4d9281
Made a small program that also SIGILLs intermittently on vmov.f16 instruction
Compiled with -g -mcpu=cortex-a55:
#include <vector>
#include <algorithm>
#include <arm_neon.h>
#include <stdio.h>
#include <sys/auxv.h>
float16_t rand_func()
{
return float16_t(double(rand()) / double(RAND_MAX));
}
int main()
{
srand( 11 );
// just to show that vmov.f16 doesn't always SIGILL
{
std::vector<float16_t> floats_small(10);
printf("float_small pre\n");
std::generate( floats_small.begin() , floats_small.end() , rand_func );
printf("float_small post\n");
}
// sometimes SIGILLS
{
std::vector<float16_t> floats_large(500000);
printf("float_large pre\n");
std::generate( floats_large.begin() , floats_large.end() , rand_func );
printf("float_large post\n");
}
return 0;
}
Execution:
$ ./a.out
float_small pre
float_small post
float_large pre
float_large post
$ ./a.out
float_small pre
float_small post
float_large pre
Illegal instruction
GDB:
Program terminated with signal SIGILL, Illegal instruction.
#0 0x00010778 in rand_func () at vmov_f16_test.c:9
9 return float16_t(double(rand()) / double(RAND_MAX));
(gdb) disassemble 0x00010770,+32
Dump of assembler code from 0x10770 to 0x10790:
0x00010770 <rand_func()+24>: vdiv.f64 d16, d17, d18
0x00010774 <rand_func()+28>: vcvtb.f16.f64 s15, d16
=> 0x00010778 <rand_func()+32>: vmov.f16 r3, s15
0x0001077c <rand_func()+36>: mov r0, r3
0x00010780 <rand_func()+40>: pop {r11, pc}
0x00010784 <rand_func()+44>: nop {0}
0x00010788 <rand_func()+48>: ; <UNDEFINED> instruction:
0xffc00000
0x0001078c <rand_func()+52>: ldrshmi pc, [pc, #255] ; 0x10893 <main()+258>
; <UNPREDICTABLE>
I have been able to run the compiled program on multiple Cortex-A55 devices.
Initially, I though that since FP16 is optional on armv8.2-a that it was truly
an illegal instruction, but if that is the case, then why would it only fail
intermittently?
