Bug 1624813

Created attachment 1480508 [details]
Reproducer

Description of problem:
FFTW code can SIGILL on Allwinner A10. I am not ARM expert, so feel free to correct me if I am wrong, but it seems the CNTVCT register for precision time measurement is not defined in ARMv7-A architecture profile [1]. It's defined in ARMv8-A architecture profile [2]. We are compiling for ARMv7-A (-march=armv7-a), so such instruction shouldn't be used. 

The problematic code is in:
kernel/cycle.h:518,
the getticks function is directly using assembly instruction 'mrrc' for CNTVCT register access, e.g.: 
mrrc    15, 1, sl, r9, cr14
which codes into 0xEC59AF1E opcode which causes the SIGILL.

It seems there is compile time check for CNTVCT support in configure.ac:727:
AC_MSG_CHECKING([armv7a has 64 bits readable CNTVCT])

But it seems to be wrong approach, because the builder may support it, but the target CPU needn't. The check should be performed runtime (like e.g openssl do) or the code should be compiled without CNTVCT support.

I tried to patch-out the CNTVCT timer code and it seems it works on my A10 without problem.

[1] https://www.cs.utexas.edu/~simon/378/resources/ARMv7-AR_TRM.pdf
[2] https://static.docs.arm.com/ddi0487/ca/DDI0487C_a_armv8_arm.pdf

Version-Release number of selected component (if applicable):
fftw-3.3.5-11.fc28.armv7hl

How reproducible:
Always on A10 and maybe others

Steps to Reproduce:
1. gcc -o test ./test.c -lfftw3f
2. ./test
3.

Actual results:
Illegal instruction (core dumped)

Expected results:
Working code

Additional info:
This causes other packages using fftw like gr-osmosdr, gqrx, ... to SIGILL on A10 and maybe others.

It seems there is an typo in the fftw source code. Sometimes the CNTVCT is referred correctly, sometimes as CNTCVT, like e.g. ARMV7A_HAS_CNTCVT, which is incorrect.