Hi Uros and Richard, I was rewriting the Alpha sched_find_first_bit implementation for the Linux Kernel, and in the process I think I've come across a gcc bug.
I rewrote the function using cmov instructions, and wrote a small
program to test its correctness and performance. I wrote the function
initially as an external .S file, and once I was reasonably sure it
was correct, converted it to C function with inline assembly.
Compiling both produce the exact same output, as shown.
<rewritten>:
ldq t0,0(a0)
clr t2
ldq t1,8(a0)
cmoveq t0,0x40,t2
cmoveq t0,t1,t0
cttz t0,t3
addq t3,t2,v0
ret
In my test program, I found that when I executed the rewritten
implementation _before_ the reference implementation that it produced
bogus results. This only happens when using the C/inline asm function.
When compiled with the external .S file, the results are correct.
Attached is a tar.gz with my test code. Compile the test program with
`gcc -O -mcpu=... find.c rewritten.S test.c -o test` with optional
-D__REWRITTEN_INLINE and -D__REWRITTEN_FIRST. At -Os, -O2, or -O3 and
-D__REWRITTEN_INLINE and -D__REWRITTEN_FIRST the program will produce
incorrect results and assert(). At -O0 or -O1 or without one or both
of the -D flags, it will produce correct results. I've tested with
gcc-4.3.4 and gcc-4.4.2.
Thanks. Let me know what I can do to help further.
Matt Turner
sched_find_first_bit.tar.gz
Description: GNU Zip compressed data
