McStas and compiler bugs

Investigations indicate that McStas-generated code is very sensitive to bugs in the optimizers of commonly used compilers! This note documents preliminary results on comparing a McStas simulation when compiled with different compilers and optimizations.

Compilers were:

• gcc-i386 - GCC 2.7.2.3 for i386-linux
• egcs - EGCS 2.90.29 for i386-linux
• pgcc - PGCC 2.90.29 for i386-linux
• cc - HP92453-01 A.10.32.03 HP ANSI C Compiler HPPA/HPUX B.10.20
• gcc-hp - GCC 2.7.2 for HPPA/HPUX B.10.20

The i386-linux compiled simulations were run on a 300MHz Pentium II. The HPPA compiled simulations were run on the fys-hp-2 workstation.

The results:

Compiler	flags	I [p/N]	MON [p/N]	neutrons/sec
gcc-i386	none	3.1e-07/393	17823.7/72244	23k
gcc-i386	-O6	3.3e-07/428	17868.9/72309	25k
egcs	-O6	3.6e-07/420	17938.8/72394	30k
pgcc	none	2.8e-07/387	17945.1/72334	21k
pgcc	-O6	4.8e-07/718	8420.16/34245	36k
cc	-Aa	2.9e-07/379	17937.6/72292	17k
cc	-Aa +Oall -Wl,-a,archive	1.7e-07/203	17949.7/72351	40k
gcc-hp	none	3.0e-07/373	17877.8/71976	19k
gcc-hp	-O6	3.2e-07/368	17726.2/71825	22k

As can be seen from the table, PGCC and HP CC with full optimization gives very large speedups on the simulations, but unfortunately also produce incorrect results!

Update

A new release of PGCC (1.1a) seems to correct the bugs and produce correct code. Using the '-mstack-align-double' switch in addition to -O6 it also achieves slightly better performance, about 31k neutrons/sec.

---

Kristian Nielsen <kristian.nielsen@risoe.dk> / Sep 3, 1998.

---

Last Modified: Tuesday, 22-Jun-2004 14:04:10 CEST