A comparison between lcc-win32, gcc, and Intel's icl compiler

Comparison details:

The best compiler for the intel processors is surely icl, the compiler of Intel Corporation. It benefits of the inmense resources and know-how of Intel Corp.

GCC is a compiler that shares with lcc-win32 its low price, but has a long development history, and a team of many people dedicated to it.

lcc-win32 has been the result of the work of just three people. There is no budget, no institution whatsoever funding this work. In light of this conditions, the performances of lcc-win32 are quite acceptable, in my opinion.

Two different source sets were used for this comparison:

• The source code of egcs 1.1b itself. Around 12 Megabytes of C source, this code represents a real complex application.
• The symbolic calculator GP of Grenoble University in France. This code was used to measure the pure CPU integer speed in a real context of complex calculations. This program doesn't do any disk access other than reading a very small source file written in the calculator's internal language. The sources represent approx 1.5 megabytes of C code.
  
• The version of gcc used was GNU C++ version egcs-2.91.57
• The version of the icl compiler was version 2.4 P97176
• The version of lcc was 2.4, compiled Nov-3 98
• The machine used was a pentium MMX (intel) at 200MHZ with 64MB RAM.

Compiling egcs 1.1b

Icl compiler was given -Os (optimize for size), gcc was given -O, and lcc was given -O. If I changed the gcc optimizations for -O9, execution speed actually increased..., so I kept just plain -O, that seems to generate better code.

Compilation time with optimizations
icl	650 sec
gcc	472 sec
lcc	213 sec

Compilation time with no optimizations
icl	236 sec
gcc	400 sec
lcc	198.4 sec

To measure the speed of the resulting executable, the same source file was used: a large file of 270.999 bytes C source, that was given to the three generated compilers. The exact command line is:

xxx.exe -O9 win32.c

All executables were started twice, and only the second time was significant to avoid the influence of the disk cache.

Execution time (optimized)
icl	33.7 sec
gcc	35.6 sec
lcc	39.8 sec

Execution time (no optimizations)
icl	43.5 sec
gcc	46.6 sec
lcc	47.8 sec

Code size (optimized)
icl	2 056 192
gcc	1 728 000
lcc	1 831 968

Code size (no optimizations)
icl	2 946 048
gcc	2 679 808
lcc	2 213 920

Compiling the symbolic calculator 'gp'

Icl compiler was given -Ox. gcc was given -O, and lcc was given -O.

Compilation time with optimizations
icl	192.3 sec
gcc	175 sec
lcc	43.5 sec

Compilation time with no optimizations
icl	73 sec
gcc	152.2 sec
lcc	36.9 sec

Code size with optimizations
icl	1 033 728
gcc	786 432
lcc	785 440

Code size with no optimizations
icl	1 218 560
gcc	1 248 256
lcc	969 248

Execution time with optimizations
icl	35 sec
gcc	40.3 sec
lcc	44 sec

Execution time with no optimizations
icl	48.8 sec
gcc	60.7 sec
lcc	51.3 sec

A floating point comparison

As measure I used a matrix multiplication program. Since the source is very small (37K) there is no point in measuring the compilation speed. I used the option -0x for intel, -02 for gcc, and -O for lcc. The results are as follows:

icl	3.9 sec
gcc	4.8 sec
lcc	5.6 sec

It must be noted that Intel's compiler comes with a very high performance floating point math library. lcc (as gcc) uses the standard library of windows

Another important point is that the results of icl and lcc were exactly identical. The results of gcc differed after the 13th digit.