On 05/05/2015 12:42 AM, Aditya K wrote:
I was able to successfully bootstrap gcc by using clang as the stage 1
compiler. I configured gcc using the following arguments.
../configure --disable-multilib --enable-bootstrap --enable-languages=c,c++
CC=/work/llvm/install-release/bin/clang
CXX=/work/llvm/install-release/bin/clang++
And the bootstrap was successful. One useful thing I got to see was clang
warnings. Clang produced several warnings (> 1000 unique ones). I have attached
two files with this email.
I also recently interested in building GCC by LLVM. Although I was
interesting in compilation speed of LLVM.
So I built GCC-5 sources with LLVM-3.6 and GCC-5 using
--disable-bootstrap --enable-languages=c to minimize time spent by built
GCC to compile libraries (or GCC-5 sources again for bootstrap). Here
are the results of make -j1 on E5-2997v3 (default -O2 was used by the
both compilers):
GCC-5 LLVM-3.6
real 20m44s 20m31s
user 15m35s 15m56s
Another interesting thing I found on SPEC2006 (only C/C++ benchmarks
were used for FP) for -O1 (the machine is 4.2Ghz i7-4790K).
x86-64 SPEC rates (more is better):
INT: LLVM GCC
-O1: 34.9 39.8 14%
FP: LLVM GCC
-O1: 35.7 48.5 36%
x86-64 Compilation time in seconds (cpu time):
INT: LLVM GCC
-O1: 196.46 153.66 -22%
FP: LLVM GCC
-O1: 110.22 86.83 -21%
So GCC-5 in -O1 mode generates 20% better code and spent 15-30% less
time on compilation than LLVM-3.6.
LLVM is popular for JITing but I believe GCC has much bigger potential
for this and recommend JIT developers to try it.
I think it helps to see alternative point of view on LLVM as "a faster
compiler than GCC".
