strstr() a lot slower with ASAN (Linux glibc x86-64)

[fuhsnn]

After enabling ASAN in the compiler project I work on, I noticed a strstr() call in the tokenization loop is responsible for significant slowdown in ASAN builds.

I believe this is worth reporting since it's a simple change that caused a huge diff, which should make an interesting case for profiling.

In the worst case, ASAN builds are about 170x slower on Skylake, 150x on Zen4, after replacing the strstr() with a loop it's back to normal 2.8x / 2.5x.
The change is fuhsnn/slimcc@eb66a46.

A reproduce script testing ASAN on-off and strstr() with-without, should work on any recent-ish glibc docker:

set -e

git clone https://github.com/fuhsnn/c_files
git clone https://github.com/fuhsnn/slimcc

cd slimcc
git checkout eb66a4619f168af7ce5c34ecc8cc37c82988d51d

cc *.c -o slimcc
echo "Asan 0, strstr 0"
time ./slimcc ../c_files/gcc.c -E -o/dev/null

cc *.c -o slimcc -fsanitize=address
echo "Asan 1 strstr 0"
time ./slimcc ../c_files/gcc.c -E -o/dev/null

git checkout HEAD~1

cc *.c -o slimcc
echo "Asan 0, strstr 1"
time ./slimcc ../c_files/gcc.c -E -o/dev/null

cc *.c -o slimcc -fsanitize=address
echo "Asan 1, strstr 1"
time ./slimcc ../c_files/gcc.c -E -o/dev/null