The original gcc I used was the minix gcc (1.37.1) by Alan W Black (and
somebody..  forgotten who?), which did some very ugly things in order to
handle floating point.  I used that to compile gcc-1.40 for minix, with
patches by Bruce Evans to clean up the floating point handling and some
of my own patches (I fooled around with gcc when learning about it, and
added a "-mstring-insns" switch which allowed gcc-1.40 to use the 386
string instructions for structure copying etc).

The minix-386 (thanks again, Bruce - without the 386 patches for minix I
would never have gotten anywhere) I was running by that time was able to
execute normal gcc binaries directly (the only major diff I did to minix
- and awb later made a better version of it), which is also the standard
linux binary format.  So the same gcc binary running under minix could
make both minix and linux binaries: the only thing that differed was the
startup routine (crt0.o) and the standard library.  My first binaries
used modified minix library routines, and the first shell I used under
linux (pre-0.01) was the minix bourne shell recompiled with those
library routines.

The binaries I later released used a library that was heavily based on
Earl Chew's (free) estdio package (thanks), with some minor routines by
me and some from various other sources (I think the minix termcap
routines were free etc).  Back then (0.01-0.11), the biggest problem
with gcc were the libraries - estdio worked (even though it had it's
quirks) but many other routines were either incomplete or missing.
Porting many programs meant I had to write new library routines or find
them from some other source (like the early and very buggy glibc.a).

Also, I wasn't able to compile binaries under linux until version 0.10
or so - all the major first binaries (notably bash) were crosscompiled
from minix.  The reason was the bad buffer-cache bugs in early linux
versions which made running gcc (which easily fills the buffer-cache)
impossible.  0.03 was able to run a gcc binary, but recompiling gcc
itself under linux led to weird crashes due to the buffer cache
corruption.

By 0.11-0.12, hlu had showed interest in maintaining gcc-2.x (I checked
out early snapshots of the alpha sources), which meant I had to
implement the math-emulation in the kernel, so that gcc wouldn't need
the patches for integer floating-point.  By that time glibc.a also
started to be partly useable, so the linux libraries started using that
more and more (with some bad bugs as a result in the early versions).
The rest is history.