Code size issues on FP-emulation on libgcc compared to LLVM's compiler_rt

[Zinovy Nis <zinovy dot nis at gmail dot com>]

Hi.


I'm building libgcc for a "iamcu" target (Pentium-like but with
soft-fp emulation, the only x86 with SoftFP I know) with
--enable-target-optspace.


It works properly but I noticed that code the size for many arithmetic
functions is much more larger than for soft-fp emulation provided by
LLVM's compiler_rt library.

For example addsf3 (addition of 2 floats) is 0.5K in compiler_rt and 1K in gcc!


I double-checked compiler options and they are correct: -Os.


Is it because of

   a) compiler_rt doesn't perform some checks (FP exception for example) or

   b) compiler_rt is incorrect in some corner cases or

   c) inefficient implementation in libgcc?


Did anyone face with this problem?

How to reproduce:

First you need to build gcc for iamcu:

  INSTALL=<your install dir>

 TARGET=i586-intel-elfiamcu

  ./configure --host=x86_64-unknown-linux-gnu --enable-target-optspace
--target=${TARGET} --prefix=${INSTALL} --with-demangler-in-ld
--enable-languages=c MAKEINFO=missing --disable-multilib
--disable-libitm --disable-libatomic --disable-libssp
--disable-libquadmath --with-newlib && make -j $(nproc) && make
install


libgcc

  Let's get size of two float addition operation:

  nm -S -t dec --defined-only
$(INSTALL)/lib/gcc/i586-intel-elfiamcu/4.9.3/libgcc.a | grep _addsf3

I get:

  00000000 00001039 T __addsf3

  So the size is ~1K

compiler_rt

  svn co http://llvm.org/svn/llvm-project/compiler-rt/trunk compiler-rt

  cd compiler_rt/lib/builtins

  Compile: $(INSTALL)/bin/i586-intel-elfiamcu-gcc -Os  addsf3.c -c

  Check function size:  nm -S -t dec addsf3.o

I get:

  00000000 00000522 T __addsf3

  So the size is ~0.5K.

Why so large difference in libgcc and compiler_rt for the same functionality?

 Thanks.

-- Zinovy Nis.