Optimization of GAS

Optimization of GAS

[Stas Kiselev]

Hello,

I want to make GAS faster. I profiled it and found out that lots of the 
time takes on working with hash. What do you thing about changing hash 
mechanism into hash_map from stl. Will it work faster ?

Thank you, Stanislav

Re: Optimization of GAS

[Ian Lance Taylor]

Stas Kiselev <stas_kiselev@bk.ru> writes:

> I want to make GAS faster. I profiled it and found out that lots of the 
> time takes on working with hash. What do you thing about changing hash 
> mechanism into hash_map from stl. Will it work faster ?

STL is C++ code.  gas is written in C.  So using hash_map directly
would be rather difficult.

I expect that it is possible to speed up the hash code in gas.  It
would also be interesting to see if it can be called less often.  A
common usage is simply looking up instruction mnemonics, in which the
set of instructions is fixed at the start of assembly, and really, if
we do appropriate tests, it is fixed when the assembler is built.  Can
we take advantage of that, perhaps by doing perfect hashing?

Ian

Re: Optimization of GAS

[Ben Elliston]

> I want to make GAS faster. I profiled it and found out that lots of the 
> time takes on working with hash. What do you thing about changing hash 
> mechanism into hash_map from stl. Will it work faster ?

GAS really isn't a bottleneck on modern machines that have plenty of buffer cache.

I changed the GCC c-torture/compile.exp tests last year so that they just compiled 
rather than including the assembly step (ie. using -S instead of -c), thinking it 
would speed up those tests.  Over that large testsuite, the saving was less than 0.1%.

If you'd like to make something faster, work on the linker!

Ben