This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Fix multiply-add regressions after expand-from-SSA

Richard Guenther writes:
> On Thu, 30 Apr 2009, Adam Nemet wrote:
> > This fixes the MIPS madd* and msub* regressions from:
> > 
> >
> > 
> > We now need to look at the TERed expressions for the operands to find the
> > underlying multiplications and casts.
> > 
> > OK if regtest/bootstrap passes on mips64octeon-linux?
> I think we should start to _not_ re-build trees here, but instead of

I am not rebuilding any tree here, just as operands are looked at I look at
the TER expression in available.  Maybe the name of the helper is wrong.  I
don't mean subsitute as in walking the tree and substituting subtrees but as
in return the original expr or subsitute it with something "better".

> > -       && TREE_CODE (TREE_OPERAND (exp, 0)) == MULT_EXPR)
> > +       && (subexp0 = substitute_ter_expr (TREE_OPERAND (exp, 0)))
> > +       && TREE_CODE (subexp0) == MULT_EXPR)
> do
>           && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME
>           && is_gimple_assign (SSA_NAME_DEF_STMT (TREE_OPERAND (exp, 0)))
>           && gimple_assign_rhs_code (SSA_NAME_DEF_STMT (TREE_OPERAND (exp, 
> 0)) == MULT_EXPR)
> etc.

Isn't that exactly what substitute_ter_expr does (maybe the name is
misleading).  I just only do the SSA_NAME_DEF_STMT if TER decided so.

> Is there a particular reason why combine doesn't catch whatever
> the expander does here (I didn't look too closely yet).

It's hard to say because at the moment it seems that even widening
multiplication is broken.  I.e. we of course need to look at the definition of
the operand to find the cast.

What's has been the design in expand-from-SSA to handle these cases?


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]