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] reload1.c: Detect and utilize implicit sets inmove2add. (take 2)

Hi Jim,

> > Attached is a patch to detect implicit sets in move2add and treat such
> > sets identically to explicit sets.  This patch is a revised version of
> The concept looks OK, however I don't think it is quite right.
> Branches can be either
> 	(if_then_else (ne cond) (pc) (label))
> or
> 	(if_then_else (ne cond) (label) (pc))
> These mean different things.  The first means we fall through if the 
> condition is NE.  The second means we branch to the label if the 
> condition is NE.  You need to check for this distinction in your code. 
> There is also the issue that
>    (if_then_else (eq cond) (pc) (label))
> means the same as
>    (if_then_else (ne cond) (label) (pc))
> So you want to accept EQ if pc is operand 1, and NE if pc is operand 2.

I think fis_get_condition handles this problem.  Specifically,
fis_get_condition() returns the condition for taking a branch, not the
one for falling through.  Here is a comment in fis_get_condition().

  /* If this branches to JUMP_LABEL when the condition is false,
     reverse the condition.  */

> Before the move2add_note_store call, you are duplicating some of the 
> tests found in move2add_note_store, specifically, the SCALAR_INT_MODE_P 
> and HARD_REGNO_NREGS checks.  It isn't clear why you are doing this.

A good point.  I will test a version of the patch without those two
lines.  (As a second thought, it might be a good idea to avoid a call
to gen_rtx_SET, which allocates memory, by doing some these simple
checks first.)

Kazu Hirata

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