This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] RTL PRE speedup
- From: Richard Biener <rguenther at suse dot de>
- To: gcc-patches at gcc dot gnu dot org
- Date: Wed, 12 Feb 2014 15:42:17 +0100 (CET)
- Subject: [PATCH] RTL PRE speedup
- Authentication-results: sourceware.org; auth=none
No magic bullet for the insn-recog.c testcase I'm looking at but
quite obvious and should save half of the canon_true_dependence calls
for blocks that end up non-transparent.
Now one of the main weakness of compute_transp is that we call it
for all expressions but for the MEM case it's certainly not
interesting to know whether it is transparent in _all_ blocks
but just a subset (dependent on the actual transform we do - either
where it's available or where its anticipated).
Committed as obvious.
Richard.
2014-02-12 Richard Biener <rguenther@suse.de>
* gcse.c (compute_transp): break from loop over canon_modify_mem_list
when we found a dependence.
Index: gcc/gcse.c
===================================================================
*** gcc/gcse.c (revision 207718)
--- gcc/gcse.c (working copy)
*************** compute_transp (const_rtx x, int indx, s
*** 1735,1741 ****
if (canon_true_dependence (dest, GET_MODE (dest),
dest_addr, x, x_addr))
! bitmap_clear_bit (bmap[bb_index], indx);
}
}
}
--- 1735,1744 ----
if (canon_true_dependence (dest, GET_MODE (dest),
dest_addr, x, x_addr))
! {
! bitmap_clear_bit (bmap[bb_index], indx);
! break;
! }
}
}
}