diff options
| author | Michael Matz <matz@suse.de> | 2009-10-20 16:07:21 +0000 |
|---|---|---|
| committer | Michael Matz <matz@gcc.gnu.org> | 2009-10-20 16:07:21 +0000 |
| commit | 1578e91079162f7ff008520487623968fc0db687 (patch) | |
| tree | 8373744cb0ba30186429688bf97d57291d6ff5b3 /gcc/loop-invariant.c | |
| parent | 5a3a8eb12ea8fa102cfbdbd2429fc119a41f5904 (diff) | |
| download | gcc-1578e91079162f7ff008520487623968fc0db687.zip gcc-1578e91079162f7ff008520487623968fc0db687.tar.gz gcc-1578e91079162f7ff008520487623968fc0db687.tar.bz2 | |
* loop-invariant.c (create_new_invariant): Use different magic number.
From-SVN: r153022
Diffstat (limited to 'gcc/loop-invariant.c')
| -rw-r--r-- | gcc/loop-invariant.c | 11 |
1 files changed, 10 insertions, 1 deletions
diff --git a/gcc/loop-invariant.c b/gcc/loop-invariant.c index 8cbfdb3..901ce51 100644 --- a/gcc/loop-invariant.c +++ b/gcc/loop-invariant.c @@ -705,8 +705,17 @@ create_new_invariant (struct def *def, rtx insn, bitmap depends_on, if (def) { inv->cost = rtx_cost (set, SET, speed); + /* ??? Try to determine cheapness of address computation. Unfortunately + the address cost is only a relative measure, we can't really compare + it with any absolute number, but only with other address costs. + But here we don't have any other addresses, so compare with a magic + number anyway. It has to be large enough to not regress PR33928 + (by avoiding to move reg+8,reg+16,reg+24 invariants), but small + enough to not regress 410.bwaves either (by still moving reg+reg + invariants). + See http://gcc.gnu.org/ml/gcc-patches/2009-10/msg01210.html . */ inv->cheap_address = address_cost (SET_SRC (set), word_mode, - speed) < COSTS_N_INSNS (1); + speed) < 3; } else { |