The following patch adds FRE after vectorization which is needed
for IVOPTs to remove redundant PHI nodes (well, I'm testing a
patch for FRE that will do it already there).
The patch also makes FRE preserve loop-closed SSA form and thus
make it suitable for use in the loop pipeline.
With the placement in the vectorizer sub-pass FRE will effectively
be enabled by -O3 only (well, or if one requests loop vectorization).
I've considered placing it after complete_unroll instead but that
would enable it at -O1 already. I have no strong opinion on the
exact placement, but it should help all passes between vectorizing
and ivopts for vectorized loops.
Yeah, it adds yet another pass and thus I don't like it very much.
But it for example improves code generated for
gfortran.dg/vect/fast-math-pr37021.f90
from
.L14:
movupd (%r11), %xmm3
addl $1, %ecx
addq %rax, %r11
movupd (%r8), %xmm0
addq %rax, %r8
unpckhpd %xmm3, %xmm3
movupd (%rdi), %xmm2
unpcklpd %xmm0, %xmm0
addq %rsi, %rdi
movupd (%rbx), %xmm1
mulpd %xmm3, %xmm2
addq %rsi, %rbx
cmpl %ecx, %ebp
palignr $8, %xmm1, %xmm1
mulpd %xmm1, %xmm0
movapd %xmm2, %xmm1
addpd %xmm0, %xmm1
subpd %xmm2, %xmm0
shufpd $2, %xmm0, %xmm1
addpd %xmm1, %xmm4
jne .L14
to
.L14:
movupd (%r8), %xmm0
addl $1, %ecx
addq %rax, %r8
movapd %xmm0, %xmm2
movupd (%rdi), %xmm1
addq %rsi, %rdi
cmpl %ecx, %r11d
unpckhpd %xmm0, %xmm2
unpcklpd %xmm0, %xmm0
mulpd %xmm1, %xmm2
palignr $8, %xmm1, %xmm1
mulpd %xmm1, %xmm0
movapd %xmm2, %xmm1
addpd %xmm0, %xmm1
subpd %xmm2, %xmm0
shufpd $2, %xmm0, %xmm1
addpd %xmm1, %xmm3
jne .L14
(yeah, the vectorizer happily generates redundant loads and one IV
for each such load)
Any other suggestions on pass placement? I can of course key
that FRE run on -O3 explicitely. Not sure if we at this point
want to start playing fancy games like setting a property
when a pass (likely) generated redundancies that are worth
fixing up and then key FRE on that one (it gets harder and
less predictable what transforms are run on code).
Bootstrap / regtest running on x86_64-unknown-linux-gnu. With
other placements I'd expect quite some testsuite fallout
eventually.
Thoughts?
Thanks,
Richard.
2015-06-10 Richard Biener <rguent...@suse.de>
* passes.def (pass_vectorize): Add pass_fre.
* tree-ssa-pre.c (eliminate_dom_walker::before_dom_children):
Preserve loop-closed SSA form.
Index: gcc/passes.def
===================================================================
*** gcc/passes.def (revision 224324)
--- gcc/passes.def (working copy)
*************** along with GCC; see the file COPYING3.
*** 252,257 ****
--- 252,258 ----
Please do not add any other passes in between. */
NEXT_PASS (pass_vectorize);
PUSH_INSERT_PASSES_WITHIN (pass_vectorize)
+ NEXT_PASS (pass_fre);
NEXT_PASS (pass_dce);
POP_INSERT_PASSES ()
NEXT_PASS (pass_predcom);
Index: gcc/tree-ssa-pre.c
===================================================================
*** gcc/tree-ssa-pre.c (revision 224324)
--- gcc/tree-ssa-pre.c (working copy)
*************** eliminate_dom_walker::before_dom_childre
*** 4013,4018 ****
--- 4013,4028 ----
tailmerging. Eventually we can reduce its reliance on SCCVN now
that we fully copy/constant-propagate (most) things. */
+ /* Compute whether this block has loop-closed PHI nodes we need
+ to preserve. */
+ bool lc_phi = false;
+ edge e;
+ if (loops_state_satisfies_p (LOOP_CLOSED_SSA)
+ && single_pred_p (b)
+ && (e = single_pred_edge (b))
+ && loop_exit_edge_p (e->src->loop_father, e))
+ lc_phi = true;
+
for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (gsi);)
{
gphi *phi = gsi.phi ();
*************** eliminate_dom_walker::before_dom_childre
*** 4026,4032 ****
tree sprime = eliminate_avail (res);
if (sprime
! && sprime != res)
{
if (dump_file && (dump_flags & TDF_DETAILS))
{
--- 4036,4043 ----
tree sprime = eliminate_avail (res);
if (sprime
! && sprime != res
! && !lc_phi)
{
if (dump_file && (dump_flags & TDF_DETAILS))
{
*************** eliminate_dom_walker::before_dom_childre
*** 4466,4472 ****
/* Replace destination PHI arguments. */
edge_iterator ei;
- edge e;
FOR_EACH_EDGE (e, ei, b->succs)
{
for (gphi_iterator gsi = gsi_start_phis (e->dest);
--- 4477,4482 ----
