Message ID | alpine.DEB.2.02.1405110032330.12138@stedding.saclay.inria.fr |
---|---|
State | New |
Headers | show |
On Sat, May 10, 2014 at 3:53 PM, Marc Glisse <marc.glisse@inria.fr> wrote: > Hello, > > in my recent phiopt patch enhancing value_replacement to optimize > x!=0?x+y:y, I forgot to check that there is no other PHI (not sure how I > managed to miss that since I copy-pasted the line just below the test). > > If there are other phi nodes (with different arguments for those 2 > branches), it would be possible to replace the phi argument and stop there > (as value_replacement does for its other transformation). However, I am > chosing to punt. The cost analysis would be different, and I wrote the > transformation assuming that this single-phi test was already done higher in > the function. I think we should have some good cost analysis because for this testcase, we should be able to get only one conditional move but right now with punting we don't. > > Bootstrap+testsuite on x86_64-linux-gnu. > > 2014-05-12 Marc Glisse <marc.glisse@inria.fr> > > PR tree-optimization/61140 > gcc/ > * tree-ssa-phiopt.c (value_replacement): Punt on multiple phis. > gcc/testsuite/ > * gcc.dg/tree-ssa/pr61140.c: New file. > > -- > Marc Glisse > Index: gcc/testsuite/gcc.dg/tree-ssa/pr61140.c > =================================================================== > --- gcc/testsuite/gcc.dg/tree-ssa/pr61140.c (revision 0) > +++ gcc/testsuite/gcc.dg/tree-ssa/pr61140.c (working copy) > @@ -0,0 +1,18 @@ > +/* { dg-do run } */ > +/* { dg-options "-O2" } */ > + > +int a[1] = { 1 }, b = 1, c; > + > +int > +main () > +{ > + for (; c < 1; c++) > + if (a[0]) > + { > + a[0] &= 1; > + b = 0; > + } > + if (b) > + __builtin_abort (); > + return 0; > +} > Index: gcc/tree-ssa-phiopt.c > =================================================================== > --- gcc/tree-ssa-phiopt.c (revision 210301) > +++ gcc/tree-ssa-phiopt.c (working copy) > @@ -842,20 +842,24 @@ value_replacement (basic_block cond_bb, > /* Now optimize (x != 0) ? x + y : y to just y. > The following condition is too restrictive, there can easily be > another > stmt in middle_bb, for instance a CONVERT_EXPR for the second > argument. */ > gimple assign = last_and_only_stmt (middle_bb); > if (!assign || gimple_code (assign) != GIMPLE_ASSIGN > || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS > || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0)) > && !POINTER_TYPE_P (TREE_TYPE (arg0)))) > return 0; > > + /* Only transform if it removes the condition. */ > + if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, > e1)) > + return 0; > + > /* Size-wise, this is always profitable. */ > if (optimize_bb_for_speed_p (cond_bb) > /* The special case is useless if it has a low probability. */ > && profile_status_for_fn (cfun) != PROFILE_ABSENT > && EDGE_PRED (middle_bb, 0)->probability < PROB_EVEN > /* If assign is cheap, there is no point avoiding it. */ > && estimate_num_insns (assign, &eni_time_weights) > >= 3 * estimate_num_insns (cond, &eni_time_weights)) > return 0; > >
On Sat, 10 May 2014, Andrew Pinski wrote: > On Sat, May 10, 2014 at 3:53 PM, Marc Glisse <marc.glisse@inria.fr> wrote: >> Hello, >> >> in my recent phiopt patch enhancing value_replacement to optimize >> x!=0?x+y:y, I forgot to check that there is no other PHI (not sure how I >> managed to miss that since I copy-pasted the line just below the test). >> >> If there are other phi nodes (with different arguments for those 2 >> branches), it would be possible to replace the phi argument and stop there >> (as value_replacement does for its other transformation). However, I am >> chosing to punt. The cost analysis would be different, and I wrote the >> transformation assuming that this single-phi test was already done higher in >> the function. > > I think we should have some good cost analysis because for this > testcase, we should be able to get only one conditional move but right > now with punting we don't. That's true. But note that the transformation is already very limited (gives up if there is a second statement in the middle bb, even a simple cast), so I would like to first quickly get the wrong-code regression out of the way, and we can make improvements afterwards (though we can of course start discussing them now). It seems like if there is only 1 extra non-singleton phi (in addition to the one we are transforming) and the target supports conditional move for this type and the direct branch has proba < 50%, with the other restrictions already in place, we could go ahead. How does that sound? Not too specialized? If there are many phis, conditional moves are out, the branch will stay, and unless the edge to the operation has a very high proba, it doesn't seem like a good idea to pull the operation out of the branch.
On 05/11/14 04:03, Marc Glisse wrote: > On Sat, 10 May 2014, Andrew Pinski wrote: > >> On Sat, May 10, 2014 at 3:53 PM, Marc Glisse <marc.glisse@inria.fr> >> wrote: >>> Hello, >>> >>> in my recent phiopt patch enhancing value_replacement to optimize >>> x!=0?x+y:y, I forgot to check that there is no other PHI (not sure how I >>> managed to miss that since I copy-pasted the line just below the test). >>> >>> If there are other phi nodes (with different arguments for those 2 >>> branches), it would be possible to replace the phi argument and stop >>> there >>> (as value_replacement does for its other transformation). However, I am >>> chosing to punt. The cost analysis would be different, and I wrote the >>> transformation assuming that this single-phi test was already done >>> higher in >>> the function. >> >> I think we should have some good cost analysis because for this >> testcase, we should be able to get only one conditional move but right >> now with punting we don't. > > That's true. But note that the transformation is already very limited > (gives up if there is a second statement in the middle bb, even a simple > cast), so I would like to first quickly get the wrong-code regression > out of the way, and we can make improvements afterwards (though we can > of course start discussing them now). > > It seems like if there is only 1 extra non-singleton phi (in addition to > the one we are transforming) and the target supports conditional move > for this type and the direct branch has proba < 50%, with the other > restrictions already in place, we could go ahead. How does that sound? > Not too specialized? If there are many phis, conditional moves are out, > the branch will stay, and unless the edge to the operation has a very > high proba, it doesn't seem like a good idea to pull the operation out > of the branch. Your call based on what you see from a codegen standpoint. Having been burned before with these kinds of transformations, I tend to be a bit conservative :-) If you decide to keep things as-is, the patch is fine. If you want to extend to handle the additional case, please repost for review. Thanks, jeff
Index: gcc/testsuite/gcc.dg/tree-ssa/pr61140.c =================================================================== --- gcc/testsuite/gcc.dg/tree-ssa/pr61140.c (revision 0) +++ gcc/testsuite/gcc.dg/tree-ssa/pr61140.c (working copy) @@ -0,0 +1,18 @@ +/* { dg-do run } */ +/* { dg-options "-O2" } */ + +int a[1] = { 1 }, b = 1, c; + +int +main () +{ + for (; c < 1; c++) + if (a[0]) + { + a[0] &= 1; + b = 0; + } + if (b) + __builtin_abort (); + return 0; +} Index: gcc/tree-ssa-phiopt.c =================================================================== --- gcc/tree-ssa-phiopt.c (revision 210301) +++ gcc/tree-ssa-phiopt.c (working copy) @@ -842,20 +842,24 @@ value_replacement (basic_block cond_bb, /* Now optimize (x != 0) ? x + y : y to just y. The following condition is too restrictive, there can easily be another stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */ gimple assign = last_and_only_stmt (middle_bb); if (!assign || gimple_code (assign) != GIMPLE_ASSIGN || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0)) && !POINTER_TYPE_P (TREE_TYPE (arg0)))) return 0; + /* Only transform if it removes the condition. */ + if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1)) + return 0; + /* Size-wise, this is always profitable. */ if (optimize_bb_for_speed_p (cond_bb) /* The special case is useless if it has a low probability. */ && profile_status_for_fn (cfun) != PROFILE_ABSENT && EDGE_PRED (middle_bb, 0)->probability < PROB_EVEN /* If assign is cheap, there is no point avoiding it. */ && estimate_num_insns (assign, &eni_time_weights) >= 3 * estimate_num_insns (cond, &eni_time_weights)) return 0;