[PATCH] Improve DOM's optimization of control statements
Until now DOM has had to be very conservative with handling control statements with known conditions. This as been an unfortunate side effect of the interaction between removing edges and recycling names via the SSA_NAME manager. Essentially DOM would have to leave control statements alone. So you'd see stuff like if (0 == 0) left around by DOM. The jump threader would thankfully come along and optimize that as a jump thread. But that's terribly inefficient, not to mention it creates unnecessary churn in the CFG and SSA_NAMEs. By optimizing that directly in DOM, including removing whatever edges are not executable, we no longer have to rely on jump threading to handle that case. Less churn in the CFG & SSA_NAMEs. There's also some chance for secondary optimizations with fewer edges left in the CFG for DOM to consider. Unfortunately, the churn caused by jump threading made it excessively difficult to analyze before/after dumps. Sadly, you can have the same code, but if the SSA_NAMEs have changed, that impacts coalescing as we leave SSA. Churn in the CFG changes labels/jumps, often without changing the actual structure, etc. I did some tests with valgrind to evaluate branching behaviour before/after effects on the resulting code and those effects were tiny, in the I doubt you could measure them range. That was expected since what we're really doing here is just capturing the optimization earlier. I had a couple more tests, but they were lost in a bit of idiocy. The test included is the one I had a second copy of lying around. Bootstrapped and regression tested on x86_64-linux-gnu. Installed on the trunk. Jeff * tree-ssa-dom.c (optimize_stmt): Collapse control flow statements with constant conditions. * tree-ssa-threadupdate.c (remove_jump_threads_starting_at): New. (remove_ctrl_stmt_and_useless_edges): No longer static. * tree-ssa-threadupdate.h (remove_jump_threads_starting_at): Prototype. (remove_ctrl_stmt_and_useless_edges): Likewise. * gcc.dg/tree-ssa/ssa-dom-branch-1.c: New test. diff --git a/gcc/tree-ssa-dom.c b/gcc/tree-ssa-dom.c index 2c51e36..a8b7038 100644 --- a/gcc/tree-ssa-dom.c +++ b/gcc/tree-ssa-dom.c @@ -1820,31 +1820,8 @@ optimize_stmt (basic_block bb, gimple_stmt_iterator si, if (is_gimple_assign (stmt)) record_equivalences_from_stmt (stmt, may_optimize_p, avail_exprs_stack); - /* If STMT is a COND_EXPR and it was modified, then we may know - where it goes. If that is the case, then mark the CFG as altered. - - This will cause us to later call remove_unreachable_blocks and - cleanup_tree_cfg when it is safe to do so. It is not safe to - clean things up here since removal of edges and such can trigger - the removal of PHI nodes, which in turn can release SSA_NAMEs to - the manager. - - That's all fine and good, except that once SSA_NAMEs are released - to the manager, we must not call create_ssa_name until all references - to released SSA_NAMEs have been eliminated. - - All references to the deleted SSA_NAMEs can not be eliminated until - we remove unreachable blocks. - - We can not remove unreachable blocks until after we have completed - any queued jump threading. - - We can not complete any queued jump threads until we have taken - appropriate variables out of SSA form. Taking variables out of - SSA form can call create_ssa_name and thus we lose. - - Ultimately I suspect we're going to need to change the interface - into the SSA_NAME manager. */ + /* If STMT is a COND_EXPR or SWITCH_EXPR and it was modified, then we may + know where it goes. */ if (gimple_modified_p (stmt) || modified_p) { tree val = NULL; @@ -1858,8 +1835,27 @@ optimize_stmt (basic_block bb, gimple_stmt_iterator si, else if (gswitch *swtch_stmt = dyn_cast (stmt)) val = gimple_switch_index (swtch_stmt); - if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val)) - cfg_altered = true; + if (val && TREE_CODE (val) == INTEGER_CST) + { + edge taken_edge = find_taken_edge (bb, val); + if (taken_edge) + { + /* Delete threads that start at BB. */ + remove_jump_threads_starting_at (bb); + + /* Now clean up the control statement at the end of +BB and remove unexecutable edges. */ + remove_ctrl_stmt_and_useless_edges (bb, taken_edge->dest); + + /* Fixup the flags on the single remaining edge. */ + taken_edge->flags + &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE | EDGE_ABNORMAL); + taken_edge->flags |= EDGE_FALLTHRU; + + /* Further simplifications may be possible. */ + cfg_altered = true; + } + } /* If we simplified a statement in such a way as to be shown that it cannot trap
Re: [PATCH] Improve DOM's optimization of control statements
Hi Jeff, Your patch causes an ICE regression. The test case is " gcc.c-torture/compile/pr27087.c", I observed it on aarch64-none-elf target when compiling the test case with '-Os' flag. A quick check shows, the cfg has been changed, but the loop information is not updated. Thus the information about the number of basic block in a loop is not reliable. Could you please have a look? Regards, Renlin On 30/09/15 21:28, Jeff Law wrote: Until now DOM has had to be very conservative with handling control statements with known conditions. This as been an unfortunate side effect of the interaction between removing edges and recycling names via the SSA_NAME manager. Essentially DOM would have to leave control statements alone. So you'd see stuff like if (0 == 0) left around by DOM. The jump threader would thankfully come along and optimize that as a jump thread. But that's terribly inefficient, not to mention it creates unnecessary churn in the CFG and SSA_NAMEs. By optimizing that directly in DOM, including removing whatever edges are not executable, we no longer have to rely on jump threading to handle that case. Less churn in the CFG & SSA_NAMEs. There's also some chance for secondary optimizations with fewer edges left in the CFG for DOM to consider. Unfortunately, the churn caused by jump threading made it excessively difficult to analyze before/after dumps. Sadly, you can have the same code, but if the SSA_NAMEs have changed, that impacts coalescing as we leave SSA. Churn in the CFG changes labels/jumps, often without changing the actual structure, etc. I did some tests with valgrind to evaluate branching behaviour before/after effects on the resulting code and those effects were tiny, in the I doubt you could measure them range. That was expected since what we're really doing here is just capturing the optimization earlier. I had a couple more tests, but they were lost in a bit of idiocy. The test included is the one I had a second copy of lying around. Bootstrapped and regression tested on x86_64-linux-gnu. Installed on the trunk. Jeff
Re: [PATCH] Improve DOM's optimization of control statements
On 10/02/2015 05:15 AM, Renlin Li wrote: Hi Jeff, Your patch causes an ICE regression. The test case is " gcc.c-torture/compile/pr27087.c", I observed it on aarch64-none-elf target when compiling the test case with '-Os' flag. A quick check shows, the cfg has been changed, but the loop information is not updated. Thus the information about the number of basic block in a loop is not reliable. Could you please have a look? Yup. Will do. Thanks for letting me know. jeff
Re: [PATCH] Improve DOM's optimization of control statements
On 10/02/2015 05:15 AM, Renlin Li wrote: Hi Jeff, Your patch causes an ICE regression. The test case is " gcc.c-torture/compile/pr27087.c", I observed it on aarch64-none-elf target when compiling the test case with '-Os' flag. A quick check shows, the cfg has been changed, but the loop information is not updated. Thus the information about the number of basic block in a loop is not reliable. Could you please have a look? Appears to be pretty simple. If we collapse a conditional inside a loop, then we need to set the loop state as needing fixups. In this specific case we have a conditional where one path eventually leads back to the loop latch block, the other path exits the loop. We statically determine the conditional will always take us to the loop exit. That has the side effect of making the block with the collapsed conditional no longer part of the loop, it's actually part of the exit path. That changes the number of nodes in the loop, what edge(s) are the exit path(s) and possibly other stuff. I'm running a fix through testing now. Thanks, jeff
Re: [PATCH] Improve DOM's optimization of control statements
On 10/02/2015 05:15 AM, Renlin Li wrote: Hi Jeff, Your patch causes an ICE regression. The test case is " gcc.c-torture/compile/pr27087.c", I observed it on aarch64-none-elf target when compiling the test case with '-Os' flag. A quick check shows, the cfg has been changed, but the loop information is not updated. Thus the information about the number of basic block in a loop is not reliable. Could you please have a look? As I mentioned, when we collapse a conditional inside a loop, we may change the # of nodes in a loop which edges are exit edges and possibly other stuff. So we need to mark loops as needing fixups. Verified this fixes the aarch64-elf regression and did a bootstrap & regression test on x86_64-linux-gnu. Installed on the trunk. jeff commit 992d281b2d1ba53a49198db44fee92a505e16f5d Author: Jeff Law Date: Fri Oct 2 15:22:04 2015 -0400 Re: [PATCH] Improve DOM's optimization of control statements * tree-ssa-dom.c (optimize_stmt): Note when loop structures need fixups. diff --git a/gcc/ChangeLog b/gcc/ChangeLog index 3f7561a..e541df3 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,8 @@ +2015-10-02 Jeff Law + + * tree-ssa-dom.c (optimize_stmt): Note when loop structures need + fixups. + 2015-10-02 Uros Bizjak * system.h (ROUND_UP): New macro definition. diff --git a/gcc/tree-ssa-dom.c b/gcc/tree-ssa-dom.c index a8b7038..d940816 100644 --- a/gcc/tree-ssa-dom.c +++ b/gcc/tree-ssa-dom.c @@ -1843,6 +1843,12 @@ optimize_stmt (basic_block bb, gimple_stmt_iterator si, /* Delete threads that start at BB. */ remove_jump_threads_starting_at (bb); + /* If BB is in a loop, then removing an outgoing edge from BB +may cause BB to move outside the loop, changes in the +loop exit edges, etc. So note that loops need fixing. */ + if (bb_loop_depth (bb) > 0) + loops_state_set (LOOPS_NEED_FIXUP); + /* Now clean up the control statement at the end of BB and remove unexecutable edges. */ remove_ctrl_stmt_and_useless_edges (bb, taken_edge->dest);
Re: [PATCH] Improve DOM's optimization of control statements
On Fri, Oct 2, 2015 at 9:30 PM, Jeff Law wrote: > On 10/02/2015 05:15 AM, Renlin Li wrote: >> >> Hi Jeff, >> >> Your patch causes an ICE regression. >> The test case is " gcc.c-torture/compile/pr27087.c", I observed it on >> aarch64-none-elf target when compiling the test case with '-Os' flag. >> >> A quick check shows, the cfg has been changed, but the loop information >> is not updated. Thus the information about the number of basic block in >> a loop is not reliable. >> >> Could you please have a look? > > As I mentioned, when we collapse a conditional inside a loop, we may change > the # of nodes in a loop which edges are exit edges and possibly other > stuff. So we need to mark loops as needing fixups. > > Verified this fixes the aarch64-elf regression and did a bootstrap & > regression test on x86_64-linux-gnu. > > Installed on the trunk. > > jeff > > commit 992d281b2d1ba53a49198db44fee92a505e16f5d > Author: Jeff Law > Date: Fri Oct 2 15:22:04 2015 -0400 > > Re: [PATCH] Improve DOM's optimization of control statements > > * tree-ssa-dom.c (optimize_stmt): Note when loop structures need > fixups. > > diff --git a/gcc/ChangeLog b/gcc/ChangeLog > index 3f7561a..e541df3 100644 > --- a/gcc/ChangeLog > +++ b/gcc/ChangeLog > @@ -1,3 +1,8 @@ > +2015-10-02 Jeff Law > + > + * tree-ssa-dom.c (optimize_stmt): Note when loop structures need > + fixups. > + > 2015-10-02 Uros Bizjak > > * system.h (ROUND_UP): New macro definition. > diff --git a/gcc/tree-ssa-dom.c b/gcc/tree-ssa-dom.c > index a8b7038..d940816 100644 > --- a/gcc/tree-ssa-dom.c > +++ b/gcc/tree-ssa-dom.c > @@ -1843,6 +1843,12 @@ optimize_stmt (basic_block bb, gimple_stmt_iterator > si, > /* Delete threads that start at BB. */ > remove_jump_threads_starting_at (bb); > > + /* If BB is in a loop, then removing an outgoing edge from BB > +may cause BB to move outside the loop, changes in the > +loop exit edges, etc. So note that loops need fixing. */ > + if (bb_loop_depth (bb) > 0) > + loops_state_set (LOOPS_NEED_FIXUP); > + I would rather do this in remove_ctrl_stmt_and_useless_edges and only if taken_edge is a loop exit. loop fixup is a pretty big hammer which we should avoid at all cost. So please try to be more specific on the cases you invoke it. Thanks, Richard. > /* Now clean up the control statement at the end of > BB and remove unexecutable edges. */ > remove_ctrl_stmt_and_useless_edges (bb, taken_edge->dest); >
Re: [PATCH] Improve DOM's optimization of control statements
On 10/05/2015 03:02 AM, Richard Biener wrote: + /* If BB is in a loop, then removing an outgoing edge from BB +may cause BB to move outside the loop, changes in the +loop exit edges, etc. So note that loops need fixing. */ + if (bb_loop_depth (bb) > 0) + loops_state_set (LOOPS_NEED_FIXUP); + I would rather do this in remove_ctrl_stmt_and_useless_edges and only if taken_edge is a loop exit. loop fixup is a pretty big hammer which we should avoid at all cost. remove_ctrl_stmt_and_useless_edges is used internally when duplicating blocks for jump threading, so we'd need to audit for that case as well. So please try to be more specific on the cases you invoke it. I'd pondered this and was still working through whether or not triggering only when one of the outgoing edges was "interesting" from a loop structure standpoint. I hadn't tested that yet and wanted to get the regression resolved, hence the large hammer. jeff
Re: [PATCH] Improve DOM's optimization of control statements
On 10/05/2015 03:02 AM, Richard Biener wrote: On Fri, Oct 2, 2015 at 9:30 PM, Jeff Law wrote: On 10/02/2015 05:15 AM, Renlin Li wrote: Hi Jeff, Your patch causes an ICE regression. The test case is " gcc.c-torture/compile/pr27087.c", I observed it on aarch64-none-elf target when compiling the test case with '-Os' flag. A quick check shows, the cfg has been changed, but the loop information is not updated. Thus the information about the number of basic block in a loop is not reliable. Could you please have a look? As I mentioned, when we collapse a conditional inside a loop, we may change the # of nodes in a loop which edges are exit edges and possibly other stuff. So we need to mark loops as needing fixups. Verified this fixes the aarch64-elf regression and did a bootstrap & regression test on x86_64-linux-gnu. Installed on the trunk. jeff commit 992d281b2d1ba53a49198db44fee92a505e16f5d Author: Jeff Law Date: Fri Oct 2 15:22:04 2015 -0400 Re: [PATCH] Improve DOM's optimization of control statements * tree-ssa-dom.c (optimize_stmt): Note when loop structures need fixups. diff --git a/gcc/ChangeLog b/gcc/ChangeLog index 3f7561a..e541df3 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,8 @@ +2015-10-02 Jeff Law + + * tree-ssa-dom.c (optimize_stmt): Note when loop structures need + fixups. + 2015-10-02 Uros Bizjak * system.h (ROUND_UP): New macro definition. diff --git a/gcc/tree-ssa-dom.c b/gcc/tree-ssa-dom.c index a8b7038..d940816 100644 --- a/gcc/tree-ssa-dom.c +++ b/gcc/tree-ssa-dom.c @@ -1843,6 +1843,12 @@ optimize_stmt (basic_block bb, gimple_stmt_iterator si, /* Delete threads that start at BB. */ remove_jump_threads_starting_at (bb); + /* If BB is in a loop, then removing an outgoing edge from BB +may cause BB to move outside the loop, changes in the +loop exit edges, etc. So note that loops need fixing. */ + if (bb_loop_depth (bb) > 0) + loops_state_set (LOOPS_NEED_FIXUP); + I would rather do this in remove_ctrl_stmt_and_useless_edges and only if taken_edge is a loop exit. loop fixup is a pretty big hammer which we should avoid at all cost. So please try to be more specific on the cases you invoke it. What's probably the most interesting is we don't actually have EDGE_LOOP_EXIT set in DOM. So we can't use that, but that also implies it doesn't need updating. Thankfully, there's an alternate test we can use instead. And after more ponderings, I'm pretty sure the only case that's of concern right now is nodes moving out of the loop, which only happens when we have a BB with a loop exit edge and we delete the *other* edges. We've got simple latches & preheaders and I don't think we destroy that property. Both properties also simplify the things we need look for. So I've minimized use of the hammer to just the case where blocks are going to be moving out of the loop. Bootstrapped and regression tested on x86_64-unknown-linux-gnu and verified the ARM test which originally spurred this change continues to work. Installed on the trunk. Jeff * tree-ssa-dom.c (optimize_stmt): Don't set LOOPS_NEED_FIXUP here. * tree-ssa-threadupdate.c (remove_ctrl_stmt_and_useless_edges): Do it here instead. Tighten test to avoid setting LOOPS_NEED_FIXUP unnecessarily. diff --git a/gcc/tree-ssa-dom.c b/gcc/tree-ssa-dom.c index 941087d..38cceff 100644 --- a/gcc/tree-ssa-dom.c +++ b/gcc/tree-ssa-dom.c @@ -1848,12 +1848,6 @@ optimize_stmt (basic_block bb, gimple_stmt_iterator si, FOR_EACH_EDGE (e, ei, bb->succs) remove_jump_threads_including (e); - /* If BB is in a loop, then removing an outgoing edge from BB -may cause BB to move outside the loop, changes in the -loop exit edges, etc. So note that loops need fixing. */ - if (bb_loop_depth (bb) > 0) - loops_state_set (LOOPS_NEED_FIXUP); - /* Now clean up the control statement at the end of BB and remove unexecutable edges. */ remove_ctrl_stmt_and_useless_edges (bb, taken_edge->dest); diff --git a/gcc/tree-ssa-threadupdate.c b/gcc/tree-ssa-threadupdate.c index 26b199b..e426c1d 100644 --- a/gcc/tree-ssa-threadupdate.c +++ b/gcc/tree-ssa-threadupdate.c @@ -300,6 +300,17 @@ remove_ctrl_stmt_and_useless_edges (basic_block bb, basic_block dest_bb) else ei_next (&ei); } + + /* If the remaining edge is a loop exit, there must have + a removed edge that was not a loop exit. + + In that case BB and possibly other blocks were previously + in the loop, but are now outside the loop. Thus, we need + to update the loop structures. */ + if (single_succ_p (
Re: [PATCH] Improve DOM's optimization of control statements
On powerpc: FAIL: gcc.c-torture/compile/pr52073.c -O2 (internal compiler error) /daten/gcc/gcc-20151006/gcc/testsuite/gcc.c-torture/compile/pr52073.c:6:1: internal compiler error: in duplicate_thread_path, at tree-ssa-threadupdate.c:2446. 0x10874ab7 duplicate_thread_path. ../../gcc/tree-ssa-threadupdate.c:2445. 0x10874ab7 thread_through_all_blocks(bool). ../../gcc/tree-ssa-threadupdate.c:2632. 0x107b0ea7 execute. ../../gcc/tree-ssa-dom.c:622. Andreas. -- Andreas Schwab, sch...@linux-m68k.org GPG Key fingerprint = 58CA 54C7 6D53 942B 1756 01D3 44D5 214B 8276 4ED5 "And now for something completely different."
Re: [PATCH] Improve DOM's optimization of control statements
On 10/07/2015 04:03 PM, Andreas Schwab wrote: On powerpc: FAIL: gcc.c-torture/compile/pr52073.c -O2 (internal compiler error) /daten/gcc/gcc-20151006/gcc/testsuite/gcc.c-torture/compile/pr52073.c:6:1: internal compiler error: in duplicate_thread_path, at tree-ssa-threadupdate.c:2446. 0x10874ab7 duplicate_thread_path. ../../gcc/tree-ssa-threadupdate.c:2445. 0x10874ab7 thread_through_all_blocks(bool). ../../gcc/tree-ssa-threadupdate.c:2632. 0x107b0ea7 execute. ../../gcc/tree-ssa-dom.c:622. Update your tree. This should have been fixed by: 2015-10-06 Jeff Law PR tree-optimization/67816 Jeff