<div dir="ltr">I have a strong suspicion this caused:<div><a href="http://lab.llvm.org:8011/builders/clang-ppc64be-linux-multistage/builds/4273">http://lab.llvm.org:8011/builders/clang-ppc64be-linux-multistage/builds/4273</a><br></div><div><br></div><div>And several other stage2 failures in the bots right now. (stack trace looks like MBP)</div><br><div class="gmail_quote"><div dir="ltr">On Sun, Jun 25, 2017 at 10:27 PM Serguei Katkov via llvm-commits <<a href="mailto:llvm-commits@lists.llvm.org">llvm-commits@lists.llvm.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Author: skatkov<br>
Date: Sun Jun 25 22:27:27 2017<br>
New Revision: 306272<br>
<br>
URL: <a href="http://llvm.org/viewvc/llvm-project?rev=306272&view=rev" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project?rev=306272&view=rev</a><br>
Log:<br>
[MBP] do not rotate loop if it creates extra branch<br>
<br>
This is a last fix for the corner case of PR32214. Actually this is not really corner case in general.<br>
<br>
We should not do a loop rotation if we create an additional branch due to it.<br>
Consider the case where we have a loop chain H, M, B, C , where<br>
H is header with viable fallthrough from pre-header and exit from the loop<br>
M - some middle block<br>
B - backedge to Header but with exit from the loop also.<br>
C - some cold block of the loop.<br>
<br>
Let's H is determined as a best exit. If we do a loop rotation M, B, C, H we can introduce the extra branch.<br>
Let's compute the change in number of branches:<br>
+1 branch from pre-header to header<br>
-1 branch from header to exit<br>
+1 branch from header to middle block if there is such<br>
-1 branch from cold bock to header if there is one<br>
<br>
So if C is not a predecessor of H then we introduce extra branch.<br>
<br>
This change actually prohibits rotation of the loop if both true<br>
1) Best Exit has next element in chain as successor.<br>
2) Last element in chain is not a predecessor of first element of chain.<br>
<br>
Reviewers: iteratee, xur<br>
Reviewed By: iteratee<br>
Subscribers: llvm-commits<br>
Differential Revision: <a href="https://reviews.llvm.org/D34271" rel="noreferrer" target="_blank">https://reviews.llvm.org/D34271</a><br>
<br>
Modified:<br>
llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp<br>
llvm/trunk/test/CodeGen/X86/block-placement.ll<br>
llvm/trunk/test/CodeGen/X86/code_placement_cold_loop_blocks.ll<br>
<br>
Modified: llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp?rev=306272&r1=306271&r2=306272&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp?rev=306272&r1=306271&r2=306272&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp (original)<br>
+++ llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp Sun Jun 25 22:27:27 2017<br>
@@ -1944,6 +1944,35 @@ void MachineBlockPlacement::rotateLoop(B<br>
if (ExitIt == LoopChain.end())<br>
return;<br>
<br>
+ // Rotating a loop exit to the bottom when there is a fallthrough to top<br>
+ // trades the entry fallthrough for an exit fallthrough.<br>
+ // If there is no bottom->top edge, but the chosen exit block does have<br>
+ // a fallthrough, we break that fallthrough for nothing in return.<br>
+<br>
+ // Let's consider an example. We have a built chain of basic blocks<br>
+ // B1, B2, ..., Bn, where Bk is a ExitingBB - chosen exit block.<br>
+ // By doing a rotation we get<br>
+ // Bk+1, ..., Bn, B1, ..., Bk<br>
+ // Break of fallthrough to B1 is compensated by a fallthrough from Bk.<br>
+ // If we had a fallthrough Bk -> Bk+1 it is broken now.<br>
+ // It might be compensated by fallthrough Bn -> B1.<br>
+ // So we have a condition to avoid creation of extra branch by loop rotation.<br>
+ // All below must be true to avoid loop rotation:<br>
+ // If there is a fallthrough to top (B1)<br>
+ // There was fallthrough from chosen exit block (Bk) to next one (Bk+1)<br>
+ // There is no fallthrough from bottom (Bn) to top (B1).<br>
+ // Please note that there is no exit fallthrough from Bn because we checked it<br>
+ // above.<br>
+ if (ViableTopFallthrough) {<br>
+ MachineBasicBlock *NextBlockInChain = *std::next(ExitIt);<br>
+ MachineBasicBlock *Bottom = *std::prev(LoopChain.end());<br>
+ if (NextBlockInChain->isPredecessor(ExitingBB))<br>
+ if (!Top->isPredecessor(Bottom))<br>
+ return;<br>
+ }<br>
+<br>
+ DEBUG(dbgs() << "Rotating loop to put exit " << getBlockName(ExitingBB)<br>
+ << " at bottom\n");<br>
std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());<br>
}<br>
<br>
<br>
Modified: llvm/trunk/test/CodeGen/X86/block-placement.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/block-placement.ll?rev=306272&r1=306271&r2=306272&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/block-placement.ll?rev=306272&r1=306271&r2=306272&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/test/CodeGen/X86/block-placement.ll (original)<br>
+++ llvm/trunk/test/CodeGen/X86/block-placement.ll Sun Jun 25 22:27:27 2017<br>
@@ -354,6 +354,7 @@ define void @unnatural_cfg2() {<br>
; single-source GCC.<br>
; CHECK-LABEL: unnatural_cfg2<br>
; CHECK: %entry<br>
+; CHECK: %loop.header<br>
; CHECK: %loop.body1<br>
; CHECK: %loop.body2<br>
; CHECK: %loop.body4<br>
@@ -361,7 +362,6 @@ define void @unnatural_cfg2() {<br>
; CHECK: %loop.inner2.begin<br>
; CHECK: %loop.body3<br>
; CHECK: %loop.inner1.begin<br>
-; CHECK: %loop.header<br>
; CHECK: %bail<br>
<br>
entry:<br>
@@ -1491,6 +1491,54 @@ ret:<br>
ret void<br>
}<br>
<br>
+define i32 @not_rotate_if_extra_branch(i32 %count) {<br>
+; Test checks that there is no loop rotation<br>
+; if it introduces extra branch.<br>
+; Specifically in this case because best exit is .header<br>
+; but it has fallthrough to .middle block and last block in<br>
+; loop chain .slow does not have afallthrough to .header.<br>
+; CHECK-LABEL: not_rotate_if_extra_branch<br>
+; CHECK: %.entry<br>
+; CHECK: %.header<br>
+; CHECK: %.middle<br>
+; CHECK: %.backedge<br>
+; CHECK: %.slow<br>
+; CHECK: %.bailout<br>
+; CHECK: %.stop<br>
+.entry:<br>
+ %sum.0 = shl nsw i32 %count, 1<br>
+ br label %.header<br>
+<br>
+.header:<br>
+ %i = phi i32 [ %i.1, %.backedge ], [ 0, %.entry ]<br>
+ %sum = phi i32 [ %sum.1, %.backedge ], [ %sum.0, %.entry ]<br>
+ %is_exc = icmp sgt i32 %i, 9000000<br>
+ br i1 %is_exc, label %.bailout, label %.middle, !prof !13<br>
+<br>
+.bailout:<br>
+ %sum.2 = add nsw i32 %count, 1<br>
+ br label %.stop<br>
+<br>
+.middle:<br>
+ %pr.1 = and i32 %i, 1023<br>
+ %pr.2 = icmp eq i32 %pr.1, 0<br>
+ br i1 %pr.2, label %.slow, label %.backedge, !prof !14<br>
+<br>
+.slow:<br>
+ tail call void @effect(i32 %sum)<br>
+ br label %.backedge<br>
+<br>
+.backedge:<br>
+ %sum.1 = add nsw i32 %i, %sum<br>
+ %i.1 = add nsw i32 %i, 1<br>
+ %end = icmp slt i32 %i.1, %count<br>
+ br i1 %end, label %.header, label %.stop, !prof !15<br>
+<br>
+.stop:<br>
+ %sum.phi = phi i32 [ %sum.1, %.backedge ], [ %sum.2, %.bailout ]<br>
+ ret i32 %sum.phi<br>
+}<br>
+<br>
declare void @effect(i32)<br>
<br>
!5 = !{!"branch_weights", i32 84, i32 16}<br>
@@ -1501,3 +1549,6 @@ declare void @effect(i32)<br>
!10 = !{!"branch_weights", i32 90, i32 10}<br>
!11 = !{!"branch_weights", i32 1, i32 1}<br>
!12 = !{!"branch_weights", i32 5, i32 3}<br>
+!13 = !{!"branch_weights", i32 1, i32 1}<br>
+!14 = !{!"branch_weights", i32 1, i32 1023}<br>
+!15 = !{!"branch_weights", i32 4095, i32 1}<br>
<br>
Modified: llvm/trunk/test/CodeGen/X86/code_placement_cold_loop_blocks.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/code_placement_cold_loop_blocks.ll?rev=306272&r1=306271&r2=306272&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/code_placement_cold_loop_blocks.ll?rev=306272&r1=306271&r2=306272&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/test/CodeGen/X86/code_placement_cold_loop_blocks.ll (original)<br>
+++ llvm/trunk/test/CodeGen/X86/code_placement_cold_loop_blocks.ll Sun Jun 25 22:27:27 2017<br>
@@ -37,7 +37,7 @@ end:<br>
ret void<br>
}<br>
<br>
-define void @nested_loop_0() !prof !1 {<br>
+define void @nested_loop_0(i1 %flag) !prof !1 {<br>
; Test if a block that is cold in the inner loop but not cold in the outer loop<br>
; will merged to the outer loop chain.<br>
;<br>
@@ -68,8 +68,7 @@ if.then:<br>
<br>
if.else:<br>
call void @e()<br>
- %call2 = call zeroext i1 @a()<br>
- br i1 %call2, label %header2, label %header, !prof !3<br>
+ br i1 %flag, label %header2, label %header, !prof !3<br>
<br>
end:<br>
call void @f()<br>
<br>
<br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
</blockquote></div></div>