[llvm] r299892 - CodeGen: BlockPlacement: Don't always tail-duplicate with no other successor.

[Kyle Butt via llvm-commits]

Author: iteratee
Date: Mon Apr 10 17:28:22 2017
New Revision: 299892

URL: http://llvm.org/viewvc/llvm-project?rev=299892&view=rev
Log:
CodeGen: BlockPlacement: Don't always tail-duplicate with no other successor.

The math works out where it can actually be counter-productive. The probability
calculations correctly handle the case where the alternative is 0 probability,
rely on those calculations.

Includes a test case that demonstrates the problem.

Added:
    llvm/trunk/test/CodeGen/X86/tail-dup-no-other-successor.ll
Modified:
    llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp
    llvm/trunk/test/CodeGen/X86/tail-dup-merge-loop-headers.ll

Modified: llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp?rev=299892&r1=299891&r2=299892&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp (original)
+++ llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp Mon Apr 10 17:28:22 2017
@@ -1489,11 +1489,7 @@ MachineBlockPlacement::selectBestSuccess
     if (DupProb < BestProb)
       break;
     if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter)
-        // If tail duplication gives us fallthrough when we otherwise wouldn't
-        // have it, that is a strict gain.
-        && (BestSucc.BB == nullptr
-            || isProfitableToTailDup(BB, Succ, BestProb, Chain,
-                                     BlockFilter))) {
+        && (isProfitableToTailDup(BB, Succ, BestProb, Chain, BlockFilter))) {
       DEBUG(
           dbgs() << "    Candidate: " << getBlockName(Succ) << ", probability: "
                  << DupProb

Modified: llvm/trunk/test/CodeGen/X86/tail-dup-merge-loop-headers.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/tail-dup-merge-loop-headers.ll?rev=299892&r1=299891&r2=299892&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/tail-dup-merge-loop-headers.ll (original)
+++ llvm/trunk/test/CodeGen/X86/tail-dup-merge-loop-headers.ll Mon Apr 10 17:28:22 2017
@@ -6,13 +6,13 @@ target triple = "x86_64-unknown-linux-gn
 ; CHECK-LABEL: tail_dup_merge_loops
 ; CHECK: # %entry
 ; CHECK-NOT: # %{{[a-zA-Z_]+}}
-; CHECK: # %exit
-; CHECK-NOT: # %{{[a-zA-Z_]+}}
 ; CHECK: # %inner_loop_exit
 ; CHECK-NOT: # %{{[a-zA-Z_]+}}
 ; CHECK: # %inner_loop_latch
 ; CHECK-NOT: # %{{[a-zA-Z_]+}}
 ; CHECK: # %inner_loop_test
+; CHECK-NOT: # %{{[a-zA-Z_]+}}
+; CHECK: # %exit
 define void @tail_dup_merge_loops(i32 %a, i8* %b, i8* %c) local_unnamed_addr #0 {
 entry:
   %notlhs674.i = icmp eq i32 %a, 0

Added: llvm/trunk/test/CodeGen/X86/tail-dup-no-other-successor.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/tail-dup-no-other-successor.ll?rev=299892&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/tail-dup-no-other-successor.ll (added)
+++ llvm/trunk/test/CodeGen/X86/tail-dup-no-other-successor.ll Mon Apr 10 17:28:22 2017
@@ -0,0 +1,53 @@
+; RUN: llc -O3 -o - %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+declare void @effect(i32);
+
+; After the loop gets laid out, loop.end is the only successor, but can't be
+; laid out because of the CFG dependency from top.fakephi. The calculations show
+; that it isn't profitable to tail-duplicate in this case, because of the
+; effects on fallthrough from %loop.end
+; CHECK-LABEL: {{^}}no_successor_still_no_taildup:
+; CHECK: %entry
+; CHECK: %loop.top
+; CHECK: %loop.latch
+; CHECK: %top.fakephi
+; CHECK: %loop.end
+; CHECK: %false
+; CHECK: %ret
+define void @no_successor_still_no_taildup (i32 %count, i32 %key) {
+entry:
+  br label %loop.top
+
+loop.top:
+  %i.loop.top = phi i32 [ %count, %entry ], [ %i.latch, %loop.latch ]
+  %cmp.top = icmp eq i32 %i.loop.top, %key
+  call void @effect(i32 0)
+  br i1 %cmp.top, label %top.fakephi, label %loop.latch, !prof !1
+
+loop.latch:
+  %i.latch = sub i32 %i.loop.top, 1
+  %cmp.latch = icmp eq i32 %i.latch, 0
+  call void @effect(i32 1)
+  br i1 %cmp.top, label %loop.top, label %loop.end, !prof !2
+
+top.fakephi:
+  call void @effect(i32 2)
+  br label %loop.end
+
+loop.end:
+  %cmp.end = icmp eq i32 %count, 0
+  br i1 %cmp.end, label %ret, label %false, !prof !3
+
+false:
+  call void @effect(i32 4)
+  br label %ret
+
+ret:
+  ret void
+}
+
+!1 = !{!"branch_weights", i32 1, i32 1}
+!2 = !{!"branch_weights", i32 5, i32 1}
+!3 = !{!"branch_weights", i32 1, i32 2}