[PATCH] D31560: CodeGen: BlockPlacement: Don't always tail-duplicate with no other successor.

[Kyle Butt via Phabricator via llvm-commits]

iteratee updated this revision to Diff 94125.

https://reviews.llvm.org/D31560

Files:
  lib/CodeGen/MachineBlockPlacement.cpp
  test/CodeGen/X86/tail-dup-merge-loop-headers.ll
  test/CodeGen/X86/tail-dup-no-other-successor.ll


Index: test/CodeGen/X86/tail-dup-no-other-successor.ll
===================================================================
--- /dev/null
+++ test/CodeGen/X86/tail-dup-no-other-successor.ll
@@ -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}
Index: test/CodeGen/X86/tail-dup-merge-loop-headers.ll
===================================================================
--- test/CodeGen/X86/tail-dup-merge-loop-headers.ll
+++ test/CodeGen/X86/tail-dup-merge-loop-headers.ll
@@ -6,13 +6,13 @@
 ; 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
Index: lib/CodeGen/MachineBlockPlacement.cpp
===================================================================
--- lib/CodeGen/MachineBlockPlacement.cpp
+++ lib/CodeGen/MachineBlockPlacement.cpp
@@ -1484,11 +1484,7 @@
     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


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D31560.94125.patch
Type: text/x-patch
Size: 3248 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170404/a2de7188/attachment.bin>