[llvm] r299890 - CodeGen: BranchFolding: Merge identical blocks, even if they are short.

[Kyle Butt via llvm-commits]

Author: iteratee
Date: Mon Apr 10 17:28:12 2017
New Revision: 299890

URL: http://llvm.org/viewvc/llvm-project?rev=299890&view=rev
Log:
CodeGen: BranchFolding: Merge identical blocks, even if they are short.

Merging identical blocks when it doesn't reduce fallthrough. It is common for
the blocks created from critical edge splitting to be identical. We would like
to merge these blocks whenever doing so would not reduce fallthrough.

Added:
    llvm/trunk/test/CodeGen/X86/tail-merge-identical.ll
Modified:
    llvm/trunk/lib/CodeGen/BranchFolding.cpp

Modified: llvm/trunk/lib/CodeGen/BranchFolding.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/BranchFolding.cpp?rev=299890&r1=299889&r2=299890&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/BranchFolding.cpp (original)
+++ llvm/trunk/lib/CodeGen/BranchFolding.cpp Mon Apr 10 17:28:12 2017
@@ -600,6 +600,22 @@ ProfitableToMerge(MachineBasicBlock *MBB
   if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin())
     return true;
 
+  // If both blocks are identical and end in a branch, merge them unless they
+  // both have a fallthrough predecessor and successor.
+  // We can only do this after block placement because it depends on whether
+  // there are fallthroughs, and we don't know until after layout.
+  if (AfterPlacement && I1 == MBB1->begin() && I2 == MBB2->begin()) {
+    auto BothFallThrough = [](MachineBasicBlock *MBB) {
+      if (MBB->succ_size() != 0 && !MBB->canFallThrough())
+        return false;
+      MachineFunction::iterator I(MBB);
+      MachineFunction *MF = MBB->getParent();
+      return (MBB != &*MF->begin()) && std::prev(I)->canFallThrough();
+    };
+    if (!BothFallThrough(MBB1) || !BothFallThrough(MBB2))
+      return true;
+  }
+
   // If both blocks have an unconditional branch temporarily stripped out,
   // count that as an additional common instruction for the following
   // heuristics. This heuristic is only accurate for single-succ blocks, so to

Added: llvm/trunk/test/CodeGen/X86/tail-merge-identical.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/tail-merge-identical.ll?rev=299890&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/tail-merge-identical.ll (added)
+++ llvm/trunk/test/CodeGen/X86/tail-merge-identical.ll Mon Apr 10 17:28:12 2017
@@ -0,0 +1,41 @@
+; RUN: llc -o - -verify-machineinstrs %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+ at data = external global [3 x i32], align 4
+ at store = external global i32, align 4
+
+; %else1 and %then2 end up lowering to identical blocks. These blocks should be
+; merged during tail-merging.
+; CHECK-LABEL: merge_identical_blocks
+; CHECK: movl $data+4
+; CHECK-NOT: movl $data+4
+; CHECK: retq
+define void @merge_identical_blocks(i1 %a, i1 %b) {
+entry:
+  br label %if1
+
+if1:                                              ; predfs = %entry
+  br i1 %a, label %else1, label %if2
+
+else1:                                            ; preds = %if1
+  %ptr.else1 = getelementptr inbounds [3 x i32], [3 x i32]* @data, i64 0, i32 1
+  br label %phi_join
+
+if2:                                              ; preds = %if1
+  br i1 %b, label %then2, label %else2
+
+then2:                                            ; preds = %if2
+  %ptr.then2 = getelementptr inbounds [3 x i32], [3 x i32]* @data, i64 0, i32 1
+  br label %phi_join
+
+else2:                                            ; preds = %if2
+  %ptr.else2 = getelementptr inbounds [3 x i32], [3 x i32]* @data, i64 0, i32 2
+  br label %phi_join
+
+phi_join:                                         ; preds = %else1, %then2, %else2
+  %val.ptr = phi i32* [ %ptr.else1, %else1 ], [ %ptr.then2, %then2 ], [ %ptr.else2, %else2 ]
+  %val = load i32, i32* %val.ptr, align 4
+  store i32 %val, i32* @store, align 4
+  ret void
+}