X86: Fold tail calls into conditional branches where possible (PR26302)

When branching to a block that immediately tail calls, it is possible to fold
the call directly into the branch if the call is direct and there is no stack
adjustment, saving one byte.

Example:

  define void @f(i32 %x, i32 %y) {
  entry:
    %p = icmp eq i32 %x, %y
    br i1 %p, label %bb1, label %bb2
  bb1:
    tail call void @foo()
    ret void
  bb2:
    tail call void @bar()
    ret void
  }

before:

  f:
          movl    4(%esp), %eax
          cmpl    8(%esp), %eax
          jne     .LBB0_2
          jmp     foo
  .LBB0_2:
          jmp     bar

after:

  f:
          movl    4(%esp), %eax
          cmpl    8(%esp), %eax
          jne     bar
  .LBB0_1:
          jmp     foo

I don't expect any significant size savings from this (on a Clang bootstrap I
saw 288 bytes), but it does make the code a little tighter.

This patch only does 32-bit, but 64-bit would work similarly.

Differential Revision: https://reviews.llvm.org/D24108

llvm-svn: 280832

1 files changed, 37 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/BranchFolding.cpp b/llvm/lib/CodeGen/BranchFolding.cpp
index c5d4df63605..06a82220875 100644
--- a/llvm/lib/CodeGen/BranchFolding.cpp
+++ b/llvm/lib/CodeGen/BranchFolding.cpp
@@ -49,6 +49,7 @@ STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
 STATISTIC(NumBranchOpts, "Number of branches optimized");
 STATISTIC(NumTailMerge , "Number of block tails merged");
 STATISTIC(NumHoist     , "Number of times common instructions are hoisted");
+STATISTIC(NumTailCalls,  "Number of tail calls optimized");
 
 static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge",
                               cl::init(cl::BOU_UNSET), cl::Hidden);
@@ -1448,6 +1449,42 @@ ReoptimizeBlock:
     }
   }
 
+  if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 &&
+      MF.getFunction()->optForSize()) {
+    // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch
+    // direction, thereby defeating careful block placement and regressing
+    // performance. Therefore, only consider this for optsize functions.
+    MachineInstr &TailCall = *MBB->getFirstNonDebugInstr();
+    if (TII->isUnconditionalTailCall(TailCall)) {
+      MachineBasicBlock *Pred = *MBB->pred_begin();
+      MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
+      SmallVector<MachineOperand, 4> PredCond;
+      bool PredAnalyzable =
+          !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true);
+
+      if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB) {
+        // The predecessor has a conditional branch to this block which consists
+        // of only a tail call. Try to fold the tail call into the conditional
+        // branch.
+        if (TII->canMakeTailCallConditional(PredCond, TailCall)) {
+          // TODO: It would be nice if analyzeBranch() could provide a pointer
+          // to the branch insturction so replaceBranchWithTailCall() doesn't
+          // have to search for it.
+          TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall);
+          ++NumTailCalls;
+          Pred->removeSuccessor(MBB);
+          MadeChange = true;
+          return MadeChange;
+        }
+      }
+      // If the predecessor is falling through to this block, we could reverse
+      // the branch condition and fold the tail call into that. However, after
+      // that we might have to re-arrange the CFG to fall through to the other
+      // block and there is a high risk of regressing code size rather than
+      // improving it.
+    }
+  }
+
   // Analyze the branch in the current block.
   MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr;
   SmallVector<MachineOperand, 4> CurCond;