Collapse DomainValues across loop back-edges.

During the initial RPO traversal of the basic blocks, remember the ones
that are incomplete because of back-edges from predecessors that haven't
been visited yet.

After the initial RPO, revisit all those loop headers so the incoming
DomainValues on the back-edges can be properly collapsed.

This will properly fix execution domains on software pipelined code,
like the included test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@144151 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/ExecutionDepsFix.cpp

@@ -160,7 +160,7 @@ private:
   void collapse(DomainValue *dv, unsigned domain);
   bool merge(DomainValue *A, DomainValue *B);
 
-  void enterBasicBlock(MachineBasicBlock*);
+  bool enterBasicBlock(MachineBasicBlock*);
   void leaveBasicBlock(MachineBasicBlock*);
   void visitInstr(MachineInstr*);
   void visitGenericInstr(MachineInstr*);
@@ -317,7 +317,13 @@ bool ExeDepsFix::merge(DomainValue *A, DomainValue *B) {
   return true;
 }
 
-void ExeDepsFix::enterBasicBlock(MachineBasicBlock *MBB) {
+// enterBasicBlock - Set up LiveRegs by merging predecessor live-out values.
+// Return true if some predecessor hasn't been processed yet (like on a loop
+// back-edge).
+bool ExeDepsFix::enterBasicBlock(MachineBasicBlock *MBB) {
+  // Detect back-edges from predecessors we haven't processed yet.
+  bool seenBackEdge = false;
+
   // Try to coalesce live-out registers from predecessors.
   for (MachineBasicBlock::livein_iterator i = MBB->livein_begin(),
          e = MBB->livein_end(); i != e; ++i) {
@@ -326,7 +332,12 @@ void ExeDepsFix::enterBasicBlock(MachineBasicBlock *MBB) {
     for (MachineBasicBlock::const_pred_iterator pi = MBB->pred_begin(),
            pe = MBB->pred_end(); pi != pe; ++pi) {
       LiveOutMap::const_iterator fi = LiveOuts.find(*pi);
-      if (fi == LiveOuts.end()) continue;
+      if (fi == LiveOuts.end()) {
+        seenBackEdge = true;
+        continue;
+      }
+      if (!fi->second)
+        continue;
       DomainValue *pdv = resolve(fi->second[rx]);
       if (!pdv) continue;
       if (!LiveRegs || !LiveRegs[rx]) {
@@ -350,12 +361,19 @@ void ExeDepsFix::enterBasicBlock(MachineBasicBlock *MBB) {
         force(rx, pdv->getFirstDomain());
     }
   }
+  return seenBackEdge;
 }
 
 void ExeDepsFix::leaveBasicBlock(MachineBasicBlock *MBB) {
   // Save live registers at end of MBB - used by enterBasicBlock().
-  if (LiveRegs)
-    LiveOuts.insert(std::make_pair(MBB, LiveRegs));
+  // Also use LiveOuts as a visited set to detect back-edges.
+  if (!LiveOuts.insert(std::make_pair(MBB, LiveRegs)).second && LiveRegs) {
+    // Insertion failed, this must be the second pass.
+    // Release all the DomainValues instead of keeping them.
+    for (unsigned i = 0, e = NumRegs; i != e; ++i)
+      release(LiveRegs[i]);
+    delete[] LiveRegs;
+  }
   LiveRegs = 0;
 }
 
@@ -545,23 +563,32 @@ bool ExeDepsFix::runOnMachineFunction(MachineFunction &mf) {
 
   MachineBasicBlock *Entry = MF->begin();
   ReversePostOrderTraversal<MachineBasicBlock*> RPOT(Entry);
+  SmallVector<MachineBasicBlock*, 16> Loops;
   for (ReversePostOrderTraversal<MachineBasicBlock*>::rpo_iterator
          MBBI = RPOT.begin(), MBBE = RPOT.end(); MBBI != MBBE; ++MBBI) {
     MachineBasicBlock *MBB = *MBBI;
-    enterBasicBlock(MBB);
+    if (enterBasicBlock(MBB))
+      Loops.push_back(MBB);
     for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;
         ++I)
       visitInstr(I);
     leaveBasicBlock(MBB);
   }
 
+  // Visit all the loop blocks again in order to merge DomainValues from
+  // back-edges.
+  for (unsigned i = 0, e = Loops.size(); i != e; ++i) {
+    MachineBasicBlock *MBB = Loops[i];
+    enterBasicBlock(MBB);
+    leaveBasicBlock(MBB);
+  }
+
   // Clear the LiveOuts vectors and collapse any remaining DomainValues.
   for (ReversePostOrderTraversal<MachineBasicBlock*>::rpo_iterator
          MBBI = RPOT.begin(), MBBE = RPOT.end(); MBBI != MBBE; ++MBBI) {
     LiveOutMap::const_iterator FI = LiveOuts.find(*MBBI);
-    if (FI == LiveOuts.end())
+    if (FI == LiveOuts.end() || !FI->second)
       continue;
-    assert(FI->second && "Null entry");
     for (unsigned i = 0, e = NumRegs; i != e; ++i)
       if (FI->second[i])
         release(FI->second[i]);

test/CodeGen/X86/sse-domains.ll

@@ -0,0 +1,45 @@
+; RUN: llc < %s | FileCheck %s
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.7"
+
+; CHECK: f
+;
+; This function contains load / store / and operations that all can execute in
+; any domain.  The only domain-specific operation is the %add = shl... operation
+; which is <4 x i32>.
+;
+; The paddd instruction can only influence the other operations through the loop
+; back-edge. Check that everything is still moved into the integer domain.
+
+define void @f(<4 x i32>* nocapture %p, i32 %n) nounwind uwtable ssp {
+entry:
+  br label %while.body
+
+; Materialize a zeroinitializer and a constant-pool load in the integer domain.
+; The order is not important.
+; CHECK: pxor
+; CHECK: movdqa
+
+; The instructions in the loop must all be integer domain as well.
+; CHECK: while.body
+; CHECK: pand
+; CHECK: movdqa
+; CHECK: movdqa
+; Finally, the controlling integer-only instruction.
+; CHECK: paddd
+while.body:
+  %p.addr.04 = phi <4 x i32>* [ %incdec.ptr, %while.body ], [ %p, %entry ]
+  %n.addr.03 = phi i32 [ %dec, %while.body ], [ %n, %entry ]
+  %x.02 = phi <4 x i32> [ %add, %while.body ], [ zeroinitializer, %entry ]
+  %dec = add nsw i32 %n.addr.03, -1
+  %and = and <4 x i32> %x.02, <i32 127, i32 127, i32 127, i32 127>
+  %incdec.ptr = getelementptr inbounds <4 x i32>* %p.addr.04, i64 1
+  store <4 x i32> %and, <4 x i32>* %p.addr.04, align 16
+  %0 = load <4 x i32>* %incdec.ptr, align 16
+  %add = shl <4 x i32> %0, <i32 1, i32 1, i32 1, i32 1>
+  %tobool = icmp eq i32 %dec, 0
+  br i1 %tobool, label %while.end, label %while.body
+
+while.end:
+  ret void
+}