[llvm] r242871 - [BranchFolding] do not iterate the aliases of virtual registers

[Jingyue Wu]

Author: jingyue
Date: Tue Jul 21 23:16:52 2015
New Revision: 242871

URL: http://llvm.org/viewvc/llvm-project?rev=242871&view=rev
Log:
[BranchFolding] do not iterate the aliases of virtual registers

Summary:
MCRegAliasIterator only works for physical registers. So, do not run it
on virtual registers.

With this issue fixed, we can resurrect the BranchFolding pass in NVPTX
backend.

Reviewers: jholewinski, bkramer

Subscribers: henryhu, meheff, llvm-commits, jholewinski

Differential Revision: http://reviews.llvm.org/D11174

Added:
    llvm/trunk/test/CodeGen/NVPTX/branch-fold.ll
Modified:
    llvm/trunk/lib/CodeGen/BranchFolding.cpp
    llvm/trunk/lib/Target/NVPTX/NVPTXTargetMachine.cpp

Modified: llvm/trunk/lib/CodeGen/BranchFolding.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/BranchFolding.cpp?rev=242871&r1=242870&r2=242871&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/BranchFolding.cpp (original)
+++ llvm/trunk/lib/CodeGen/BranchFolding.cpp Tue Jul 21 23:16:52 2015
@@ -12,7 +12,8 @@
 // it then removes.
 //
 // Note that this pass must be run after register allocation, it cannot handle
-// SSA form.
+// SSA form. It also must handle virtual registers for targets that emit virtual
+// ISA (e.g. NVPTX).
 //
 //===----------------------------------------------------------------------===//
 
@@ -1573,6 +1574,17 @@ static MachineBasicBlock *findFalseBlock
   return nullptr;
 }
 
+template <class Container>
+static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI,
+                                Container &Set) {
+  if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
+    for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
+      Set.insert(*AI);
+  } else {
+    Set.insert(Reg);
+  }
+}
+
 /// findHoistingInsertPosAndDeps - Find the location to move common instructions
 /// in successors to. The location is usually just before the terminator,
 /// however if the terminator is a conditional branch and its previous
@@ -1598,8 +1610,7 @@ MachineBasicBlock::iterator findHoisting
     if (!Reg)
       continue;
     if (MO.isUse()) {
-      for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
-        Uses.insert(*AI);
+      addRegAndItsAliases(Reg, TRI, Uses);
     } else {
       if (!MO.isDead())
         // Don't try to hoist code in the rare case the terminator defines a
@@ -1608,8 +1619,7 @@ MachineBasicBlock::iterator findHoisting
 
       // If the terminator defines a register, make sure we don't hoist
       // the instruction whose def might be clobbered by the terminator.
-      for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
-        Defs.insert(*AI);
+      addRegAndItsAliases(Reg, TRI, Defs);
     }
   }
 
@@ -1665,15 +1675,15 @@ MachineBasicBlock::iterator findHoisting
     if (!Reg)
       continue;
     if (MO.isUse()) {
-      for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
-        Uses.insert(*AI);
+      addRegAndItsAliases(Reg, TRI, Uses);
     } else {
       if (Uses.erase(Reg)) {
-        for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
-          Uses.erase(*SubRegs); // Use sub-registers to be conservative
+        if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
+          for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
+            Uses.erase(*SubRegs); // Use sub-registers to be conservative
+        }
       }
-      for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
-        Defs.insert(*AI);
+      addRegAndItsAliases(Reg, TRI, Defs);
     }
   }
 
@@ -1800,8 +1810,12 @@ bool BranchFolder::HoistCommonCodeInSucc
       unsigned Reg = MO.getReg();
       if (!Reg || !LocalDefsSet.count(Reg))
         continue;
-      for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
-        LocalDefsSet.erase(*AI);
+      if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
+        for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
+          LocalDefsSet.erase(*AI);
+      } else {
+        LocalDefsSet.erase(Reg);
+      }
     }
 
     // Track local defs so we can update liveins.
@@ -1813,8 +1827,7 @@ bool BranchFolder::HoistCommonCodeInSucc
       if (!Reg)
         continue;
       LocalDefs.push_back(Reg);
-      for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
-        LocalDefsSet.insert(*AI);
+      addRegAndItsAliases(Reg, TRI, LocalDefsSet);
     }
 
     HasDups = true;

Modified: llvm/trunk/lib/Target/NVPTX/NVPTXTargetMachine.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/NVPTX/NVPTXTargetMachine.cpp?rev=242871&r1=242870&r2=242871&view=diff
==============================================================================
--- llvm/trunk/lib/Target/NVPTX/NVPTXTargetMachine.cpp (original)
+++ llvm/trunk/lib/Target/NVPTX/NVPTXTargetMachine.cpp Tue Jul 21 23:16:52 2015
@@ -163,7 +163,6 @@ void NVPTXPassConfig::addIRPasses() {
   // NVPTXPrologEpilog pass (see NVPTXPrologEpilogPass.cpp).
   disablePass(&PrologEpilogCodeInserterID);
   disablePass(&MachineCopyPropagationID);
-  disablePass(&BranchFolderPassID);
   disablePass(&TailDuplicateID);
 
   addPass(createNVPTXImageOptimizerPass());

Added: llvm/trunk/test/CodeGen/NVPTX/branch-fold.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/NVPTX/branch-fold.ll?rev=242871&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/NVPTX/branch-fold.ll (added)
+++ llvm/trunk/test/CodeGen/NVPTX/branch-fold.ll Tue Jul 21 23:16:52 2015
@@ -0,0 +1,24 @@
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 -disable-cgp | FileCheck %s
+; Disable CGP which also folds branches, so that only BranchFolding is under
+; the spotlight.
+
+target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-nvidia-cuda"
+
+define void @foo(i32 %x, float* %output) {
+; CHECK-LABEL: .visible .func foo(
+; CHECK-NOT: bra.uni
+; CHECK-NOT: LBB0_
+  %1 = icmp eq i32 %x, 1
+  br i1 %1, label %then, label %else
+
+then:
+  br label %merge
+
+else:
+  br label %merge
+
+merge:
+  store float 2.0, float* %output
+  ret void
+}