R600: Calculate store mask instead of using switch.

[oota-llvm.git] / lib / CodeGen / AggressiveAntiDepBreaker.cpp

diff --git a/lib/CodeGen/AggressiveAntiDepBreaker.cpp b/lib/CodeGen/AggressiveAntiDepBreaker.cpp
index c50f8b5a42ad9e02a5e56586cc6e67568f91d816..25c438c9615a43eec4e6ec9f91fcc62808196048 100644 (file)
--- a/lib/CodeGen/AggressiveAntiDepBreaker.cpp
+++ b/lib/CodeGen/AggressiveAntiDepBreaker.cpp
@@ -247,8 +247,8 @@ void AggressiveAntiDepBreaker::GetPassthruRegs(MachineInstr *MI,
     if ((MO.isDef() && MI->isRegTiedToUseOperand(i)) ||
         IsImplicitDefUse(MI, MO)) {
       const unsigned Reg = MO.getReg();
-      PassthruRegs.insert(Reg);
-      for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs)
+      for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
+           SubRegs.isValid(); ++SubRegs)
         PassthruRegs.insert(*SubRegs);
     }
   }
@@ -403,8 +403,18 @@ void AggressiveAntiDepBreaker::PrescanInstruction(MachineInstr *MI,
       continue;
 
     // Update def for Reg and aliases.
-    for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
+    for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) {
+      // We need to be careful here not to define already-live super registers.
+      // If the super register is already live, then this definition is not
+      // a definition of the whole super register (just a partial insertion
+      // into it). Earlier subregister definitions (which we've not yet visited
+      // because we're iterating bottom-up) need to be linked to the same group
+      // as this definition.
+      if (TRI->isSuperRegister(Reg, *AI) && State->IsLive(*AI))
+        continue;
+
       DefIndices[*AI] = Count;
+    }
   }
 }
 
@@ -782,7 +792,7 @@ unsigned AggressiveAntiDepBreaker::BreakAntiDependencies(
     if (MI == CriticalPathMI) {
       CriticalPathSU = CriticalPathStep(CriticalPathSU);
       CriticalPathMI = (CriticalPathSU) ? CriticalPathSU->getInstr() : 0;
-    } else {
+    } else if (CriticalPathSet.any()) {
       ExcludeRegs = &CriticalPathSet;
     }