Don't taint relaxed loads that immediately comes before an AcqRel read-modify-write op

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

diff --git a/lib/CodeGen/AtomicExpandPass.cpp b/lib/CodeGen/AtomicExpandPass.cpp
index d12fdb246984e9db8949340221ec2935d99883ba..077c52b19a7a3d1907cdd3dcda9eb967ccaf0f51 100644 (file)
--- a/lib/CodeGen/AtomicExpandPass.cpp
+++ b/lib/CodeGen/AtomicExpandPass.cpp
@@ -15,6 +15,11 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "llvm/ADT/SetOperations.h"
+#include "llvm/ADT/SetVector.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Analysis/MemoryLocation.h"
 #include "llvm/CodeGen/AtomicExpandUtils.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/IR/Function.h"
@@ -23,11 +28,13 @@
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/Module.h"
+#include "llvm/IR/NoFolder.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Target/TargetLowering.h"
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetSubtargetInfo.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
 
 using namespace llvm;
 
@@ -64,6 +71,33 @@ namespace {
     bool isIdempotentRMW(AtomicRMWInst *AI);
     bool simplifyIdempotentRMW(AtomicRMWInst *AI);
   };
+
+
+  // If 'LI' is a relaxed load, and it is immediately followed by a
+// atomic read-modify-write that has acq_rel parameter, we don't have to do
+// anything since the rmw serves as a natural barrier.
+void MarkRelaxedLoadBeforeAcqrelRMW(LoadInst* LI) {
+  auto* BB = LI->getParent();
+  auto BBI = LI->getIterator();
+  for (BBI++; BBI != BB->end(); BBI++) {
+    Instruction* CurInst = &*BBI;
+    if (!CurInst) {
+      return;
+    }
+    if (!CurInst->isAtomic()) {
+      continue;
+    }
+    auto* RMW = dyn_cast<AtomicRMWInst>(CurInst);
+    if (!RMW) {
+      return;
+    }
+    if (RMW->getOrdering() == AcquireRelease ||
+        RMW->getOrdering() == SequentiallyConsistent) {
+      LI->setHasSubsequentAcqlRMW(true);
+    }
+  }
+}
+
 }
 
 char AtomicExpand::ID = 0;
@@ -82,12 +116,61 @@ bool AtomicExpand::runOnFunction(Function &F) {
   TLI = TM->getSubtargetImpl(F)->getTargetLowering();
 
   SmallVector<Instruction *, 1> AtomicInsts;
+  SmallVector<LoadInst*, 1> MonotonicLoadInsts;
 
   // Changing control-flow while iterating through it is a bad idea, so gather a
   // list of all atomic instructions before we start.
   for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
-    if (I->isAtomic())
+    // XXX-update: For relaxed loads, change them to acquire. This includes
+    // relaxed loads, relaxed atomic RMW & relaxed atomic compare exchange.
+    if (I->isAtomic()) {
+      switch (I->getOpcode()) {
+        case Instruction::AtomicCmpXchg: {
+          // XXX-comment: AtomicCmpXchg in AArch64 will be translated to a
+          // conditional branch that contains the value of the load anyway, so
+          // we don't need to do anything.
+          /*
+          auto* CmpXchg = dyn_cast<AtomicCmpXchgInst>(&*I);
+          auto SuccOrdering = CmpXchg->getSuccessOrdering();
+          if (SuccOrdering == Monotonic) {
+            CmpXchg->setSuccessOrdering(Acquire);
+          } else if (SuccOrdering == Release) {
+            CmpXchg->setSuccessOrdering(AcquireRelease);
+          }
+          */
+          break;
+        }
+        case Instruction::AtomicRMW: {
+          // XXX-comment: Similar to AtomicCmpXchg. These instructions in
+          // AArch64 will be translated to a loop whose condition depends on the
+          // store status, which further depends on the load value.
+          /*
+          auto* RMW = dyn_cast<AtomicRMWInst>(&*I);
+          if (RMW->getOrdering() == Monotonic) {
+            RMW->setOrdering(Acquire);
+          }
+          */
+          break;
+        }
+        case Instruction::Load: {
+          auto* LI = dyn_cast<LoadInst>(&*I);
+          if (LI->getOrdering() == Monotonic) {
+            /*
+            DEBUG(dbgs() << "Transforming relaxed loads to acquire loads: "
+                         << *LI << '\n');
+            LI->setOrdering(Acquire);
+            */
+//            MonotonicLoadInsts.push_back(LI);
+            MarkRelaxedLoadBeforeAcqrelRMW(LI);
+          }
+          break;
+        }
+        default: {
+          break;
+        }
+      }
       AtomicInsts.push_back(&*I);
+    }
   }
 
   bool MadeChange = false;
@@ -104,7 +187,7 @@ bool AtomicExpand::runOnFunction(Function &F) {
     if (TLI->getInsertFencesForAtomic()) {
       if (LI && isAtLeastAcquire(LI->getOrdering())) {
         FenceOrdering = LI->getOrdering();
-        LI->setOrdering(Monotonic);
+//        AddFakeConditionalBranch(
         IsStore = false;
         IsLoad = true;
       } else if (SI && isAtLeastRelease(SI->getOrdering())) {
@@ -172,6 +255,7 @@ bool AtomicExpand::runOnFunction(Function &F) {
       MadeChange |= expandAtomicCmpXchg(CASI);
     }
   }
+
   return MadeChange;
 }