Remove duplicated DMB instructions

ARM specific optimiztion, finding places in ARM machine code where 2 dmbs
follow one another, and eliminating one of them.

Patch by Reinoud Elhorst.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205409 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARM.h

@@ -42,6 +42,7 @@ FunctionPass *createARMGlobalMergePass(const TargetLowering* tli);
 FunctionPass *createARMConstantIslandPass();
 FunctionPass *createMLxExpansionPass();
 FunctionPass *createThumb2ITBlockPass();
+FunctionPass *createARMOptimizeBarriersPass();
 FunctionPass *createThumb2SizeReductionPass();
 
 /// \brief Creates an ARM-specific Target Transformation Info pass.

lib/Target/ARM/ARMOptimizeBarriersPass.cpp

@@ -0,0 +1,101 @@
+//===-- ARMOptimizeBarriersPass - two DMBs without a memory access in between,
+//removed one -===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===------------------------------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "double barriers"
+
+#include "ARM.h"
+#include "ARMMachineFunctionInfo.h"
+#include "ARMInstrInfo.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+using namespace llvm;
+
+STATISTIC(NumDMBsRemoved, "Number of DMBs removed");
+
+namespace {
+class ARMOptimizeBarriersPass : public MachineFunctionPass {
+public:
+  static char ID;
+  ARMOptimizeBarriersPass() : MachineFunctionPass(ID) {}
+
+  virtual bool runOnMachineFunction(MachineFunction &Fn);
+
+  virtual const char *getPassName() const {
+    return "optimise barriers pass";
+  }
+
+private:
+};
+char ARMOptimizeBarriersPass::ID = 0;
+}
+
+// Returns whether the instruction can safely move past a DMB instruction
+// The current implementation allows this iif MI does not have any possible
+// memory access
+static bool CanMovePastDMB(const MachineInstr *MI) {
+  return !(MI->mayLoad() ||
+          MI->mayStore() ||
+          MI->hasUnmodeledSideEffects() ||
+          MI->isCall() ||
+          MI->isReturn());
+}
+
+bool ARMOptimizeBarriersPass::runOnMachineFunction(MachineFunction &MF) {
+  // Vector to store the DMBs we will remove after the first iteration
+  std::vector<MachineInstr *> ToRemove;
+  // DMBType is the Imm value of the first operand. It determines whether it's a
+  // DMB ish, dmb sy, dmb osh, etc
+  int64_t DMBType = -1;
+
+  // Find a dmb. If we can move it until the next dmb, tag the second one for
+  // removal
+  for (auto &MBB : MF) {
+    // Will be true when we have seen a DMB, and not seen any instruction since
+    // that cannot move past a DMB
+    bool IsRemovableNextDMB = false;
+    for (auto &MI : MBB) {
+      if (MI.getOpcode() == ARM::DMB) {
+        if (IsRemovableNextDMB) {
+          // If the Imm of this DMB is the same as that of the last DMB, we can
+          // tag this second DMB for removal
+          if (MI.getOperand(0).getImm() == DMBType) {
+            ToRemove.push_back(&MI);
+          } else {
+            // If it has a different DMBType, we cannot remove it, but will scan
+            // for the next DMB, recording this DMB's type as last seen DMB type
+            DMBType = MI.getOperand(0).getImm();
+          }
+        } else {
+          // After we see a DMB, a next one is removable
+          IsRemovableNextDMB = true;
+          DMBType = MI.getOperand(0).getImm();
+        }
+      } else if (!CanMovePastDMB(&MI)) {
+        // If we find an instruction unable to pass past a DMB, a next DMB is
+        // not removable
+        IsRemovableNextDMB = false;
+      }
+    }
+  }
+  // Remove the tagged DMB
+  for (auto MI : ToRemove) {
+    MI->eraseFromParent();
+    ++NumDMBsRemoved;
+  }
+
+  return NumDMBsRemoved > 0;
+}
+
+/// createARMOptimizeBarriersPass - Returns an instance of the remove double
+/// barriers
+/// pass.
+FunctionPass *llvm::createARMOptimizeBarriersPass() {
+  return new ARMOptimizeBarriersPass();
+}

lib/Target/ARM/ARMTargetMachine.cpp

@@ -296,6 +296,7 @@ bool ARMPassConfig::addPreEmitPass() {
     addPass(&UnpackMachineBundlesID);
   }
 
+  addPass(createARMOptimizeBarriersPass());
   addPass(createARMConstantIslandPass());
 
   return true;

lib/Target/ARM/CMakeLists.txt

@@ -34,6 +34,7 @@ add_llvm_target(ARMCodeGen
   ARMMCInstLower.cpp
   ARMMachineFunctionInfo.cpp
   ARMRegisterInfo.cpp
+  ARMOptimizeBarriersPass.cpp
   ARMSelectionDAGInfo.cpp
   ARMSubtarget.cpp
   ARMTargetMachine.cpp

test/CodeGen/ARM/optimize-dmbs-v7.ll

@@ -0,0 +1,74 @@
+; RUN: llc < %s -mtriple=armv7 -mattr=+db | FileCheck %s
+
+@x1 = global i32 0, align 4
+@x2 = global i32 0, align 4
+
+define void @test() {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %i.013 = phi i32 [ 1, %entry ], [ %inc6, %for.body ]
+  store atomic i32 %i.013, i32* @x1 seq_cst, align 4
+  store atomic i32 %i.013, i32* @x1 seq_cst, align 4
+  store atomic i32 %i.013, i32* @x2 seq_cst, align 4
+  %inc6 = add nsw i32 %i.013, 1
+  %exitcond = icmp eq i32 %inc6, 2
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  ret void
+
+; The for.body contains 3 seq_cst stores.
+; Hence it should have 3 dmb;str;dmb sequences with the middle dmbs collapsed
+; CHECK: %for.body
+; CHECK-NOT: str
+; CHECK: dmb
+; CHECK-NOT: dmb
+; CHECK: str
+
+; CHECK-NOT: str
+; CHECK: dmb
+; CHECK-NOT: dmb
+; CHECK: str
+
+; CHECK-NOT: str
+; CHECK: dmb
+; CHECK-NOT: dmb
+; CHECK: str
+
+; CHECK-NOT: str
+; CHECK: dmb
+; CHECK-NOT: dmb
+; CHECK-NOT: str
+; CHECK: %for.end
+}
+
+define void @test2() {
+  call void @llvm.arm.dmb(i32 11)
+  tail call void @test()
+  call void @llvm.arm.dmb(i32 11)
+  ret void
+; the call should prevent the two dmbs from collapsing
+; CHECK: test2:
+; CHECK: dmb
+; CHECK-NEXT: bl
+; CHECK-NEXT: dmb
+}
+
+define void @test3() {
+  call void @llvm.arm.dmb(i32 11)
+  call void @llvm.arm.dsb(i32 9)
+  call void @llvm.arm.dmb(i32 11)
+  ret void
+; the call should prevent the two dmbs from collapsing
+; CHECK: test3:
+; CHECK: dmb
+; CHECK-NEXT: dsb
+; CHECK-NEXT: dmb
+
+}
+
+
+declare void @llvm.arm.dmb(i32)
+declare void @llvm.arm.dsb(i32)