ARM: implement @llvm.readcyclecounter intrinsic

This implements the @llvm.readcyclecounter intrinsic as the specific
MRC instruction specified in the ARM manuals for CPUs with the Power
Management extensions.

Older CPUs had slightly different methods which may also have to be
implemented eventually, but this should cover all v7 cases.

rdar://problem/13939186

llvm-svn: 182603

3 files changed, 70 insertions, 1 deletions

diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst
index 4eecd15a6fc..5e32eba7f72 100644
--- a/llvm/docs/LangRef.rst
+++ b/llvm/docs/LangRef.rst
@@ -6663,6 +6663,9 @@ memory. Implementations are allowed to either return a application
 specific value or a system wide value. On backends without support, this
 is lowered to a constant 0.
 
+Note that runtime support may be conditional on the privilege-level code is
+running at and the host platform.
+
 Standard C Library Intrinsics
 -----------------------------
 
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index a1443d1b020..4238c72c139 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -681,6 +681,8 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
   setOperationAction(ISD::CTTZ_ZERO_UNDEF  , MVT::i32  , Expand);
   setOperationAction(ISD::CTLZ_ZERO_UNDEF  , MVT::i32  , Expand);
 
+  setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
+
   // Only ARMv6 has BSWAP.
   if (!Subtarget->hasV6Ops())
     setOperationAction(ISD::BSWAP, MVT::i32, Expand);
@@ -5702,7 +5704,6 @@ static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
   return SDValue();
 }
 
-
 static void
 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
                     SelectionDAG &DAG, unsigned NewOp) {
@@ -5736,6 +5737,44 @@ ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
   Results.push_back(Result.getValue(2));
 }
 
+static void ReplaceREADCYCLECOUNTER(SDNode *N,
+                                    SmallVectorImpl<SDValue> &Results,
+                                    SelectionDAG &DAG,
+                                    const ARMSubtarget *Subtarget) {
+  DebugLoc DL = N->getDebugLoc();
+  SDValue Cycles32, OutChain;
+
+  if (Subtarget->hasPerfMon()) {
+    // Under Power Management extensions, the cycle-count is:
+    //    mrc p15, #0, <Rt>, c9, c13, #0
+    SDValue Ops[] = { N->getOperand(0), // Chain
+                      DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
+                      DAG.getConstant(15, MVT::i32),
+                      DAG.getConstant(0, MVT::i32),
+                      DAG.getConstant(9, MVT::i32),
+                      DAG.getConstant(13, MVT::i32),
+                      DAG.getConstant(0, MVT::i32)
+    };
+
+    Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
+                           DAG.getVTList(MVT::i32, MVT::Other), &Ops[0],
+                           array_lengthof(Ops));
+    OutChain = Cycles32.getValue(1);
+  } else {
+    // Intrinsic is defined to return 0 on unsupported platforms. Technically
+    // there are older ARM CPUs that have implementation-specific ways of
+    // obtaining this information (FIXME!).
+    Cycles32 = DAG.getConstant(0, MVT::i32);
+    OutChain = DAG.getEntryNode();
+  }
+
+
+  SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
+                                 Cycles32, DAG.getConstant(0, MVT::i32));
+  Results.push_back(Cycles64);
+  Results.push_back(OutChain);
+}
+
 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   switch (Op.getOpcode()) {
   default: llvm_unreachable("Don't know how to custom lower this!");
@@ -5813,6 +5852,9 @@ void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
   case ISD::SRA:
     Res = Expand64BitShift(N, DAG, Subtarget);
     break;
+  case ISD::READCYCLECOUNTER:
+    ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
+    return;
   case ISD::ATOMIC_LOAD_ADD:
     ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
     return;
diff --git a/llvm/test/CodeGen/ARM/readcyclecounter.ll b/llvm/test/CodeGen/ARM/readcyclecounter.ll
new file mode 100644
index 00000000000..db47ad355d0
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/readcyclecounter.ll
@@ -0,0 +1,24 @@
+; RUN: llc -mtriple=armv7-none-linux-gnueabi < %s | FileCheck %s
+; RUN: llc -mtriple=thumbv7-none-linux-gnueabi < %s | FileCheck %s
+; RUN: llc -mtriple=armv7-none-linux-gnueabi -mattr=-perfmon < %s | FileCheck %s --check-prefix=CHECK-NO-PERFMON
+; RUN: llc -mtriple=armv6-none-linux-gnueabi < %s | FileCheck %s --check-prefix=CHECK-NO-PERFMON
+
+; The performance monitor we're looking for is an ARMv7 extension. It should be
+; possible to disable it, but realistically present on at least every v7-A
+; processor (but not on v6, at least by default).
+
+declare i64 @llvm.readcyclecounter()
+
+define i64 @get_count() {
+  %val = call i64 @llvm.readcyclecounter()
+  ret i64 %val
+
+  ; As usual, exact registers only sort of matter but the cycle-count had better
+  ; end up in r0 in the end.
+
+; CHECK: mrc p15, #0, r0, c9, c13, #0
+; CHECK: {{movs?}} r1, #0
+
+; CHECK-NO-PERFMON: {{movs?}} r0, #0
+; CHECK-NO-PERFMON: {{movs?}} r1, #0
+}