[AMDGPU] Convert shl (add) into add (shl)

shl (or|add x, c2), c1 => or|add (shl x, c1), (c2 << c1)
This allows to fold a constant into an address in some cases as
well as to eliminate second shift if the expression is used as
an address and second shift is a result of a GEP.

Differential Revision: https://reviews.llvm.org/D33432

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@303641 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 5ec46a8294c0c07355ac5990deceb950543c1502..9d7562d176d1acaefad0e385aaca5df5d7b156a6 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -127,6 +127,29 @@ EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
   return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
 }
 
+bool AMDGPUTargetLowering::isOrEquivalentToAdd(SelectionDAG &DAG, SDValue Op)
+{
+  assert(Op.getOpcode() == ISD::OR);
+
+  SDValue N0 = Op->getOperand(0);
+  SDValue N1 = Op->getOperand(1);
+  EVT VT = N0.getValueType();
+
+  if (VT.isInteger() && !VT.isVector()) {
+    KnownBits LHSKnown, RHSKnown;
+    DAG.computeKnownBits(N0, LHSKnown);
+
+    if (LHSKnown.Zero.getBoolValue()) {
+      DAG.computeKnownBits(N1, RHSKnown);
+
+      if (!(~RHSKnown.Zero & ~LHSKnown.Zero))
+        return true;
+    }
+  }
+
+  return false;
+}
+
 AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
                                            const AMDGPUSubtarget &STI)
     : TargetLowering(TM), Subtarget(&STI) {
@@ -2596,8 +2619,6 @@ SDValue AMDGPUTargetLowering::splitBinaryBitConstantOpImpl(
 SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
   EVT VT = N->getValueType(0);
-  if (VT != MVT::i64)
-    return SDValue();
 
   ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
   if (!RHS)
@@ -2618,6 +2639,8 @@ SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
   case ISD::SIGN_EXTEND:
   case ISD::ANY_EXTEND: {
     // shl (ext x) => zext (shl x), if shift does not overflow int
+    if (VT != MVT::i64)
+      break;
     KnownBits Known;
     SDValue X = LHS->getOperand(0);
     DAG.computeKnownBits(X, Known);
@@ -2628,8 +2651,23 @@ SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
     SDValue Shl = DAG.getNode(ISD::SHL, SL, XVT, X, SDValue(RHS, 0));
     return DAG.getZExtOrTrunc(Shl, SL, VT);
   }
+  case ISD::OR:  if (!isOrEquivalentToAdd(DAG, LHS)) break;
+  case ISD::ADD: { // Fall through from above
+    // shl (or|add x, c2), c1 => or|add (shl x, c1), (c2 << c1)
+    if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(LHS->getOperand(1))) {
+      SDValue Shl = DAG.getNode(ISD::SHL, SL, VT, LHS->getOperand(0),
+                                SDValue(RHS, 0));
+      SDValue C2V = DAG.getConstant(C2->getAPIntValue() << RHSVal,
+                                    SDLoc(C2), VT);
+      return DAG.getNode(LHS->getOpcode(), SL, VT, Shl, C2V);
+    }
+    break;
+  }
   }
 
+  if (VT != MVT::i64)
+    return SDValue();
+
   // i64 (shl x, C) -> (build_pair 0, (shl x, C -32))
 
   // On some subtargets, 64-bit shift is a quarter rate instruction. In the

diff --git a/lib/Target/AMDGPU/AMDGPUISelLowering.h b/lib/Target/AMDGPU/AMDGPUISelLowering.h
index fb2f15022d259d02b3bb10b07901a84de0f510b7..0d066cdbdff4d8c58a97e34fa6b3135bd2974937 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -34,6 +34,9 @@ private:
   /// compare.
   SDValue getFFBH_U32(SelectionDAG &DAG, SDValue Op, const SDLoc &DL) const;
 
+public:
+  static bool isOrEquivalentToAdd(SelectionDAG &DAG, SDValue Op);
+
 protected:
   const AMDGPUSubtarget *Subtarget;
   AMDGPUAS AMDGPUASI;

diff --git a/test/CodeGen/AMDGPU/shl-add-to-add-shl.ll b/test/CodeGen/AMDGPU/shl-add-to-add-shl.ll
new file mode 100644 (file)
index 0000000..1cdfec9
--- /dev/null
+++ b/test/CodeGen/AMDGPU/shl-add-to-add-shl.ll
@@ -0,0 +1,40 @@
+; RUN: llc -march=amdgcn -mcpu=fiji < %s | FileCheck %s
+
+; Check transformation shl (or|add x, c2), c1 => or|add (shl x, c1), (c2 << c1)
+; Only one shift if expected, GEP shall not produce a separate shift
+
+; CHECK-LABEL: {{^}}add_const_offset:
+; CHECK: v_lshlrev_b32_e32 v[[SHL:[0-9]+]], 4, v0
+; CHECK: v_add_i32_e32 v[[ADD:[0-9]+]], vcc, 0xc80, v[[SHL]]
+; CHECK-NOT: v_lshl
+; CHECK: v_add_i32_e32 v[[ADDRLO:[0-9]+]], vcc, s{{[0-9]+}}, v[[ADD]]
+; CHECK: load_dword v{{[0-9]+}}, v{{\[}}[[ADDRLO]]:
+define amdgpu_kernel void @add_const_offset(i32 addrspace(1)* nocapture %arg) {
+bb:
+  %id = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %add = add i32 %id, 200
+  %shl = shl i32 %add, 2
+  %ptr = getelementptr inbounds i32, i32 addrspace(1)* %arg, i32 %shl
+  %val = load i32, i32 addrspace(1)* %ptr, align 4
+  store i32 %val, i32 addrspace(1)* %arg, align 4
+  ret void
+}
+
+; CHECK-LABEL: {{^}}or_const_offset:
+; CHECK: v_lshlrev_b32_e32 v[[SHL:[0-9]+]], 4, v0
+; CHECK: v_or_b32_e32 v[[OR:[0-9]+]], 0x1000, v[[SHL]]
+; CHECK-NOT: v_lshl
+; CHECK: v_add_i32_e32 v[[ADDRLO:[0-9]+]], vcc, s{{[0-9]+}}, v[[OR]]
+; CHECK: load_dword v{{[0-9]+}}, v{{\[}}[[ADDRLO]]:
+define amdgpu_kernel void @or_const_offset(i32 addrspace(1)* nocapture %arg) {
+bb:
+  %id = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %add = or i32 %id, 256
+  %shl = shl i32 %add, 2
+  %ptr = getelementptr inbounds i32, i32 addrspace(1)* %arg, i32 %shl
+  %val = load i32, i32 addrspace(1)* %ptr, align 4
+  store i32 %val, i32 addrspace(1)* %arg, align 4
+  ret void
+}
+
+declare i32 @llvm.amdgcn.workitem.id.x()