Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp =================================================================== --- lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -4544,16 +4544,20 @@ ConstantSDNode *BinOpCst = getAsNonOpaqueConstant(LHS->getOperand(1)); if (!BinOpCst) return SDValue(); - // FIXME: disable this unless the input to the binop is a shift by a constant. - // If it is not a shift, it pessimizes some common cases like: - // - // void foo(int *X, int i) { X[i & 1235] = 1; } - // int bar(int *X, int i) { return X[i & 255]; } + // FIXME: disable this unless the input to the binop is a shift by a constant + // or is copy/select.Enable this in other cases when figure out it's exactly profitable. SDNode *BinOpLHSVal = LHS->getOperand(0).getNode(); - if ((BinOpLHSVal->getOpcode() != ISD::SHL && - BinOpLHSVal->getOpcode() != ISD::SRA && - BinOpLHSVal->getOpcode() != ISD::SRL) || - !isa(BinOpLHSVal->getOperand(1))) + bool isShift = BinOpLHSVal->getOpcode() == ISD::SHL || + BinOpLHSVal->getOpcode() == ISD::SRA || + BinOpLHSVal->getOpcode() == ISD::SRL; + bool isCopyOrSelect = BinOpLHSVal->getOpcode() == ISD::CopyFromReg || + BinOpLHSVal->getOpcode() == ISD::SELECT; + + if ((!isShift || !isa(BinOpLHSVal->getOperand(1))) && + !isCopyOrSelect) + return SDValue(); + + if (isCopyOrSelect && N->hasOneUse()) return SDValue(); EVT VT = N->getValueType(0); Index: test/CodeGen/ARM/shift-combine.ll =================================================================== --- /dev/null +++ test/CodeGen/ARM/shift-combine.ll @@ -0,0 +1,31 @@ +; RUN: llc -mtriple=armv7-linux-gnueabihf %s -o - | FileCheck %s + +@array = weak global [4 x i32] zeroinitializer + +define i32 @test_lshr_and1(i32 %x) { +entry: +;CHECK-LABLE: test_lshr_and1: +;CHECK: movw r1, :lower16:array +;CHECK-NEXT: and r0, r0, #12 +;CHECK-NEXT: movt r1, :upper16:array +;CHECK-NEXT: ldr r0, [r1, r0] +;CHECK-NEXT: bx lr + %tmp2 = lshr i32 %x, 2 + %tmp3 = and i32 %tmp2, 3 + %tmp4 = getelementptr [4 x i32], [4 x i32]* @array, i32 0, i32 %tmp3 + %tmp5 = load i32, i32* %tmp4, align 4 + ret i32 %tmp5 +} +define i32 @test_lshr_and2(i32 %x) { +entry: +;CHECK-LABLE: test_lshr_and2: +;CHECK: ubfx r0, r0, #1, #15 +;CHECK-NEXT: add r0, r0, r0 +;CHECK-NEXT: bx lr + %a = and i32 %x, 65534 + %b = lshr i32 %a, 1 + %c = and i32 %x, 65535 + %d = lshr i32 %c, 1 + %e = add i32 %b, %d + ret i32 %e +} Index: test/CodeGen/Thumb2/machine-licm.ll =================================================================== --- test/CodeGen/Thumb2/machine-licm.ll +++ test/CodeGen/Thumb2/machine-licm.ll @@ -85,10 +85,9 @@ ; CHECK-LABEL: t3: bb.nph: ; CHECK: bb.nph -; CHECK: movw {{(r[0-9])|(lr)}}, #32768 +; CHECK: movw {{(r[0-9]+)|(lr)}}, #32768 ; CHECK: movs {{(r[0-9]+)|(lr)}}, #0 ; CHECK: movw [[REGISTER:(r[0-9]+)|(lr)]], #16386 -; CHECK: movw {{(r[0-9]+)|(lr)}}, #65534 ; CHECK: movt {{(r[0-9]+)|(lr)}}, #65535 br label %bb @@ -97,7 +96,6 @@ ; CHECK: eor.w ; CHECK: eorne.w {{(r[0-9])|(lr)}}, {{(r[0-9])|(lr)}}, [[REGISTER]] ; CHECK-NOT: eor -; CHECK: and %data_addr.013 = phi i8 [ %data, %bb.nph ], [ %8, %bb ] ; [#uses=2] %crc_addr.112 = phi i16 [ %crc, %bb.nph ], [ %crc_addr.2, %bb ] ; [#uses=3] %i.011 = phi i8 [ 0, %bb.nph ], [ %7, %bb ] ; [#uses=1]