[llvm] r290410 - Make the canonicalisation on shifts benifit to more case.

[Zijiao Ma via llvm-commits]

Author: zijiaoma
Date: Thu Dec 22 20:56:07 2016
New Revision: 290410

URL: http://llvm.org/viewvc/llvm-project?rev=290410&view=rev
Log:
Make the canonicalisation on shifts benifit to more case.

1.Fix pessimized case in FIXME.
2.Add tests for it.
3.The canonicalisation on shifts results in different sequence for
  tests of machine-licm.Correct some check lines.

Differential Revision: https://reviews.llvm.org/D27916

Added:
    llvm/trunk/test/CodeGen/ARM/shift-combine.ll
Modified:
    llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
    llvm/trunk/test/CodeGen/Thumb2/machine-licm.ll

Modified: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp?rev=290410&r1=290409&r2=290410&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp Thu Dec 22 20:56:07 2016
@@ -4544,16 +4544,20 @@ SDValue DAGCombiner::visitShiftByConstan
   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<ConstantSDNode>(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<ConstantSDNode>(BinOpLHSVal->getOperand(1))) &&
+      !isCopyOrSelect)
+    return SDValue();
+
+  if (isCopyOrSelect && N->hasOneUse())
     return SDValue();
 
   EVT VT = N->getValueType(0);

Added: llvm/trunk/test/CodeGen/ARM/shift-combine.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/shift-combine.ll?rev=290410&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/ARM/shift-combine.ll (added)
+++ llvm/trunk/test/CodeGen/ARM/shift-combine.ll Thu Dec 22 20:56:07 2016
@@ -0,0 +1,31 @@
+; RUN: llc -mtriple=armv7-linux-gnueabihf %s -o - | FileCheck %s
+
+ at 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
+}

Modified: llvm/trunk/test/CodeGen/Thumb2/machine-licm.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/Thumb2/machine-licm.ll?rev=290410&r1=290409&r2=290410&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/Thumb2/machine-licm.ll (original)
+++ llvm/trunk/test/CodeGen/Thumb2/machine-licm.ll Thu Dec 22 20:56:07 2016
@@ -85,10 +85,9 @@ define zeroext i16 @t3(i8 zeroext %data,
 ; 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 @@ bb:
 ; 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 ] ; <i8> [#uses=2]
   %crc_addr.112 = phi i16 [ %crc, %bb.nph ], [ %crc_addr.2, %bb ] ; <i16> [#uses=3]
   %i.011 = phi i8 [ 0, %bb.nph ], [ %7, %bb ]     ; <i8> [#uses=1]