Fix PR15155: lost vadd/vsplat optimization.

During lowering of a BUILD_VECTOR, we look for opportunities to use a
vector splat.  When the splatted value fits in 5 signed bits, a single
splat does the job.  When it doesn't fit in 5 bits but does fit in 6,
and is an even value, we can splat on half the value and add the result
to itself.

This last optimization hasn't been working recently because of improved
constant folding.  To circumvent this, create a pseudo VADD_SPLAT that
can be expanded during instruction selection.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@175632 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
index 0f943e82badff068d9ace4bf2d7de989e13e236b..01d731a80957e35fb129083c75552216c11cb54b 100644 (file)
--- a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
+++ b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
@@ -1322,6 +1322,36 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) {
     return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
                                   SDValue(Tmp, 0), GA);
   }
+  case PPCISD::VADD_SPLAT: {
+    // Convert: VADD_SPLAT elt, size
+    // Into:    tmp = VSPLTIS[BHW] elt
+    //          VADDU[BHW]M tmp, tmp
+    // Where:   [BHW] = B for size = 1, H for size = 2, W for size = 4
+    assert(isa<ConstantSDNode>(N->getOperand(0)) &&
+           isa<ConstantSDNode>(N->getOperand(1)) &&
+           "Invalid operand on VADD_SPLAT!");
+    int EltSize = N->getConstantOperandVal(1);
+    unsigned Opc1, Opc2;
+    EVT VT;
+    if (EltSize == 1) {
+      Opc1 = PPC::VSPLTISB;
+      Opc2 = PPC::VADDUBM;
+      VT = MVT::v16i8;
+    } else if (EltSize == 2) {
+      Opc1 = PPC::VSPLTISH;
+      Opc2 = PPC::VADDUHM;
+      VT = MVT::v8i16;
+    } else {
+      assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!");
+      Opc1 = PPC::VSPLTISW;
+      Opc2 = PPC::VADDUWM;
+      VT = MVT::v4i32;
+    }
+    SDValue Elt = getI32Imm(N->getConstantOperandVal(0));
+    SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, Elt);
+    SDValue TmpVal = SDValue(Tmp, 0);
+    return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal);
+  }
   }
 
   return SelectCode(N);

diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index e789112de210d6ef027c7197a1c30394066cc49e..f3ef38a3dca9017449a546fe8edf5aeb612c5179 100644 (file)
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -594,6 +594,7 @@ const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case PPCISD::GET_TLSLD_ADDR:  return "PPCISD::GET_TLSLD_ADDR";
   case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA";
   case PPCISD::ADDI_DTPREL_L:   return "PPCISD::ADDI_DTPREL_L";
+  case PPCISD::VADD_SPLAT:      return "PPCISD::VADD_SPLAT";
   }
 }
 
@@ -5020,14 +5021,13 @@ SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
   // If this value is in the range [-32,30] and is even, use:
   //    tmp = VSPLTI[bhw], result = add tmp, tmp
   if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
-    // FIXME: This is currently disabled because the ADD will be folded back
-    // into an invalid BUILD_VECTOR immediately.
-    return SDValue();
-#if 0
-    SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
-    Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
-    return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
-#endif
+    // To avoid having the optimization undone by constant folding, we
+    // convert to a pseudo that will be expanded later.
+    SDValue Elt = DAG.getConstant(SextVal >> 1, MVT::i32);
+    EVT VT = Op.getValueType();
+    int Size = VT == MVT::v16i8 ? 1 : (VT == MVT::v8i16 ? 2 : 4);
+    SDValue EltSize = DAG.getConstant(Size, MVT::i32);
+    return DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize);
   }
 
   // If this is 0x8000_0000 x 4, turn into vspltisw + vslw.  If it is

diff --git a/lib/Target/PowerPC/PPCISelLowering.h b/lib/Target/PowerPC/PPCISelLowering.h
index 12b3df7c9a54709ca5ad0f75e8cd6e0c447894d6..7cc2d1ac322d96d25a19779eb2efb3d0bde7e670 100644 (file)
--- a/lib/Target/PowerPC/PPCISelLowering.h
+++ b/lib/Target/PowerPC/PPCISelLowering.h
@@ -237,6 +237,11 @@ namespace llvm {
       /// sym@got@dtprel@l.
       ADDI_DTPREL_L,
 
+      /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
+      /// into an ADD of a VSPLTI with itself during instruction selection.
+      /// Necessary to avoid losing this optimization due to constant folds.
+      VADD_SPLAT,
+
       /// STD_32 - This is the STD instruction for use with "32-bit" registers.
       STD_32 = ISD::FIRST_TARGET_MEMORY_OPCODE,
 

diff --git a/test/CodeGen/PowerPC/vaddsplat.ll b/test/CodeGen/PowerPC/vaddsplat.ll
new file mode 100644 (file)
index 0000000..b4c16c1
--- /dev/null
+++ b/test/CodeGen/PowerPC/vaddsplat.ll
@@ -0,0 +1,77 @@
+; RUN: llc -O0 -mcpu=pwr7 <%s | FileCheck %s
+
+; Test optimization of build_vector into vadd/vsplt for 6-bit immediates.
+
+target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f128:128:128-v128:128:128-n32:64"
+target triple = "powerpc64-unknown-linux-gnu"
+
+%v4i32 = type <4 x i32>
+%v8i16 = type <8 x i16>
+%v16i8 = type <16 x i8>
+
+define void @test_v4i32_pos(%v4i32* %P, %v4i32* %S) {
+       %p = load %v4i32* %P
+       %r = add %v4i32 %p, < i32 18, i32 18, i32 18, i32 18 >
+       store %v4i32 %r, %v4i32* %S
+       ret void
+}
+
+; CHECK: test_v4i32_pos:
+; CHECK: vspltisw [[REG1:[0-9]+]], 9
+; CHECK: vadduwm {{[0-9]+}}, [[REG1]], [[REG1]]
+
+define void @test_v4i32_neg(%v4i32* %P, %v4i32* %S) {
+       %p = load %v4i32* %P
+       %r = add %v4i32 %p, < i32 -28, i32 -28, i32 -28, i32 -28 >
+       store %v4i32 %r, %v4i32* %S
+       ret void
+}
+
+; CHECK: test_v4i32_neg:
+; CHECK: vspltisw [[REG1:[0-9]+]], -14
+; CHECK: vadduwm {{[0-9]+}}, [[REG1]], [[REG1]]
+
+define void @test_v8i16_pos(%v8i16* %P, %v8i16* %S) {
+       %p = load %v8i16* %P
+       %r = add %v8i16 %p, < i16 30, i16 30, i16 30, i16 30, i16 30, i16 30, i16 30, i16 30 >
+       store %v8i16 %r, %v8i16* %S
+       ret void
+}
+
+; CHECK: test_v8i16_pos:
+; CHECK: vspltish [[REG1:[0-9]+]], 15
+; CHECK: vadduhm {{[0-9]+}}, [[REG1]], [[REG1]]
+
+define void @test_v8i16_neg(%v8i16* %P, %v8i16* %S) {
+       %p = load %v8i16* %P
+       %r = add %v8i16 %p, < i16 -32, i16 -32, i16 -32, i16 -32, i16 -32, i16 -32, i16 -32, i16 -32 >
+       store %v8i16 %r, %v8i16* %S
+       ret void
+}
+
+; CHECK: test_v8i16_neg:
+; CHECK: vspltish [[REG1:[0-9]+]], -16
+; CHECK: vadduhm {{[0-9]+}}, [[REG1]], [[REG1]]
+
+define void @test_v16i8_pos(%v16i8* %P, %v16i8* %S) {
+       %p = load %v16i8* %P
+       %r = add %v16i8 %p, < i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16, i8 16 >
+       store %v16i8 %r, %v16i8* %S
+       ret void
+}
+
+; CHECK: test_v16i8_pos:
+; CHECK: vspltisb [[REG1:[0-9]+]], 8
+; CHECK: vaddubm {{[0-9]+}}, [[REG1]], [[REG1]]
+
+define void @test_v16i8_neg(%v16i8* %P, %v16i8* %S) {
+       %p = load %v16i8* %P
+       %r = add %v16i8 %p, < i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18, i8 -18 >
+       store %v16i8 %r, %v16i8* %S
+       ret void
+}
+
+; CHECK: test_v16i8_neg:
+; CHECK: vspltisb [[REG1:[0-9]+]], -9
+; CHECK: vaddubm {{[0-9]+}}, [[REG1]], [[REG1]]
+