Add support for NEON VMVN immediate instructions.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@108324 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 9e9192b..0091df7 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -625,6 +625,7 @@
case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
+ case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
case ARMISD::VDUP: return "ARMISD::VDUP";
case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
case ARMISD::VEXT: return "ARMISD::VEXT";
@@ -2925,6 +2926,8 @@
switch (SplatBitSize) {
case 8:
+ if (!isVMOV)
+ return SDValue();
// Any 1-byte value is OK. Op=0, Cmode=1110.
assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
OpCmode = 0xe;
@@ -3006,9 +3009,9 @@
return SDValue();
case 64: {
- // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
if (!isVMOV)
return SDValue();
+ // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
uint64_t BitMask = 0xff;
uint64_t Val = 0;
unsigned ImmMask = 1;
@@ -3248,6 +3251,17 @@
SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
}
+
+ // Try an immediate VMVN.
+ uint64_t NegatedImm = (SplatBits.getZExtValue() ^
+ ((1LL << SplatBitSize) - 1));
+ Val = isNEONModifiedImm(NegatedImm,
+ SplatUndef.getZExtValue(), SplatBitSize,
+ DAG, VmovVT, VT.is128BitVector(), false);
+ if (Val.getNode()) {
+ SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vmov);
+ }
}
}
@@ -4232,14 +4246,15 @@
/// ARMISD::VDUPLANE.
static SDValue PerformVDUPLANECombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI) {
- // If the source is already a VMOVIMM splat, the VDUPLANE is redundant.
+ // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
+ // redundant.
SDValue Op = N->getOperand(0);
EVT VT = N->getValueType(0);
// Ignore bit_converts.
while (Op.getOpcode() == ISD::BIT_CONVERT)
Op = Op.getOperand(0);
- if (Op.getOpcode() != ARMISD::VMOVIMM)
+ if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
return SDValue();
// Make sure the VMOV element size is not bigger than the VDUPLANE elements.
diff --git a/lib/Target/ARM/ARMISelLowering.h b/lib/Target/ARM/ARMISelLowering.h
index 81120c8..128b72e 100644
--- a/lib/Target/ARM/ARMISelLowering.h
+++ b/lib/Target/ARM/ARMISelLowering.h
@@ -124,8 +124,11 @@
VGETLANEu, // zero-extend vector extract element
VGETLANEs, // sign-extend vector extract element
- // Vector duplicate:
+ // Vector move immediate and move negated immediate:
VMOVIMM,
+ VMVNIMM,
+
+ // Vector duplicate:
VDUP,
VDUPLANE,
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index bde806d..7f7eb98 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -65,9 +65,10 @@
def NEONvgetlaneu : SDNode<"ARMISD::VGETLANEu", SDTARMVGETLN>;
def NEONvgetlanes : SDNode<"ARMISD::VGETLANEs", SDTARMVGETLN>;
-def NEONvmovImm : SDNode<"ARMISD::VMOVIMM",
- SDTypeProfile<1, 1, [SDTCisVec<0>,
- SDTCisVT<1, i32>]>>;
+def SDTARMVMOVIMM : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVT<1, i32>]>;
+def NEONvmovImm : SDNode<"ARMISD::VMOVIMM", SDTARMVMOVIMM>;
+def NEONvmvnImm : SDNode<"ARMISD::VMVNIMM", SDTARMVMOVIMM>;
+
def NEONvdup : SDNode<"ARMISD::VDUP", SDTypeProfile<1, 1, [SDTCisVec<0>]>>;
// VDUPLANE can produce a quad-register result from a double-register source,
@@ -2383,6 +2384,28 @@
[(set QPR:$dst, (v4i32 (or QPR:$src1,
(vnotq QPR:$src2))))]>;
+// VMVN : Vector Bitwise NOT (Immediate)
+
+let isReMaterializable = 1 in {
+def VMVNv4i16 : N1ModImm<1, 0b000, {1,0,?,0}, 0, 0, 1, 1, (outs DPR:$dst),
+ (ins nModImm:$SIMM), IIC_VMOVImm,
+ "vmvn", "i16", "$dst, $SIMM", "",
+ [(set DPR:$dst, (v4i16 (NEONvmvnImm timm:$SIMM)))]>;
+def VMVNv8i16 : N1ModImm<1, 0b000, {1,0,?,0}, 0, 1, 1, 1, (outs QPR:$dst),
+ (ins nModImm:$SIMM), IIC_VMOVImm,
+ "vmvn", "i16", "$dst, $SIMM", "",
+ [(set QPR:$dst, (v8i16 (NEONvmvnImm timm:$SIMM)))]>;
+
+def VMVNv2i32 : N1ModImm<1, 0b000, {?,?,?,?}, 0, 0, 1, 1, (outs DPR:$dst),
+ (ins nModImm:$SIMM), IIC_VMOVImm,
+ "vmvn", "i32", "$dst, $SIMM", "",
+ [(set DPR:$dst, (v2i32 (NEONvmvnImm timm:$SIMM)))]>;
+def VMVNv4i32 : N1ModImm<1, 0b000, {?,?,?,?}, 0, 1, 1, 1, (outs QPR:$dst),
+ (ins nModImm:$SIMM), IIC_VMOVImm,
+ "vmvn", "i32", "$dst, $SIMM", "",
+ [(set QPR:$dst, (v4i32 (NEONvmvnImm timm:$SIMM)))]>;
+}
+
// VMVN : Vector Bitwise NOT
def VMVNd : N2VX<0b11, 0b11, 0b00, 0b00, 0b01011, 0, 0,
(outs DPR:$dst), (ins DPR:$src), IIC_VSUBiD,