[llvm-commits] [llvm] r107730 - in /llvm/trunk/lib/Target/ARM: ARMISelDAGToDAG.cpp ARMInstrNEON.td
Bob Wilson
bob.wilson at apple.com
Tue Jul 6 16:36:25 PDT 2010
Author: bwilson
Date: Tue Jul 6 18:36:25 2010
New Revision: 107730
URL: http://llvm.org/viewvc/llvm-project?rev=107730&view=rev
Log:
Use REG_SEQUENCE nodes to make the table registers for VTBL instructions be
allocated to consecutive registers.
Modified:
llvm/trunk/lib/Target/ARM/ARMISelDAGToDAG.cpp
llvm/trunk/lib/Target/ARM/ARMInstrNEON.td
Modified: llvm/trunk/lib/Target/ARM/ARMISelDAGToDAG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/ARM/ARMISelDAGToDAG.cpp?rev=107730&r1=107729&r2=107730&view=diff
==============================================================================
--- llvm/trunk/lib/Target/ARM/ARMISelDAGToDAG.cpp (original)
+++ llvm/trunk/lib/Target/ARM/ARMISelDAGToDAG.cpp Tue Jul 6 18:36:25 2010
@@ -143,6 +143,11 @@
unsigned *DOpcodes, unsigned *QOpcodes0,
unsigned *QOpcodes1);
+ /// SelectVTBL - Select NEON VTBL intrinsics. NumVecs should be 2, 3 or 4.
+ /// These are custom-selected so that a REG_SEQUENCE can be generated to
+ /// force the table registers to be consecutive.
+ SDNode *SelectVTBL(SDNode *N, unsigned NumVecs, unsigned Opc);
+
/// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
@@ -1197,7 +1202,7 @@
SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
unsigned *DOpcodes, unsigned *QOpcodes0,
unsigned *QOpcodes1) {
- assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
+ assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
DebugLoc dl = N->getDebugLoc();
SDValue MemAddr, Align;
@@ -1524,6 +1529,42 @@
return NULL;
}
+SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, unsigned NumVecs, unsigned Opc) {
+ assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range");
+ DebugLoc dl = N->getDebugLoc();
+ EVT VT = N->getValueType(0);
+
+ // Form a REG_SEQUENCE to force register allocation.
+ SDValue RegSeq;
+ SDValue V0 = N->getOperand(1);
+ SDValue V1 = N->getOperand(2);
+ if (NumVecs == 2)
+ RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
+ else {
+ SDValue V2 = N->getOperand(3);
+ // If it's a vtbl3, form a quad D-register and leave the last part as
+ // an undef.
+ SDValue V3 = (NumVecs == 3)
+ ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
+ : N->getOperand(4);
+ RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
+ }
+
+ // Now extract the D registers back out.
+ SmallVector<SDValue, 5> Ops;
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
+ if (NumVecs > 2)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT, RegSeq));
+ if (NumVecs > 3)
+ Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT, RegSeq));
+
+ Ops.push_back(N->getOperand(NumVecs+1));
+ Ops.push_back(getAL(CurDAG)); // predicate
+ Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register
+ return CurDAG->getMachineNode(Opc, dl, VT, Ops.data(), NumVecs+3);
+}
+
SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
bool isSigned) {
if (!Subtarget->hasV6T2Ops())
@@ -2281,6 +2322,22 @@
break;
}
+ case ISD::INTRINSIC_WO_CHAIN: {
+ unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
+ switch (IntNo) {
+ default:
+ break;
+
+ case Intrinsic::arm_neon_vtbl2:
+ return SelectVTBL(N, 2, ARM::VTBL2);
+ case Intrinsic::arm_neon_vtbl3:
+ return SelectVTBL(N, 3, ARM::VTBL3);
+ case Intrinsic::arm_neon_vtbl4:
+ return SelectVTBL(N, 4, ARM::VTBL4);
+ }
+ break;
+ }
+
case ISD::CONCAT_VECTORS:
return SelectConcatVector(N);
}
Modified: llvm/trunk/lib/Target/ARM/ARMInstrNEON.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/ARM/ARMInstrNEON.td?rev=107730&r1=107729&r2=107730&view=diff
==============================================================================
--- llvm/trunk/lib/Target/ARM/ARMInstrNEON.td (original)
+++ llvm/trunk/lib/Target/ARM/ARMInstrNEON.td Tue Jul 6 18:36:25 2010
@@ -3291,22 +3291,16 @@
def VTBL2
: N3V<1,1,0b11,0b1001,0,0, (outs DPR:$dst),
(ins DPR:$tbl1, DPR:$tbl2, DPR:$src), NVTBLFrm, IIC_VTB2,
- "vtbl", "8", "$dst, \\{$tbl1, $tbl2\\}, $src", "",
- [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl2
- DPR:$tbl1, DPR:$tbl2, DPR:$src)))]>;
+ "vtbl", "8", "$dst, \\{$tbl1, $tbl2\\}, $src", "", []>;
def VTBL3
: N3V<1,1,0b11,0b1010,0,0, (outs DPR:$dst),
(ins DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$src), NVTBLFrm, IIC_VTB3,
- "vtbl", "8", "$dst, \\{$tbl1, $tbl2, $tbl3\\}, $src", "",
- [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl3
- DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$src)))]>;
+ "vtbl", "8", "$dst, \\{$tbl1, $tbl2, $tbl3\\}, $src", "", []>;
def VTBL4
: N3V<1,1,0b11,0b1011,0,0, (outs DPR:$dst),
(ins DPR:$tbl1, DPR:$tbl2, DPR:$tbl3, DPR:$tbl4, DPR:$src),
NVTBLFrm, IIC_VTB4,
- "vtbl", "8", "$dst, \\{$tbl1, $tbl2, $tbl3, $tbl4\\}, $src", "",
- [(set DPR:$dst, (v8i8 (int_arm_neon_vtbl4 DPR:$tbl1, DPR:$tbl2,
- DPR:$tbl3, DPR:$tbl4, DPR:$src)))]>;
+ "vtbl", "8", "$dst, \\{$tbl1, $tbl2, $tbl3, $tbl4\\}, $src", "", []>;
} // hasExtraSrcRegAllocReq = 1
// VTBX : Vector Table Extension
More information about the llvm-commits
mailing list