Add direct moves to/from VSR and exploit them for FP/INT conversions
This patch corresponds to review:
http://reviews.llvm.org/D8928
It adds direct move instructions to/from VSX registers to GPR's. These are
exploited for FP <-> INT conversions.
llvm-svn: 234682
diff --git a/llvm/lib/Target/PowerPC/PPC.td b/llvm/lib/Target/PowerPC/PPC.td
index d97a5a1..e8c8239 100644
--- a/llvm/lib/Target/PowerPC/PPC.td
+++ b/llvm/lib/Target/PowerPC/PPC.td
@@ -122,6 +122,10 @@
def FeatureP8Vector : SubtargetFeature<"power8-vector", "HasP8Vector", "true",
"Enable POWER8 vector instructions",
[FeatureVSX, FeatureP8Altivec]>;
+def FeatureDirectMove :
+ SubtargetFeature<"direct-move", "HasDirectMove", "true",
+ "Enable Power8 direct move instructions",
+ [FeatureVSX]>;
def FeaturePartwordAtomic : SubtargetFeature<"partword-atomics",
"HasPartwordAtomics", "true",
"Enable l[bh]arx and st[bh]cx.">;
@@ -164,7 +168,7 @@
DeprecatedMFTB, DeprecatedDST];
list<SubtargetFeature> Power8SpecificFeatures =
[DirectivePwr8, FeatureP8Altivec, FeatureP8Vector, FeatureP8Crypto,
- FeatureHTM, FeatureICBT];
+ FeatureHTM, FeatureDirectMove, FeatureICBT];
list<SubtargetFeature> Power8FeatureList =
!listconcat(Power7FeatureList, Power8SpecificFeatures);
}
diff --git a/llvm/lib/Target/PowerPC/PPCFastISel.cpp b/llvm/lib/Target/PowerPC/PPCFastISel.cpp
index 8d3f308..002616b 100644
--- a/llvm/lib/Target/PowerPC/PPCFastISel.cpp
+++ b/llvm/lib/Target/PowerPC/PPCFastISel.cpp
@@ -958,6 +958,8 @@
}
// Attempt to fast-select an integer-to-floating-point conversion.
+// FIXME: Once fast-isel has better support for VSX, conversions using
+// direct moves should be implemented.
bool PPCFastISel::SelectIToFP(const Instruction *I, bool IsSigned) {
MVT DstVT;
Type *DstTy = I->getType();
@@ -1065,6 +1067,8 @@
}
// Attempt to fast-select a floating-point-to-integer conversion.
+// FIXME: Once fast-isel has better support for VSX, conversions using
+// direct moves should be implemented.
bool PPCFastISel::SelectFPToI(const Instruction *I, bool IsSigned) {
MVT DstVT, SrcVT;
Type *DstTy = I->getType();
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 2df4c48..b4b65ca 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -996,6 +996,9 @@
case PPCISD::EH_SJLJ_SETJMP: return "PPCISD::EH_SJLJ_SETJMP";
case PPCISD::EH_SJLJ_LONGJMP: return "PPCISD::EH_SJLJ_LONGJMP";
case PPCISD::MFOCRF: return "PPCISD::MFOCRF";
+ case PPCISD::MFVSR: return "PPCISD::MFVSR";
+ case PPCISD::MTVSRA: return "PPCISD::MTVSRA";
+ case PPCISD::MTVSRZ: return "PPCISD::MTVSRZ";
case PPCISD::VCMP: return "PPCISD::VCMP";
case PPCISD::VCMPo: return "PPCISD::VCMPo";
case PPCISD::LBRX: return "PPCISD::LBRX";
@@ -5911,8 +5914,46 @@
RLI.MPI = MPI;
}
+/// \brief Custom lowers floating point to integer conversions to use
+/// the direct move instructions available in ISA 2.07 to avoid the
+/// need for load/store combinations.
+SDValue PPCTargetLowering::LowerFP_TO_INTDirectMove(SDValue Op,
+ SelectionDAG &DAG,
+ SDLoc dl) const {
+ assert(Op.getOperand(0).getValueType().isFloatingPoint());
+ SDValue Src = Op.getOperand(0);
+
+ if (Src.getValueType() == MVT::f32)
+ Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
+
+ SDValue Tmp;
+ switch (Op.getSimpleValueType().SimpleTy) {
+ default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
+ case MVT::i32:
+ Tmp = DAG.getNode(
+ Op.getOpcode() == ISD::FP_TO_SINT
+ ? PPCISD::FCTIWZ
+ : (Subtarget.hasFPCVT() ? PPCISD::FCTIWUZ : PPCISD::FCTIDZ),
+ dl, MVT::f64, Src);
+ Tmp = DAG.getNode(PPCISD::MFVSR, dl, MVT::i32, Tmp);
+ break;
+ case MVT::i64:
+ assert((Op.getOpcode() == ISD::FP_TO_SINT || Subtarget.hasFPCVT()) &&
+ "i64 FP_TO_UINT is supported only with FPCVT");
+ Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIDZ :
+ PPCISD::FCTIDUZ,
+ dl, MVT::f64, Src);
+ Tmp = DAG.getNode(PPCISD::MFVSR, dl, MVT::i64, Tmp);
+ break;
+ }
+ return Tmp;
+}
+
SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
SDLoc dl) const {
+ if (Subtarget.hasDirectMove() && Subtarget.isPPC64())
+ return LowerFP_TO_INTDirectMove(Op, DAG, dl);
+
ReuseLoadInfo RLI;
LowerFP_TO_INTForReuse(Op, RLI, DAG, dl);
@@ -5990,6 +6031,38 @@
DAG.UpdateNodeOperands(TF.getNode(), ResChain, NewResChain);
}
+/// \brief Custom lowers integer to floating point conversions to use
+/// the direct move instructions available in ISA 2.07 to avoid the
+/// need for load/store combinations.
+SDValue PPCTargetLowering::LowerINT_TO_FPDirectMove(SDValue Op,
+ SelectionDAG &DAG,
+ SDLoc dl) const {
+ assert((Op.getValueType() == MVT::f32 ||
+ Op.getValueType() == MVT::f64) &&
+ "Invalid floating point type as target of conversion");
+ assert(Subtarget.hasFPCVT() &&
+ "Int to FP conversions with direct moves require FPCVT");
+ SDValue FP;
+ SDValue Src = Op.getOperand(0);
+ bool SinglePrec = Op.getValueType() == MVT::f32;
+ bool WordInt = Src.getSimpleValueType().SimpleTy == MVT::i32;
+ bool Signed = Op.getOpcode() == ISD::SINT_TO_FP;
+ unsigned ConvOp = Signed ? (SinglePrec ? PPCISD::FCFIDS : PPCISD::FCFID) :
+ (SinglePrec ? PPCISD::FCFIDUS : PPCISD::FCFIDU);
+
+ if (WordInt) {
+ FP = DAG.getNode(Signed ? PPCISD::MTVSRA : PPCISD::MTVSRZ,
+ dl, MVT::f64, Src);
+ FP = DAG.getNode(ConvOp, dl, SinglePrec ? MVT::f32 : MVT::f64, FP);
+ }
+ else {
+ FP = DAG.getNode(PPCISD::MTVSRA, dl, MVT::f64, Src);
+ FP = DAG.getNode(ConvOp, dl, SinglePrec ? MVT::f32 : MVT::f64, FP);
+ }
+
+ return FP;
+}
+
SDValue PPCTargetLowering::LowerINT_TO_FP(SDValue Op,
SelectionDAG &DAG) const {
SDLoc dl(Op);
@@ -6025,6 +6098,11 @@
DAG.getConstantFP(1.0, Op.getValueType()),
DAG.getConstantFP(0.0, Op.getValueType()));
+ // If we have direct moves, we can do all the conversion, skip the store/load
+ // however, without FPCVT we can't do most conversions.
+ if (Subtarget.hasDirectMove() && Subtarget.isPPC64() && Subtarget.hasFPCVT())
+ return LowerINT_TO_FPDirectMove(Op, DAG, dl);
+
assert((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) &&
"UINT_TO_FP is supported only with FPCVT");
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.h b/llvm/lib/Target/PowerPC/PPCISelLowering.h
index 695fc3f..7e2ebd4 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.h
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.h
@@ -119,6 +119,15 @@
/// resultant GPR. Bits corresponding to other CR regs are undefined.
MFOCRF,
+ /// Direct move from a VSX register to a GPR
+ MFVSR,
+
+ /// Direct move from a GPR to a VSX register (algebraic)
+ MTVSRA,
+
+ /// Direct move from a GPR to a VSX register (zero)
+ MTVSRZ,
+
// FIXME: Remove these once the ANDI glue bug is fixed:
/// i1 = ANDIo_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
/// eq or gt bit of CR0 after executing andi. x, 1. This is used to
@@ -645,6 +654,10 @@
void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
SelectionDAG &DAG, SDLoc dl) const;
+ SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
+ SDLoc dl) const;
+ SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
+ SDLoc dl) const;
SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
diff --git a/llvm/lib/Target/PowerPC/PPCInstrFormats.td b/llvm/lib/Target/PowerPC/PPCInstrFormats.td
index b7a7a1f..43c2158 100644
--- a/llvm/lib/Target/PowerPC/PPCInstrFormats.td
+++ b/llvm/lib/Target/PowerPC/PPCInstrFormats.td
@@ -764,6 +764,12 @@
let Inst{31} = XT{5};
}
+class XX1_RS6_RD5_XO<bits<6> opcode, bits<10> xo, dag OOL, dag IOL,
+ string asmstr, InstrItinClass itin, list<dag> pattern>
+ : XX1Form<opcode, xo, OOL, IOL, asmstr, itin, pattern> {
+ let B = 0;
+}
+
class XX2Form<bits<6> opcode, bits<9> xo, dag OOL, dag IOL, string asmstr,
InstrItinClass itin, list<dag> pattern>
: I<opcode, OOL, IOL, asmstr, itin> {
diff --git a/llvm/lib/Target/PowerPC/PPCInstrVSX.td b/llvm/lib/Target/PowerPC/PPCInstrVSX.td
index ec04da4..a98e58f 100644
--- a/llvm/lib/Target/PowerPC/PPCInstrVSX.td
+++ b/llvm/lib/Target/PowerPC/PPCInstrVSX.td
@@ -41,6 +41,9 @@
def PPCstxvd2x : SDNode<"PPCISD::STXVD2X", SDT_PPCstxvd2x,
[SDNPHasChain, SDNPMayStore]>;
def PPCxxswapd : SDNode<"PPCISD::XXSWAPD", SDT_PPCxxswapd, [SDNPHasChain]>;
+def PPCmfvsr : SDNode<"PPCISD::MFVSR", SDTUnaryOp, []>;
+def PPCmtvsra : SDNode<"PPCISD::MTVSRA", SDTUnaryOp, []>;
+def PPCmtvsrz : SDNode<"PPCISD::MTVSRZ", SDTUnaryOp, []>;
multiclass XX3Form_Rcr<bits<6> opcode, bits<7> xo, dag OOL, dag IOL,
string asmbase, string asmstr, InstrItinClass itin,
@@ -946,6 +949,7 @@
when the elements are larger than i32.
*/
def HasP8Vector : Predicate<"PPCSubTarget->hasP8Vector()">;
+def HasDirectMove : Predicate<"PPCSubTarget->hasDirectMove()">;
let Predicates = [HasP8Vector] in {
let AddedComplexity = 400 in { // Prefer VSX patterns over non-VSX patterns.
let isCommutable = 1 in {
@@ -965,3 +969,24 @@
[(set v4i32:$XT, (or v4i32:$XA, (vnot_ppc v4i32:$XB)))]>;
} // AddedComplexity = 500
} // HasP8Vector
+
+let Predicates = [HasDirectMove, HasVSX] in {
+// VSX direct move instructions
+def MFVSRD : XX1_RS6_RD5_XO<31, 51, (outs g8rc:$rA), (ins vsfrc:$XT),
+ "mfvsrd $rA, $XT", IIC_VecGeneral,
+ [(set i64:$rA, (PPCmfvsr f64:$XT))]>,
+ Requires<[In64BitMode]>;
+def MFVSRWZ : XX1_RS6_RD5_XO<31, 115, (outs gprc:$rA), (ins vsfrc:$XT),
+ "mfvsrwz $rA, $XT", IIC_VecGeneral,
+ [(set i32:$rA, (PPCmfvsr f64:$XT))]>;
+def MTVSRD : XX1_RS6_RD5_XO<31, 179, (outs vsfrc:$XT), (ins g8rc:$rA),
+ "mtvsrd $XT, $rA", IIC_VecGeneral,
+ [(set f64:$XT, (PPCmtvsra i64:$rA))]>,
+ Requires<[In64BitMode]>;
+def MTVSRWA : XX1_RS6_RD5_XO<31, 211, (outs vsfrc:$XT), (ins gprc:$rA),
+ "mtvsrwa $XT, $rA", IIC_VecGeneral,
+ [(set f64:$XT, (PPCmtvsra i32:$rA))]>;
+def MTVSRWZ : XX1_RS6_RD5_XO<31, 243, (outs vsfrc:$XT), (ins gprc:$rA),
+ "mtvsrwz $XT, $rA", IIC_VecGeneral,
+ [(set f64:$XT, (PPCmtvsrz i32:$rA))]>;
+} // HasDirectMove, HasVSX
diff --git a/llvm/lib/Target/PowerPC/PPCSubtarget.cpp b/llvm/lib/Target/PowerPC/PPCSubtarget.cpp
index 71365af..f313b0a 100644
--- a/llvm/lib/Target/PowerPC/PPCSubtarget.cpp
+++ b/llvm/lib/Target/PowerPC/PPCSubtarget.cpp
@@ -97,6 +97,7 @@
HasICBT = false;
HasInvariantFunctionDescriptors = false;
HasPartwordAtomics = false;
+ HasDirectMove = false;
IsQPXStackUnaligned = false;
HasHTM = false;
}
diff --git a/llvm/lib/Target/PowerPC/PPCSubtarget.h b/llvm/lib/Target/PowerPC/PPCSubtarget.h
index 6941be6..8d95508 100644
--- a/llvm/lib/Target/PowerPC/PPCSubtarget.h
+++ b/llvm/lib/Target/PowerPC/PPCSubtarget.h
@@ -117,6 +117,7 @@
bool HasICBT;
bool HasInvariantFunctionDescriptors;
bool HasPartwordAtomics;
+ bool HasDirectMove;
bool HasHTM;
/// When targeting QPX running a stock PPC64 Linux kernel where the stack
@@ -243,6 +244,7 @@
return HasInvariantFunctionDescriptors;
}
bool hasPartwordAtomics() const { return HasPartwordAtomics; }
+ bool hasDirectMove() const { return HasDirectMove; }
bool isQPXStackUnaligned() const { return IsQPXStackUnaligned; }
unsigned getPlatformStackAlignment() const {