[Hexagon] Use integer instructions for floating point immediates
Floating point instructions use general purpose registers, so the few
instructions that can put floating point immediates into registers are,
in fact, integer instruction. Use them explicitly instead of having
pseudo-instructions specifically for dealing with floating point values.
Simplify the constant loading instructions (from sdata) to have only two:
one for 32-bit values and one for 64-bit values: CONST32 and CONST64.
llvm-svn: 278244
diff --git a/llvm/lib/Target/Hexagon/AsmParser/HexagonAsmParser.cpp b/llvm/lib/Target/Hexagon/AsmParser/HexagonAsmParser.cpp
index 4aa108f..ba223a7 100644
--- a/llvm/lib/Target/Hexagon/AsmParser/HexagonAsmParser.cpp
+++ b/llvm/lib/Target/Hexagon/AsmParser/HexagonAsmParser.cpp
@@ -1658,13 +1658,9 @@
// Translate a "$Rx = CONST32(#imm)" to "$Rx = memw(gp+#LABEL) "
case Hexagon::CONST32:
- case Hexagon::CONST32_Float_Real:
- case Hexagon::CONST32_Int_Real:
is32bit = true;
// Translate a "$Rx:y = CONST64(#imm)" to "$Rx:y = memd(gp+#LABEL) "
- case Hexagon::CONST64_Float_Real:
- case Hexagon::CONST64_Int_Real:
-
+ case Hexagon::CONST64:
// FIXME: need better way to detect AsmStreamer (upstream removed getKind())
if (!Parser.getStreamer().hasRawTextSupport()) {
MCELFStreamer *MES = static_cast<MCELFStreamer *>(&Parser.getStreamer());
diff --git a/llvm/lib/Target/Hexagon/HexagonAsmPrinter.cpp b/llvm/lib/Target/Hexagon/HexagonAsmPrinter.cpp
index 8a3bd22..c95c893 100644
--- a/llvm/lib/Target/Hexagon/HexagonAsmPrinter.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonAsmPrinter.cpp
@@ -283,8 +283,7 @@
}
// "$dst = CONST64(#$src1)",
- case Hexagon::CONST64_Float_Real:
- case Hexagon::CONST64_Int_Real:
+ case Hexagon::CONST64:
if (!OutStreamer->hasRawTextSupport()) {
const MCOperand &Imm = MappedInst.getOperand(1);
MCSectionSubPair Current = OutStreamer->getCurrentSection();
@@ -303,8 +302,6 @@
}
break;
case Hexagon::CONST32:
- case Hexagon::CONST32_Float_Real:
- case Hexagon::CONST32_Int_Real:
if (!OutStreamer->hasRawTextSupport()) {
MCOperand &Imm = MappedInst.getOperand(1);
MCSectionSubPair Current = OutStreamer->getCurrentSection();
@@ -465,21 +462,6 @@
MappedInst = TmpInst;
return;
}
- case Hexagon::TFRI_f:
- MappedInst.setOpcode(Hexagon::A2_tfrsi);
- return;
- case Hexagon::TFRI_cPt_f:
- MappedInst.setOpcode(Hexagon::C2_cmoveit);
- return;
- case Hexagon::TFRI_cNotPt_f:
- MappedInst.setOpcode(Hexagon::C2_cmoveif);
- return;
- case Hexagon::MUX_ri_f:
- MappedInst.setOpcode(Hexagon::C2_muxri);
- return;
- case Hexagon::MUX_ir_f:
- MappedInst.setOpcode(Hexagon::C2_muxir);
- return;
// Translate a "$Rdd = #imm" to "$Rdd = combine(#[-1,0], #imm)"
case Hexagon::A2_tfrpi: {
diff --git a/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp b/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
index d1c75b3..6de62f1 100644
--- a/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
@@ -1353,8 +1353,8 @@
case Hexagon::A2_tfrpi:
case Hexagon::TFR_PdTrue:
case Hexagon::TFR_PdFalse:
- case Hexagon::CONST32_Int_Real:
- case Hexagon::CONST64_Int_Real:
+ case Hexagon::CONST32:
+ case Hexagon::CONST64:
return true;
}
return false;
@@ -1389,7 +1389,7 @@
return Reg;
}
- BuildMI(B, At, DL, HII.get(Hexagon::CONST64_Int_Real), Reg)
+ BuildMI(B, At, DL, HII.get(Hexagon::CONST64), Reg)
.addImm(C);
return Reg;
}
diff --git a/llvm/lib/Target/Hexagon/HexagonBitTracker.cpp b/llvm/lib/Target/Hexagon/HexagonBitTracker.cpp
index 78e1dce..25a8da7 100644
--- a/llvm/lib/Target/Hexagon/HexagonBitTracker.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonBitTracker.cpp
@@ -146,8 +146,7 @@
// These instructions may be marked as mayLoad, but they are generating
// immediate values, so skip them.
case CONST32:
- case CONST32_Int_Real:
- case CONST64_Int_Real:
+ case CONST64:
break;
default:
return evaluateLoad(MI, Inputs, Outputs);
@@ -257,8 +256,7 @@
case A2_tfrsi:
case A2_tfrpi:
case CONST32:
- case CONST32_Int_Real:
- case CONST64_Int_Real:
+ case CONST64:
return rr0(eIMM(im(1), W0), Outputs);
case TFR_PdFalse:
return rr0(RegisterCell(W0).fill(0, W0, BT::BitValue::Zero), Outputs);
diff --git a/llvm/lib/Target/Hexagon/HexagonConstPropagation.cpp b/llvm/lib/Target/Hexagon/HexagonConstPropagation.cpp
index 172df0c..c4a0a62 100644
--- a/llvm/lib/Target/Hexagon/HexagonConstPropagation.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonConstPropagation.cpp
@@ -1971,14 +1971,13 @@
default:
return false;
case Hexagon::A2_tfrsi:
- case Hexagon::CONST32:
case Hexagon::A2_tfrpi:
- case Hexagon::CONST32_Int_Real:
- case Hexagon::CONST64_Int_Real:
+ case Hexagon::CONST32:
+ case Hexagon::CONST64:
{
const MachineOperand &VO = MI.getOperand(1);
- // The operand of CONST32_Int_Real can be a blockaddress, e.g.
- // %vreg0<def> = CONST32_Int_Real <blockaddress(@eat, %L)>
+ // The operand of CONST32 can be a blockaddress, e.g.
+ // %vreg0<def> = CONST32 <blockaddress(@eat, %L)>
// Do this check for all instructions for safety.
if (!VO.isImm())
return false;
@@ -2326,10 +2325,9 @@
default:
break;
case Hexagon::A2_tfrsi:
- case Hexagon::CONST32:
case Hexagon::A2_tfrpi:
- case Hexagon::CONST32_Int_Real:
- case Hexagon::CONST64_Int_Real:
+ case Hexagon::CONST32:
+ case Hexagon::CONST64:
case Hexagon::TFR_PdTrue:
case Hexagon::TFR_PdFalse:
return false;
@@ -2921,7 +2919,7 @@
.addImm(Hi)
.addImm(Lo);
} else {
- NewD = &HII.get(Hexagon::CONST64_Int_Real);
+ NewD = &HII.get(Hexagon::CONST64);
NewMI = BuildMI(B, At, DL, *NewD, NewR)
.addImm(V);
}
diff --git a/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp b/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
index 6409852..a472443 100644
--- a/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
@@ -628,8 +628,7 @@
int64_t V = HiOperand.getImm();
V = (V << 32) | (0x0ffffffffLL & LoOperand.getImm());
- BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::CONST64_Int_Real),
- DoubleDestReg)
+ BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::CONST64), DoubleDestReg)
.addImm(V);
}
diff --git a/llvm/lib/Target/Hexagon/HexagonFrameLowering.cpp b/llvm/lib/Target/Hexagon/HexagonFrameLowering.cpp
index 0577344..2530b5f 100644
--- a/llvm/lib/Target/Hexagon/HexagonFrameLowering.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonFrameLowering.cpp
@@ -563,7 +563,7 @@
// Subtract offset from frame pointer.
// We use a caller-saved non-parameter register for that.
unsigned CallerSavedReg = HRI.getFirstCallerSavedNonParamReg();
- BuildMI(MBB, InsertPt, dl, HII.get(Hexagon::CONST32_Int_Real),
+ BuildMI(MBB, InsertPt, dl, HII.get(Hexagon::CONST32),
CallerSavedReg).addImm(NumBytes);
BuildMI(MBB, InsertPt, dl, HII.get(Hexagon::A2_sub), SP)
.addReg(SP)
diff --git a/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp b/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp
index 608f2918..02e08ba 100644
--- a/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp
@@ -1490,8 +1490,8 @@
case TargetOpcode::COPY:
case Hexagon::A2_tfrsi:
case Hexagon::A2_tfrpi:
- case Hexagon::CONST32_Int_Real:
- case Hexagon::CONST64_Int_Real: {
+ case Hexagon::CONST32:
+ case Hexagon::CONST64: {
// Call recursively to avoid an extra check whether operand(1) is
// indeed an immediate (it could be a global address, for example),
// plus we can handle COPY at the same time.
diff --git a/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp b/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp
index ad4aece..d6faa47 100644
--- a/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp
@@ -823,7 +823,7 @@
Bit <<= ES;
}
SDValue Ones = CurDAG->getTargetConstant(MV, dl, MVT::i64);
- SDNode *OnesReg = CurDAG->getMachineNode(Hexagon::CONST64_Int_Real, dl,
+ SDNode *OnesReg = CurDAG->getMachineNode(Hexagon::CONST64, dl,
MVT::i64, Ones);
if (ExVT.getSizeInBits() == 32) {
SDNode *And = CurDAG->getMachineNode(Hexagon::A2_andp, dl, MVT::i64,
@@ -918,19 +918,15 @@
void HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) {
SDLoc dl(N);
ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N);
- const APFloat &APF = CN->getValueAPF();
+ APInt A = CN->getValueAPF().bitcastToAPInt();
if (N->getValueType(0) == MVT::f32) {
- ReplaceNode(
- N, CurDAG->getMachineNode(Hexagon::TFRI_f, dl, MVT::f32,
- CurDAG->getTargetConstantFP(
- APF.convertToFloat(), dl, MVT::f32)));
+ SDValue V = CurDAG->getTargetConstant(A.getZExtValue(), dl, MVT::i32);
+ ReplaceNode(N, CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl, MVT::f32, V));
return;
}
if (N->getValueType(0) == MVT::f64) {
- ReplaceNode(
- N, CurDAG->getMachineNode(Hexagon::CONST64_Float_Real, dl, MVT::f64,
- CurDAG->getTargetConstantFP(
- APF.convertToDouble(), dl, MVT::f64)));
+ SDValue V = CurDAG->getTargetConstant(A.getZExtValue(), dl, MVT::i64);
+ ReplaceNode(N, CurDAG->getMachineNode(Hexagon::CONST64, dl, MVT::f64, V));
return;
}
diff --git a/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp b/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp
index 92137ec..b1e2375 100644
--- a/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp
@@ -1301,20 +1301,6 @@
case Hexagon::TCRETURNr:
MI.setDesc(get(Hexagon::J2_jumpr));
return true;
- case Hexagon::TFRI_f:
- case Hexagon::TFRI_cPt_f:
- case Hexagon::TFRI_cNotPt_f: {
- unsigned Opx = (Opc == Hexagon::TFRI_f) ? 1 : 2;
- APFloat FVal = MI.getOperand(Opx).getFPImm()->getValueAPF();
- APInt IVal = FVal.bitcastToAPInt();
- MI.RemoveOperand(Opx);
- unsigned NewOpc = (Opc == Hexagon::TFRI_f) ? Hexagon::A2_tfrsi :
- (Opc == Hexagon::TFRI_cPt_f) ? Hexagon::C2_cmoveit :
- Hexagon::C2_cmoveif;
- MI.setDesc(get(NewOpc));
- MI.addOperand(MachineOperand::CreateImm(IVal.getZExtValue()));
- return true;
- }
}
return false;
@@ -1942,7 +1928,7 @@
// object we are going to end up with here for now.
// In the future we probably should add isSymbol(), etc.
if (MO.isGlobal() || MO.isSymbol() || MO.isBlockAddress() ||
- MO.isJTI() || MO.isCPI())
+ MO.isJTI() || MO.isCPI() || MO.isFPImm())
return true;
// If the extendable operand is not 'Immediate' type, the instruction should
@@ -3424,14 +3410,6 @@
if (CondOpcode >= 0) // Valid Conditional opcode/instruction
return CondOpcode;
- // This switch case will be removed once all the instructions have been
- // modified to use relation maps.
- switch(Opc) {
- case Hexagon::TFRI_f:
- return !invertPredicate ? Hexagon::TFRI_cPt_f :
- Hexagon::TFRI_cNotPt_f;
- }
-
llvm_unreachable("Unexpected predicable instruction");
}
diff --git a/llvm/lib/Target/Hexagon/HexagonInstrInfo.td b/llvm/lib/Target/Hexagon/HexagonInstrInfo.td
index fda1c98..094cec6 100644
--- a/llvm/lib/Target/Hexagon/HexagonInstrInfo.td
+++ b/llvm/lib/Target/Hexagon/HexagonInstrInfo.td
@@ -4799,29 +4799,18 @@
"$dst.l = #LO($global@GOTREL)",
[]>;
-// This pattern is incorrect. When we add small data, we should change
-// this pattern to use memw(#foo).
-// This is for sdata.
-let isMoveImm = 1, isAsmParserOnly = 1 in
-def CONST32 : CONSTLDInst<(outs IntRegs:$dst), (ins globaladdress:$global),
- "$dst = CONST32(#$global)",
- [(set (i32 IntRegs:$dst),
- (load (HexagonCONST32 tglobaltlsaddr:$global)))]>;
+let isReMaterializable = 1, isMoveImm = 1 in
+def CONST32 : CONSTLDInst<(outs IntRegs:$Rd), (ins i32imm:$v),
+ "$Rd = CONST32(#$v)", [(set I32:$Rd, imm:$v)]>;
-let isReMaterializable = 1, isMoveImm = 1, isAsmParserOnly = 1 in
-def CONST32_Int_Real : CONSTLDInst<(outs IntRegs:$dst), (ins i32imm:$global),
- "$dst = CONST32(#$global)",
- [(set (i32 IntRegs:$dst), imm:$global) ]>;
+let isReMaterializable = 1, isMoveImm = 1 in
+def CONST64 : CONSTLDInst<(outs DoubleRegs:$Rd), (ins i64imm:$v),
+ "$Rd = CONST64(#$v)", [(set I64:$Rd, imm:$v)]>;
-// Map TLS addressses to a CONST32 instruction
+// Map TLS addressses to A2_tfrsi.
def: Pat<(HexagonCONST32 tglobaltlsaddr:$addr), (A2_tfrsi s16Ext:$addr)>;
def: Pat<(HexagonCONST32 bbl:$label), (A2_tfrsi s16Ext:$label)>;
-let isReMaterializable = 1, isMoveImm = 1, isAsmParserOnly = 1 in
-def CONST64_Int_Real : CONSTLDInst<(outs DoubleRegs:$dst), (ins i64imm:$global),
- "$dst = CONST64(#$global)",
- [(set (i64 DoubleRegs:$dst), imm:$global)]>;
-
let hasSideEffects = 0, isReMaterializable = 1, isPseudo = 1,
isCodeGenOnly = 1 in
def TFR_PdTrue : SInst<(outs PredRegs:$dst), (ins), "",
diff --git a/llvm/lib/Target/Hexagon/HexagonInstrInfoV5.td b/llvm/lib/Target/Hexagon/HexagonInstrInfoV5.td
index 8088f4d..3d96864 100644
--- a/llvm/lib/Target/Hexagon/HexagonInstrInfoV5.td
+++ b/llvm/lib/Target/Hexagon/HexagonInstrInfoV5.td
@@ -11,6 +11,15 @@
//
//===----------------------------------------------------------------------===//
+def f32ImmPred : PatLeaf<(f32 fpimm:$F)>;
+def f64ImmPred : PatLeaf<(f64 fpimm:$F)>;
+
+def ftoi : SDNodeXForm<fpimm, [{
+ APInt I = N->getValueAPF().bitcastToAPInt();
+ return CurDAG->getTargetConstant(I.getZExtValue(), SDLoc(N),
+ MVT::getIntegerVT(I.getBitWidth()));
+}]>;
+
//===----------------------------------------------------------------------===//
// XTYPE/MPY
//===----------------------------------------------------------------------===//
@@ -67,46 +76,6 @@
let Inst{20,13,7,4} = 0b1111;
}
-let isReMaterializable = 1, isMoveImm = 1, isAsmParserOnly = 1 in
-def CONST64_Float_Real : LDInst<(outs DoubleRegs:$dst), (ins f64imm:$src1),
- "$dst = CONST64(#$src1)",
- [(set F64:$dst, fpimm:$src1)]>,
- Requires<[HasV5T]>;
-
-let isReMaterializable = 1, isMoveImm = 1, isAsmParserOnly = 1 in
-def CONST32_Float_Real : LDInst<(outs IntRegs:$dst), (ins f32imm:$src1),
- "$dst = CONST32(#$src1)",
- [(set F32:$dst, fpimm:$src1)]>,
- Requires<[HasV5T]>;
-
-// Transfer immediate float.
-// Only works with single precision fp value.
-// For double precision, use CONST64_float_real, as 64bit transfer
-// can only hold 40-bit values - 32 from const ext + 8 bit immediate.
-// Make sure that complexity is more than the CONST32 pattern in
-// HexagonInstrInfo.td patterns.
-let isExtended = 1, opExtendable = 1, isMoveImm = 1, isReMaterializable = 1,
- isPredicable = 1, AddedComplexity = 30, validSubTargets = HasV5SubT,
- isCodeGenOnly = 1, isPseudo = 1 in
-def TFRI_f : ALU32_ri<(outs IntRegs:$dst), (ins f32Ext:$src1),
- "$dst = #$src1",
- [(set F32:$dst, fpimm:$src1)]>,
- Requires<[HasV5T]>;
-
-let isExtended = 1, opExtendable = 2, isPredicated = 1, hasSideEffects = 0,
- validSubTargets = HasV5SubT, isCodeGenOnly = 1, isPseudo = 1 in
-def TFRI_cPt_f : ALU32_ri<(outs IntRegs:$dst),
- (ins PredRegs:$src1, f32Ext:$src2),
- "if ($src1) $dst = #$src2", []>,
- Requires<[HasV5T]>;
-
-let isExtended = 1, opExtendable = 2, isPredicated = 1, isPredicatedFalse = 1,
- hasSideEffects = 0, validSubTargets = HasV5SubT, isPseudo = 1 in
-def TFRI_cNotPt_f : ALU32_ri<(outs IntRegs:$dst),
- (ins PredRegs:$src1, f32Ext:$src2),
- "if (!$src1) $dst = #$src2", []>,
- Requires<[HasV5T]>;
-
def SDTHexagonI32I64: SDTypeProfile<1, 1, [SDTCisVT<0, i32>,
SDTCisVT<1, i64>]>;
@@ -380,40 +349,40 @@
let Predicates = [HasV5T] in {
def: Pat<(i1 (seto F32:$src1, F32:$src2)),
(C2_not (F2_sfcmpuo F32:$src2, F32:$src1))>;
- def: Pat<(i1 (seto F32:$src1, fpimm:$src2)),
- (C2_not (F2_sfcmpuo (TFRI_f fpimm:$src2), F32:$src1))>;
+ def: Pat<(i1 (seto F32:$src1, f32ImmPred:$src2)),
+ (C2_not (F2_sfcmpuo (f32 (A2_tfrsi (ftoi $src2))), F32:$src1))>;
def: Pat<(i1 (seto F64:$src1, F64:$src2)),
(C2_not (F2_dfcmpuo F64:$src2, F64:$src1))>;
- def: Pat<(i1 (seto F64:$src1, fpimm:$src2)),
- (C2_not (F2_dfcmpuo (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
+ def: Pat<(i1 (seto F64:$src1, f64ImmPred:$src2)),
+ (C2_not (F2_dfcmpuo (CONST64 (ftoi $src2)), F64:$src1))>;
}
// Ordered lt.
let Predicates = [HasV5T] in {
def: Pat<(i1 (setolt F32:$src1, F32:$src2)),
(F2_sfcmpgt F32:$src2, F32:$src1)>;
- def: Pat<(i1 (setolt F32:$src1, fpimm:$src2)),
- (F2_sfcmpgt (f32 (TFRI_f fpimm:$src2)), F32:$src1)>;
+ def: Pat<(i1 (setolt F32:$src1, f32ImmPred:$src2)),
+ (F2_sfcmpgt (f32 (A2_tfrsi (ftoi $src2))), F32:$src1)>;
def: Pat<(i1 (setolt F64:$src1, F64:$src2)),
(F2_dfcmpgt F64:$src2, F64:$src1)>;
- def: Pat<(i1 (setolt F64:$src1, fpimm:$src2)),
- (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
+ def: Pat<(i1 (setolt F64:$src1, f64ImmPred:$src2)),
+ (F2_dfcmpgt (CONST64 (ftoi $src2)), F64:$src1)>;
}
// Unordered lt.
let Predicates = [HasV5T] in {
def: Pat<(i1 (setult F32:$src1, F32:$src2)),
- (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
+ (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
(F2_sfcmpgt F32:$src2, F32:$src1))>;
- def: Pat<(i1 (setult F32:$src1, fpimm:$src2)),
- (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
- (F2_sfcmpgt (TFRI_f fpimm:$src2), F32:$src1))>;
+ def: Pat<(i1 (setult F32:$src1, f32ImmPred:$src2)),
+ (C2_or (F2_sfcmpuo F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))),
+ (F2_sfcmpgt (f32 (A2_tfrsi (ftoi $src2))), F32:$src1))>;
def: Pat<(i1 (setult F64:$src1, F64:$src2)),
- (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
+ (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
(F2_dfcmpgt F64:$src2, F64:$src1))>;
- def: Pat<(i1 (setult F64:$src1, fpimm:$src2)),
- (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
- (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
+ def: Pat<(i1 (setult F64:$src1, f64ImmPred:$src2)),
+ (C2_or (F2_dfcmpuo F64:$src1, (CONST64 (ftoi $src2))),
+ (F2_dfcmpgt (CONST64 (ftoi $src2)), F64:$src1))>;
}
// Ordered le.
@@ -421,31 +390,31 @@
// rs <= rt -> rt >= rs.
def: Pat<(i1 (setole F32:$src1, F32:$src2)),
(F2_sfcmpge F32:$src2, F32:$src1)>;
- def: Pat<(i1 (setole F32:$src1, fpimm:$src2)),
- (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1)>;
+ def: Pat<(i1 (setole F32:$src1, f32ImmPred:$src2)),
+ (F2_sfcmpge (f32 (A2_tfrsi (ftoi $src2))), F32:$src1)>;
// Rss <= Rtt -> Rtt >= Rss.
def: Pat<(i1 (setole F64:$src1, F64:$src2)),
(F2_dfcmpge F64:$src2, F64:$src1)>;
- def: Pat<(i1 (setole F64:$src1, fpimm:$src2)),
- (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
+ def: Pat<(i1 (setole F64:$src1, f64ImmPred:$src2)),
+ (F2_dfcmpge (CONST64 (ftoi $src2)), F64:$src1)>;
}
// Unordered le.
let Predicates = [HasV5T] in {
// rs <= rt -> rt >= rs.
def: Pat<(i1 (setule F32:$src1, F32:$src2)),
- (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
+ (C2_or (F2_sfcmpuo F32:$src1, F32:$src2),
(F2_sfcmpge F32:$src2, F32:$src1))>;
- def: Pat<(i1 (setule F32:$src1, fpimm:$src2)),
- (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
- (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1))>;
+ def: Pat<(i1 (setule F32:$src1, f32ImmPred:$src2)),
+ (C2_or (F2_sfcmpuo F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))),
+ (F2_sfcmpge (f32 (A2_tfrsi (ftoi $src2))), F32:$src1))>;
def: Pat<(i1 (setule F64:$src1, F64:$src2)),
- (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
+ (C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
(F2_dfcmpge F64:$src2, F64:$src1))>;
- def: Pat<(i1 (setule F64:$src1, fpimm:$src2)),
- (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
- (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1))>;
+ def: Pat<(i1 (setule F64:$src1, f64ImmPred:$src2)),
+ (C2_or (F2_dfcmpuo F64:$src1, (CONST64 (ftoi $src2))),
+ (F2_dfcmpge (CONST64 (ftoi $src2)), F64:$src1))>;
}
// Ordered ne.
@@ -454,10 +423,10 @@
(C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
def: Pat<(i1 (setone F64:$src1, F64:$src2)),
(C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
- def: Pat<(i1 (setone F32:$src1, fpimm:$src2)),
- (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2)))>;
- def: Pat<(i1 (setone F64:$src1, fpimm:$src2)),
- (C2_not (F2_dfcmpeq F64:$src1, (CONST64_Float_Real fpimm:$src2)))>;
+ def: Pat<(i1 (setone F32:$src1, f32ImmPred:$src2)),
+ (C2_not (F2_sfcmpeq F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))))>;
+ def: Pat<(i1 (setone F64:$src1, f64ImmPred:$src2)),
+ (C2_not (F2_dfcmpeq F64:$src1, (CONST64 (ftoi $src2))))>;
}
// Unordered ne.
@@ -468,13 +437,14 @@
def: Pat<(i1 (setune F64:$src1, F64:$src2)),
(C2_or (F2_dfcmpuo F64:$src1, F64:$src2),
(C2_not (F2_dfcmpeq F64:$src1, F64:$src2)))>;
- def: Pat<(i1 (setune F32:$src1, fpimm:$src2)),
- (C2_or (F2_sfcmpuo F32:$src1, (TFRI_f fpimm:$src2)),
- (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2))))>;
- def: Pat<(i1 (setune F64:$src1, fpimm:$src2)),
- (C2_or (F2_dfcmpuo F64:$src1, (CONST64_Float_Real fpimm:$src2)),
+ def: Pat<(i1 (setune F32:$src1, f32ImmPred:$src2)),
+ (C2_or (F2_sfcmpuo F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))),
+ (C2_not (F2_sfcmpeq F32:$src1,
+ (f32 (A2_tfrsi (ftoi $src2))))))>;
+ def: Pat<(i1 (setune F64:$src1, f64ImmPred:$src2)),
+ (C2_or (F2_dfcmpuo F64:$src1, (CONST64 (ftoi $src2))),
(C2_not (F2_dfcmpeq F64:$src1,
- (CONST64_Float_Real fpimm:$src2))))>;
+ (CONST64 (ftoi $src2)))))>;
}
// Besides set[o|u][comparions], we also need set[comparisons].
@@ -482,35 +452,35 @@
// lt.
def: Pat<(i1 (setlt F32:$src1, F32:$src2)),
(F2_sfcmpgt F32:$src2, F32:$src1)>;
- def: Pat<(i1 (setlt F32:$src1, fpimm:$src2)),
- (F2_sfcmpgt (TFRI_f fpimm:$src2), F32:$src1)>;
+ def: Pat<(i1 (setlt F32:$src1, f32ImmPred:$src2)),
+ (F2_sfcmpgt (f32 (A2_tfrsi (ftoi $src2))), F32:$src1)>;
def: Pat<(i1 (setlt F64:$src1, F64:$src2)),
(F2_dfcmpgt F64:$src2, F64:$src1)>;
- def: Pat<(i1 (setlt F64:$src1, fpimm:$src2)),
- (F2_dfcmpgt (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
+ def: Pat<(i1 (setlt F64:$src1, f64ImmPred:$src2)),
+ (F2_dfcmpgt (CONST64 (ftoi $src2)), F64:$src1)>;
// le.
// rs <= rt -> rt >= rs.
def: Pat<(i1 (setle F32:$src1, F32:$src2)),
(F2_sfcmpge F32:$src2, F32:$src1)>;
- def: Pat<(i1 (setle F32:$src1, fpimm:$src2)),
- (F2_sfcmpge (TFRI_f fpimm:$src2), F32:$src1)>;
+ def: Pat<(i1 (setle F32:$src1, f32ImmPred:$src2)),
+ (F2_sfcmpge (f32 (A2_tfrsi (ftoi $src2))), F32:$src1)>;
// Rss <= Rtt -> Rtt >= Rss.
def: Pat<(i1 (setle F64:$src1, F64:$src2)),
(F2_dfcmpge F64:$src2, F64:$src1)>;
- def: Pat<(i1 (setle F64:$src1, fpimm:$src2)),
- (F2_dfcmpge (CONST64_Float_Real fpimm:$src2), F64:$src1)>;
+ def: Pat<(i1 (setle F64:$src1, f64ImmPred:$src2)),
+ (F2_dfcmpge (CONST64 (ftoi $src2)), F64:$src1)>;
// ne.
def: Pat<(i1 (setne F32:$src1, F32:$src2)),
(C2_not (F2_sfcmpeq F32:$src1, F32:$src2))>;
def: Pat<(i1 (setne F64:$src1, F64:$src2)),
(C2_not (F2_dfcmpeq F64:$src1, F64:$src2))>;
- def: Pat<(i1 (setne F32:$src1, fpimm:$src2)),
- (C2_not (F2_sfcmpeq F32:$src1, (TFRI_f fpimm:$src2)))>;
- def: Pat<(i1 (setne F64:$src1, fpimm:$src2)),
- (C2_not (F2_dfcmpeq F64:$src1, (CONST64_Float_Real fpimm:$src2)))>;
+ def: Pat<(i1 (setne F32:$src1, f32ImmPred:$src2)),
+ (C2_not (F2_sfcmpeq F32:$src1, (f32 (A2_tfrsi (ftoi $src2)))))>;
+ def: Pat<(i1 (setne F64:$src1, f64ImmPred:$src2)),
+ (C2_not (F2_dfcmpeq F64:$src1, (CONST64 (ftoi $src2))))>;
}
// F2 convert template classes:
@@ -708,7 +678,7 @@
def F2_sffma_lib: T_sfmpy_acc <0, 1>;
def F2_sffms_lib: T_sfmpy_acc <1, 1>;
-def : Pat <(f32 (fma F32:$src2, F32:$src3, F32:$src1)),
+def : Pat <(fma F32:$src2, F32:$src3, F32:$src1),
(F2_sffma F32:$src1, F32:$src2, F32:$src3)>;
// Floating-point fused multiply add w/ additional scaling (2**pu).
@@ -735,20 +705,12 @@
let Inst{4-0} = Rx;
}
-let isExtended = 1, isExtentSigned = 1, opExtentBits = 8, opExtendable = 3,
- isPseudo = 1, InputType = "imm" in
-def MUX_ir_f : ALU32_rr<(outs IntRegs:$dst),
- (ins PredRegs:$src1, IntRegs:$src2, f32Ext:$src3),
- "$dst = mux($src1, $src2, #$src3)",
- [(set F32:$dst, (f32 (select I1:$src1, F32:$src2, fpimm:$src3)))]>,
+def: Pat<(select I1:$Pu, F32:$Rs, f32ImmPred:$imm),
+ (C2_muxir I1:$Pu, F32:$Rs, (ftoi $imm))>,
Requires<[HasV5T]>;
-let isExtended = 1, isExtentSigned = 1, opExtentBits = 8, opExtendable = 2,
- isPseudo = 1, InputType = "imm" in
-def MUX_ri_f : ALU32_rr<(outs IntRegs:$dst),
- (ins PredRegs:$src1, f32Ext:$src2, IntRegs:$src3),
- "$dst = mux($src1, #$src2, $src3)",
- [(set F32:$dst, (f32 (select I1:$src1, fpimm:$src2, F32:$src3)))]>,
+def: Pat<(select I1:$Pu, f32ImmPred:$imm, F32:$Rt),
+ (C2_muxri I1:$Pu, (ftoi $imm), F32:$Rt)>,
Requires<[HasV5T]>;
def: Pat<(select I1:$src1, F32:$src2, F32:$src3),
@@ -768,15 +730,15 @@
Requires<[HasV5T]>;
// Map from p0 = pnot(p0); r0 = select(p0, #i, r1)
-// => r0 = MUX_ir_f(p0, #i, r1)
-def: Pat<(select (not I1:$src1), fpimm:$src2, F32:$src3),
- (MUX_ir_f I1:$src1, F32:$src3, fpimm:$src2)>,
+// => r0 = mux(p0, #i, r1)
+def: Pat<(select (not I1:$src1), f32ImmPred:$src2, F32:$src3),
+ (C2_muxir I1:$src1, F32:$src3, (ftoi $src2))>,
Requires<[HasV5T]>;
// Map from p0 = pnot(p0); r0 = mux(p0, r1, #i)
-// => r0 = MUX_ri_f(p0, r1, #i)
-def: Pat<(select (not I1:$src1), F32:$src2, fpimm:$src3),
- (MUX_ri_f I1:$src1, fpimm:$src3, F32:$src2)>,
+// => r0 = mux(p0, r1, #i)
+def: Pat<(select (not I1:$src1), F32:$src2, f32ImmPred:$src3),
+ (C2_muxri I1:$src1, (ftoi $src3), F32:$src2)>,
Requires<[HasV5T]>;
def: Pat<(i32 (fp_to_sint F64:$src1)),
@@ -873,7 +835,7 @@
// Classify floating-point value
let isFP = 1 in
- def F2_sfclass : T_TEST_BIT_IMM<"sfclass", 0b111>;
+def F2_sfclass : T_TEST_BIT_IMM<"sfclass", 0b111>;
let isFP = 1 in
def F2_dfclass: ALU64Inst<(outs PredRegs:$Pd), (ins DoubleRegs:$Rss, u5Imm:$u5),
@@ -910,8 +872,8 @@
}
let hasNewValue = 1, opNewValue = 0 in {
-def F2_sfimm_p : T_fimm <"sfmake", IntRegs, 0b0110, 0>;
-def F2_sfimm_n : T_fimm <"sfmake", IntRegs, 0b0110, 1>;
+ def F2_sfimm_p : T_fimm <"sfmake", IntRegs, 0b0110, 0>;
+ def F2_sfimm_n : T_fimm <"sfmake", IntRegs, 0b0110, 1>;
}
def F2_dfimm_p : T_fimm <"dfmake", DoubleRegs, 0b1001, 0>;
diff --git a/llvm/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp b/llvm/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp
index 5a94cce..a7c3173 100644
--- a/llvm/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp
@@ -8,7 +8,7 @@
//===----------------------------------------------------------------------===//
//
// When the compiler is invoked with no small data, for instance, with the -G0
-// command line option, then all CONST32_* opcodes should be broken down into
+// command line option, then all CONST* opcodes should be broken down into
// appropriate LO and HI instructions. This splitting is done by this pass.
// The only reason this is not done in the DAG lowering itself is that there
// is no simple way of getting the register allocator to allot the same hard
@@ -88,8 +88,7 @@
while (MII != MIE) {
MachineInstr &MI = *MII;
int Opc = MI.getOpcode();
- if (Opc == Hexagon::CONST32_Int_Real &&
- MI.getOperand(1).isBlockAddress()) {
+ if (Opc == Hexagon::CONST32 && MI.getOperand(1).isBlockAddress()) {
int DestReg = MI.getOperand(0).getReg();
MachineOperand &Symbol = MI.getOperand(1);
@@ -103,40 +102,21 @@
continue;
}
- else if (Opc == Hexagon::CONST32_Int_Real ||
- Opc == Hexagon::CONST32_Float_Real) {
+ else if (Opc == Hexagon::CONST32) {
int DestReg = MI.getOperand(0).getReg();
// We have to convert an FP immediate into its corresponding integer
// representation
- int64_t ImmValue;
- if (Opc == Hexagon::CONST32_Float_Real) {
- APFloat Val = MI.getOperand(1).getFPImm()->getValueAPF();
- ImmValue = *Val.bitcastToAPInt().getRawData();
- }
- else
- ImmValue = MI.getOperand(1).getImm();
-
+ int64_t ImmValue = MI.getOperand(1).getImm();
BuildMI(*MBB, MII, MI.getDebugLoc(), TII->get(Hexagon::A2_tfrsi),
DestReg)
.addImm(ImmValue);
MII = MBB->erase(&MI);
continue;
}
- else if (Opc == Hexagon::CONST64_Int_Real ||
- Opc == Hexagon::CONST64_Float_Real) {
+ else if (Opc == Hexagon::CONST64) {
int DestReg = MI.getOperand(0).getReg();
-
- // We have to convert an FP immediate into its corresponding integer
- // representation
- int64_t ImmValue;
- if (Opc == Hexagon::CONST64_Float_Real) {
- APFloat Val = MI.getOperand(1).getFPImm()->getValueAPF();
- ImmValue = *Val.bitcastToAPInt().getRawData();
- }
- else
- ImmValue = MI.getOperand(1).getImm();
-
+ int64_t ImmValue = MI.getOperand(1).getImm();
unsigned DestLo = TRI->getSubReg(DestReg, Hexagon::subreg_loreg);
unsigned DestHi = TRI->getSubReg(DestReg, Hexagon::subreg_hireg);
diff --git a/llvm/lib/Target/Hexagon/HexagonSplitDouble.cpp b/llvm/lib/Target/Hexagon/HexagonSplitDouble.cpp
index 25b2aff..7ca80b0 100644
--- a/llvm/lib/Target/Hexagon/HexagonSplitDouble.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonSplitDouble.cpp
@@ -170,7 +170,7 @@
case Hexagon::A4_combineii:
case Hexagon::A4_combineri:
case Hexagon::A2_combinew:
- case Hexagon::CONST64_Int_Real:
+ case Hexagon::CONST64:
case Hexagon::A2_sxtw:
@@ -319,7 +319,7 @@
return 2;
case Hexagon::A2_tfrpi:
- case Hexagon::CONST64_Int_Real: {
+ case Hexagon::CONST64: {
uint64_t D = MI->getOperand(1).getImm();
unsigned Lo = D & 0xFFFFFFFFULL;
unsigned Hi = D >> 32;
@@ -995,7 +995,7 @@
break;
case A2_tfrpi:
- case CONST64_Int_Real:
+ case CONST64:
splitImmediate(MI, PairMap);
Split = true;
break;
