[PATCH] [ARMv8] CodeGen for VSEL
Joey Gouly
joey.gouly at arm.com
Thu Aug 22 05:25:10 PDT 2013
Big cleanup, thanks to Tim's CMOV work.
http://llvm-reviews.chandlerc.com/D1445
CHANGE SINCE LAST DIFF
http://llvm-reviews.chandlerc.com/D1445?vs=3592&id=3664#toc
Files:
lib/Target/ARM/ARMISelLowering.cpp
lib/Target/ARM/ARMInstrVFP.td
Index: lib/Target/ARM/ARMISelLowering.cpp
===================================================================
--- lib/Target/ARM/ARMISelLowering.cpp
+++ lib/Target/ARM/ARMISelLowering.cpp
@@ -3178,6 +3178,61 @@
SelectTrue, SelectFalse, ISD::SETNE);
}
+static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
+ if (CC == ISD::SETNE)
+ return ISD::SETEQ;
+ return ISD::getSetCCSwappedOperands(CC);
+}
+
+static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
+ bool &swpCmpOps, bool &swpVselOps) {
+ // Start by selecting the GE condition code for opcodes that return true for
+ // 'equality'
+ if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
+ CC == ISD::SETULE)
+ CondCode = ARMCC::GE;
+
+ // and GT for opcodes that return false for 'equality'.
+ else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
+ CC == ISD::SETULT)
+ CondCode = ARMCC::GT;
+
+ // Since we are constrained to GE/GT, if the opcode contains 'less', we need
+ // to swap the compare operands.
+ if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
+ CC == ISD::SETULT)
+ swpCmpOps = true;
+
+ // Both GT and GE are ordered comparisons, and return false for 'unordered'.
+ // If we have an unordered opcode, we need to swap the operands to the VSEL
+ // instruction (effectively negating the condition).
+ //
+ // This also has the effect of swapping which one of 'less' or 'greater'
+ // returns true, so we also swap the compare operands. It also switches
+ // whether we return true for 'equality', so we compensate by picking the
+ // opposite condition code to our original choice.
+ if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
+ CC == ISD::SETUGT) {
+ swpCmpOps = !swpCmpOps;
+ swpVselOps = !swpVselOps;
+ CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
+ }
+
+ // 'ordered' is 'anything but unordered', so use the VS condition code and
+ // swap the VSEL operands.
+ if (CC == ISD::SETO) {
+ CondCode = ARMCC::VS;
+ swpVselOps = true;
+ }
+
+ // 'unordered or not equal' is 'anything but equal', so use the EQ condition
+ // code and swap the VSEL operands.
+ if (CC == ISD::SETUNE) {
+ CondCode = ARMCC::EQ;
+ swpVselOps = true;
+ }
+}
+
SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
EVT VT = Op.getValueType();
SDValue LHS = Op.getOperand(0);
@@ -3188,15 +3243,52 @@
SDLoc dl(Op);
if (LHS.getValueType() == MVT::i32) {
+ // Try to generate VSEL on ARMv8.
+ // The VSEL instruction can't use all the usual ARM condition
+ // codes: it only has two bits to select the condition code, so it's
+ // constrained to use only GE, GT, VS and EQ.
+ //
+ // To implement all the various ISD::SETXXX opcodes, we sometimes need to
+ // swap the operands of the previous compare instruction (effectively
+ // inverting the compare condition, swapping 'less' and 'greater') and
+ // sometimes need to swap the operands to the VSEL (which inverts the
+ // condition in the sense of firing whenever the previous condition didn't)
+ if (getSubtarget()->hasV8FP() && (TrueVal.getValueType() == MVT::f32 ||
+ TrueVal.getValueType() == MVT::f64)) {
+ ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
+ if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
+ CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
+ CC = getInverseCCForVSEL(CC);
+ std::swap(TrueVal, FalseVal);
+ }
+ }
+
SDValue ARMcc;
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
- return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
+ return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
+ Cmp);
}
ARMCC::CondCodes CondCode, CondCode2;
FPCCToARMCC(CC, CondCode, CondCode2);
+ // Try to generate VSEL on ARMv8.
+ if (getSubtarget()->hasV8FP() && (TrueVal.getValueType() == MVT::f32 ||
+ TrueVal.getValueType() == MVT::f64)) {
+ bool swpCmpOps = false;
+ bool swpVselOps = false;
+ checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
+
+ if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
+ CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
+ if (swpCmpOps)
+ std::swap(LHS, RHS);
+ if (swpVselOps)
+ std::swap(TrueVal, FalseVal);
+ }
+ }
+
SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
Index: lib/Target/ARM/ARMInstrVFP.td
===================================================================
--- lib/Target/ARM/ARMInstrVFP.td
+++ lib/Target/ARM/ARMInstrVFP.td
@@ -333,8 +333,8 @@
let D = VFPNeonA8Domain;
}
-multiclass vsel_inst<string op, bits<2> opc> {
- let DecoderNamespace = "VFPV8", PostEncoderMethod = "" in {
+multiclass vsel_inst<string op, bits<2> opc, int CC> {
+ let DecoderNamespace = "VFPV8", PostEncoderMethod = "", Uses = [CPSR] in {
def S : ASbInp<0b11100, opc, 0,
(outs SPR:$Sd), (ins SPR:$Sn, SPR:$Sm),
NoItinerary, !strconcat("vsel", op, ".f32\t$Sd, $Sn, $Sm"),
@@ -345,12 +345,20 @@
NoItinerary, !strconcat("vsel", op, ".f64\t$Dd, $Dn, $Dm"),
[]>, Requires<[HasV8FP]>;
}
+
+ let AddedComplexity = 4 in {
+ def : Pat<(ARMcmov SPR:$Sn, SPR:$Sm, CC),
+ (!cast<Instruction>(NAME#"S") SPR:$Sm, SPR:$Sn)>;
+ def : Pat<(ARMcmov (f64 DPR:$Dn), (f64 DPR:$Dm), CC),
+ (!cast<Instruction>(NAME#"D") DPR:$Dm, DPR:$Dn)>;
+ }
}
-defm VSELGT : vsel_inst<"gt", 0b11>;
-defm VSELGE : vsel_inst<"ge", 0b10>;
-defm VSELEQ : vsel_inst<"eq", 0b00>;
-defm VSELVS : vsel_inst<"vs", 0b01>;
+// The CC constants here match ARMCC::CondCodes.
+defm VSELGT : vsel_inst<"gt", 0b11, 12>;
+defm VSELGE : vsel_inst<"ge", 0b10, 10>;
+defm VSELEQ : vsel_inst<"eq", 0b00, 0>;
+defm VSELVS : vsel_inst<"vs", 0b01, 6>;
multiclass vmaxmin_inst<string op, bit opc> {
let DecoderNamespace = "VFPV8", PostEncoderMethod = "" in {
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D1445.2.patch
Type: text/x-patch
Size: 6355 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130822/6d9a92f8/attachment.bin>
More information about the llvm-commits
mailing list