Rewrite fast-isel integer cast handling to handle more cases, and to be simpler and more consistent.
The practical effects here are that x86-64 fast-isel can now handle trunc from i8 to i1, and ARM fast-isel can handle many more constructs involving integers narrower than 32 bits (including loads, stores, and many integer casts).
rdar://9437928 .
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@132099 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMFastISel.cpp b/lib/Target/ARM/ARMFastISel.cpp
index 790a6e3..2d2ac67 100644
--- a/lib/Target/ARM/ARMFastISel.cpp
+++ b/lib/Target/ARM/ARMFastISel.cpp
@@ -167,6 +167,7 @@
bool SelectCall(const Instruction *I);
bool SelectSelect(const Instruction *I);
bool SelectRet(const Instruction *I);
+ bool SelectIntCast(const Instruction *I);
// Utility routines.
private:
@@ -1129,7 +1130,7 @@
} else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
MVT SourceVT;
if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
- (isTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
+ (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
unsigned TstOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
unsigned OpReg = getRegForValue(TI->getOperand(0));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
@@ -1948,6 +1949,77 @@
}
+bool ARMFastISel::SelectIntCast(const Instruction *I) {
+ // On ARM, in general, integer casts don't involve legal types; this code
+ // handles promotable integers. The high bits for a type smaller than
+ // the register size are assumed to be undefined.
+ const Type *DestTy = I->getType();
+ Value *Op = I->getOperand(0);
+ const Type *SrcTy = Op->getType();
+
+ EVT SrcVT, DestVT;
+ SrcVT = TLI.getValueType(SrcTy, true);
+ DestVT = TLI.getValueType(DestTy, true);
+
+ if (isa<TruncInst>(I)) {
+ if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
+ return false;
+ if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
+ return false;
+
+ unsigned SrcReg = getRegForValue(Op);
+ if (!SrcReg) return false;
+
+ // Because the high bits are undefined, a truncate doesn't generate
+ // any code.
+ UpdateValueMap(I, SrcReg);
+ return true;
+ }
+ if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
+ return false;
+
+ unsigned Opc;
+ bool isZext = isa<ZExtInst>(I);
+ bool isBoolZext = false;
+ switch (SrcVT.getSimpleVT().SimpleTy) {
+ default: return false;
+ case MVT::i16:
+ if (isZext)
+ Opc = isThumb ? ARM::t2UXTHr : ARM::UXTHr;
+ else
+ Opc = isThumb ? ARM::t2SXTHr : ARM::SXTHr;
+ break;
+ case MVT::i8:
+ if (isZext)
+ Opc = isThumb ? ARM::t2UXTBr : ARM::UXTBr;
+ else
+ Opc = isThumb ? ARM::t2SXTBr : ARM::SXTBr;
+ break;
+ case MVT::i1:
+ if (isZext) {
+ Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
+ isBoolZext = true;
+ break;
+ }
+ return false;
+ }
+
+ // FIXME: We could save an instruction in many cases by special-casing
+ // load instructions.
+ unsigned SrcReg = getRegForValue(Op);
+ if (!SrcReg) return false;
+
+ unsigned DestReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+ MachineInstrBuilder MIB;
+ MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
+ .addReg(SrcReg);
+ if (isBoolZext)
+ MIB.addImm(1);
+ AddOptionalDefs(MIB);
+ UpdateValueMap(I, DestReg);
+ return true;
+}
+
// TODO: SoftFP support.
bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
@@ -1985,6 +2057,10 @@
return SelectSelect(I);
case Instruction::Ret:
return SelectRet(I);
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ return SelectIntCast(I);
default: break;
}
return false;
diff --git a/lib/Target/X86/X86FastISel.cpp b/lib/Target/X86/X86FastISel.cpp
index 56a777f..5162bc5 100644
--- a/lib/Target/X86/X86FastISel.cpp
+++ b/lib/Target/X86/X86FastISel.cpp
@@ -936,18 +936,31 @@
bool X86FastISel::X86SelectZExt(const Instruction *I) {
// Handle zero-extension from i1 to i8, which is common.
- if (I->getType()->isIntegerTy(8) &&
- I->getOperand(0)->getType()->isIntegerTy(1)) {
- unsigned ResultReg = getRegForValue(I->getOperand(0));
- if (ResultReg == 0) return false;
- // Set the high bits to zero.
- ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
- if (ResultReg == 0) return false;
- UpdateValueMap(I, ResultReg);
- return true;
+ if (!I->getOperand(0)->getType()->isIntegerTy(1))
+ return false;
+
+ EVT DstVT = TLI.getValueType(I->getType());
+ if (!TLI.isTypeLegal(DstVT))
+ return false;
+
+ unsigned ResultReg = getRegForValue(I->getOperand(0));
+ if (ResultReg == 0)
+ return false;
+
+ // Set the high bits to zero.
+ ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
+ if (ResultReg == 0)
+ return false;
+
+ if (DstVT != MVT::i8) {
+ ResultReg = FastEmit_r(MVT::i8, DstVT.getSimpleVT(), ISD::ZERO_EXTEND,
+ ResultReg, /*Kill=*/true);
+ if (ResultReg == 0)
+ return false;
}
- return false;
+ UpdateValueMap(I, ResultReg);
+ return true;
}
@@ -1229,18 +1242,13 @@
}
bool X86FastISel::X86SelectTrunc(const Instruction *I) {
- if (Subtarget->is64Bit())
- // All other cases should be handled by the tblgen generated code.
- return false;
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
EVT DstVT = TLI.getValueType(I->getType());
- // This code only handles truncation to byte right now.
+ // This code only handles truncation to byte.
if (DstVT != MVT::i8 && DstVT != MVT::i1)
- // All other cases should be handled by the tblgen generated code.
return false;
- if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
- // All other cases should be handled by the tblgen generated code.
+ if (!TLI.isTypeLegal(SrcVT))
return false;
unsigned InputReg = getRegForValue(I->getOperand(0));
@@ -1248,16 +1256,26 @@
// Unhandled operand. Halt "fast" selection and bail.
return false;
- // First issue a copy to GR16_ABCD or GR32_ABCD.
- const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
- ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
- unsigned CopyReg = createResultReg(CopyRC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- CopyReg).addReg(InputReg);
+ if (SrcVT == MVT::i8) {
+ // Truncate from i8 to i1; no code needed.
+ UpdateValueMap(I, InputReg);
+ return true;
+ }
- // Then issue an extract_subreg.
+ if (!Subtarget->is64Bit()) {
+ // If we're on x86-32; we can't extract an i8 from a general register.
+ // First issue a copy to GR16_ABCD or GR32_ABCD.
+ const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
+ ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
+ unsigned CopyReg = createResultReg(CopyRC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
+ CopyReg).addReg(InputReg);
+ InputReg = CopyReg;
+ }
+
+ // Issue an extract_subreg.
unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
- CopyReg, /*Kill=*/true,
+ InputReg, /*Kill=*/true,
X86::sub_8bit);
if (!ResultReg)
return false;