#define DEBUG_TYPE "t2-reduce-size"
#include "ARM.h"
-#include "ARMAddressingModes.h"
-#include "ARMBaseRegisterInfo.h"
#include "ARMBaseInstrInfo.h"
+#include "ARMBaseRegisterInfo.h"
#include "ARMSubtarget.h"
+#include "MCTargetDesc/ARMAddressingModes.h"
#include "Thumb2InstrInfo.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/PostOrderIterator.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/IR/Function.h" // To access Function attributes
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/Statistic.h"
using namespace llvm;
STATISTIC(NumNarrows, "Number of 32-bit instrs reduced to 16-bit ones");
/// ReduceTable - A static table with information on mapping from wide
/// opcodes to narrow
struct ReduceEntry {
- unsigned WideOpc; // Wide opcode
- unsigned NarrowOpc1; // Narrow opcode to transform to
- unsigned NarrowOpc2; // Narrow opcode when it's two-address
+ uint16_t WideOpc; // Wide opcode
+ uint16_t NarrowOpc1; // Narrow opcode to transform to
+ uint16_t NarrowOpc2; // Narrow opcode when it's two-address
uint8_t Imm1Limit; // Limit of immediate field (bits)
uint8_t Imm2Limit; // Limit of immediate field when it's two-address
unsigned LowRegs1 : 1; // Only possible if low-registers are used
unsigned PredCC2 : 2;
unsigned PartFlag : 1; // 16-bit instruction does partial flag update
unsigned Special : 1; // Needs to be dealt with specially
+ unsigned AvoidMovs: 1; // Avoid movs with shifter operand (for Swift)
};
static const ReduceEntry ReduceTable[] = {
- // Wide, Narrow1, Narrow2, imm1,imm2, lo1, lo2, P/C, PF, S
- { ARM::t2ADCrr, 0, ARM::tADC, 0, 0, 0, 1, 0,0, 0,0 },
- { ARM::t2ADDri, ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 0,0, 0,1 },
- { ARM::t2ADDrr, ARM::tADDrr, ARM::tADDhirr, 0, 0, 1, 0, 0,1, 0,0 },
- { ARM::t2ADDSri,ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 2,2, 0,1 },
- { ARM::t2ADDSrr,ARM::tADDrr, 0, 0, 0, 1, 0, 2,0, 0,1 },
- { ARM::t2ANDrr, 0, ARM::tAND, 0, 0, 0, 1, 0,0, 1,0 },
- { ARM::t2ASRri, ARM::tASRri, 0, 5, 0, 1, 0, 0,0, 1,0 },
- { ARM::t2ASRrr, 0, ARM::tASRrr, 0, 0, 0, 1, 0,0, 1,0 },
- { ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 1,0 },
- //FIXME: Disable CMN, as CCodes are backwards from compare expectations
- //{ ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0,0 },
- { ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 2,0, 0,0 },
- { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 0,1 },
- { ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 1,0 },
- // FIXME: adr.n immediate offset must be multiple of 4.
- //{ ARM::t2LEApcrelJT,ARM::tLEApcrelJT, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2LSLri, ARM::tLSLri, 0, 5, 0, 1, 0, 0,0, 1,0 },
- { ARM::t2LSLrr, 0, ARM::tLSLrr, 0, 0, 0, 1, 0,0, 1,0 },
- { ARM::t2LSRri, ARM::tLSRri, 0, 5, 0, 1, 0, 0,0, 1,0 },
- { ARM::t2LSRrr, 0, ARM::tLSRrr, 0, 0, 0, 1, 0,0, 1,0 },
- // FIXME: tMOVi8 and tMVN also partially update CPSR but they are less
- // likely to cause issue in the loop. As a size / performance workaround,
- // they are not marked as such.
- { ARM::t2MOVi, ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 0,0 },
- { ARM::t2MOVi16,ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 0,1 },
- // FIXME: Do we need the 16-bit 'S' variant?
- { ARM::t2MOVr,ARM::tMOVgpr2gpr,0, 0, 0, 0, 0, 1,0, 0,0 },
- { ARM::t2MOVCCr,0, ARM::tMOVCCr, 0, 0, 0, 0, 0,1, 0,0 },
- { ARM::t2MOVCCi,0, ARM::tMOVCCi, 0, 8, 0, 1, 0,1, 0,0 },
- { ARM::t2MUL, 0, ARM::tMUL, 0, 0, 0, 1, 0,0, 1,0 },
- { ARM::t2MVNr, ARM::tMVN, 0, 0, 0, 1, 0, 0,0, 0,0 },
- { ARM::t2ORRrr, 0, ARM::tORR, 0, 0, 0, 1, 0,0, 1,0 },
- { ARM::t2REV, ARM::tREV, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2REV16, ARM::tREV16, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2REVSH, ARM::tREVSH, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2RORrr, 0, ARM::tROR, 0, 0, 0, 1, 0,0, 1,0 },
- { ARM::t2RSBri, ARM::tRSB, 0, 0, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2RSBSri,ARM::tRSB, 0, 0, 0, 1, 0, 2,0, 0,1 },
- { ARM::t2SBCrr, 0, ARM::tSBC, 0, 0, 0, 1, 0,0, 0,0 },
- { ARM::t2SUBri, ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 0,0, 0,0 },
- { ARM::t2SUBrr, ARM::tSUBrr, 0, 0, 0, 1, 0, 0,0, 0,0 },
- { ARM::t2SUBSri,ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 2,2, 0,0 },
- { ARM::t2SUBSrr,ARM::tSUBrr, 0, 0, 0, 1, 0, 2,0, 0,0 },
- { ARM::t2SXTBr, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2SXTHr, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2TSTrr, ARM::tTST, 0, 0, 0, 1, 0, 2,0, 0,0 },
- { ARM::t2UXTBr, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0,0 },
- { ARM::t2UXTHr, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0,0 },
-
- // FIXME: Clean this up after splitting each Thumb load / store opcode
- // into multiple ones.
- { ARM::t2LDRi12,ARM::tLDRi, ARM::tLDRspi, 5, 8, 1, 0, 0,0, 0,1 },
- { ARM::t2LDRs, ARM::tLDRr, 0, 0, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2LDRBi12,ARM::tLDRBi, 0, 5, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2LDRBs, ARM::tLDRBr, 0, 0, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2LDRHi12,ARM::tLDRHi, 0, 5, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2LDRHs, ARM::tLDRHr, 0, 0, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2LDRSBs,ARM::tLDRSB, 0, 0, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2LDRSHs,ARM::tLDRSH, 0, 0, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2STRi12,ARM::tSTRi, ARM::tSTRspi, 5, 8, 1, 0, 0,0, 0,1 },
- { ARM::t2STRs, ARM::tSTRr, 0, 0, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2STRBi12,ARM::tSTRBi, 0, 5, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2STRBs, ARM::tSTRBr, 0, 0, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2STRHi12,ARM::tSTRHi, 0, 5, 0, 1, 0, 0,0, 0,1 },
- { ARM::t2STRHs, ARM::tSTRHr, 0, 0, 0, 1, 0, 0,0, 0,1 },
-
- { ARM::t2LDMIA, ARM::tLDMIA, 0, 0, 0, 1, 1, 1,1, 0,1 },
- { ARM::t2LDMIA_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 0,1 },
- { ARM::t2LDMIA_UPD,ARM::tLDMIA_UPD,ARM::tPOP,0, 0, 1, 1, 1,1, 0,1 },
- // ARM::t2STM (with no basereg writeback) has no Thumb1 equivalent
- { ARM::t2STMIA_UPD,ARM::tSTMIA_UPD, 0, 0, 0, 1, 1, 1,1, 0,1 },
- { ARM::t2STMDB_UPD, 0, ARM::tPUSH, 0, 0, 1, 1, 1,1, 0,1 },
+ // Wide, Narrow1, Narrow2, imm1,imm2, lo1, lo2, P/C,PF,S,AM
+ { ARM::t2ADCrr, 0, ARM::tADC, 0, 0, 0, 1, 0,0, 0,0,0 },
+ { ARM::t2ADDri, ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 0,0, 0,1,0 },
+ { ARM::t2ADDrr, ARM::tADDrr, ARM::tADDhirr, 0, 0, 1, 0, 0,1, 0,0,0 },
+ { ARM::t2ADDSri,ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 2,2, 0,1,0 },
+ { ARM::t2ADDSrr,ARM::tADDrr, 0, 0, 0, 1, 0, 2,0, 0,1,0 },
+ { ARM::t2ANDrr, 0, ARM::tAND, 0, 0, 0, 1, 0,0, 1,0,0 },
+ { ARM::t2ASRri, ARM::tASRri, 0, 5, 0, 1, 0, 0,0, 1,0,1 },
+ { ARM::t2ASRrr, 0, ARM::tASRrr, 0, 0, 0, 1, 0,0, 1,0,1 },
+ { ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 1,0,0 },
+ //FIXME: Disable CMN, as CCodes are backwards from compare expectations
+ //{ ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0,0,0 },
+ { ARM::t2CMNzrr, ARM::tCMNz, 0, 0, 0, 1, 0, 2,0, 0,0,0 },
+ { ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 2,0, 0,0,0 },
+ { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 0,1,0 },
+ { ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 1,0,0 },
+ // FIXME: adr.n immediate offset must be multiple of 4.
+ //{ ARM::t2LEApcrelJT,ARM::tLEApcrelJT, 0, 0, 0, 1, 0, 1,0, 0,0,0 },
+ { ARM::t2LSLri, ARM::tLSLri, 0, 5, 0, 1, 0, 0,0, 1,0,1 },
+ { ARM::t2LSLrr, 0, ARM::tLSLrr, 0, 0, 0, 1, 0,0, 1,0,1 },
+ { ARM::t2LSRri, ARM::tLSRri, 0, 5, 0, 1, 0, 0,0, 1,0,1 },
+ { ARM::t2LSRrr, 0, ARM::tLSRrr, 0, 0, 0, 1, 0,0, 1,0,1 },
+ { ARM::t2MOVi, ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 1,0,0 },
+ { ARM::t2MOVi16,ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 1,1,0 },
+ // FIXME: Do we need the 16-bit 'S' variant?
+ { ARM::t2MOVr,ARM::tMOVr, 0, 0, 0, 0, 0, 1,0, 0,0,0 },
+ { ARM::t2MUL, 0, ARM::tMUL, 0, 0, 0, 1, 0,0, 1,0,0 },
+ { ARM::t2MVNr, ARM::tMVN, 0, 0, 0, 1, 0, 0,0, 0,0,0 },
+ { ARM::t2ORRrr, 0, ARM::tORR, 0, 0, 0, 1, 0,0, 1,0,0 },
+ { ARM::t2REV, ARM::tREV, 0, 0, 0, 1, 0, 1,0, 0,0,0 },
+ { ARM::t2REV16, ARM::tREV16, 0, 0, 0, 1, 0, 1,0, 0,0,0 },
+ { ARM::t2REVSH, ARM::tREVSH, 0, 0, 0, 1, 0, 1,0, 0,0,0 },
+ { ARM::t2RORrr, 0, ARM::tROR, 0, 0, 0, 1, 0,0, 1,0,0 },
+ { ARM::t2RSBri, ARM::tRSB, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2RSBSri,ARM::tRSB, 0, 0, 0, 1, 0, 2,0, 0,1,0 },
+ { ARM::t2SBCrr, 0, ARM::tSBC, 0, 0, 0, 1, 0,0, 0,0,0 },
+ { ARM::t2SUBri, ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 0,0, 0,0,0 },
+ { ARM::t2SUBrr, ARM::tSUBrr, 0, 0, 0, 1, 0, 0,0, 0,0,0 },
+ { ARM::t2SUBSri,ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 2,2, 0,0,0 },
+ { ARM::t2SUBSrr,ARM::tSUBrr, 0, 0, 0, 1, 0, 2,0, 0,0,0 },
+ { ARM::t2SXTB, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0,1,0 },
+ { ARM::t2SXTH, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0,1,0 },
+ { ARM::t2TSTrr, ARM::tTST, 0, 0, 0, 1, 0, 2,0, 0,0,0 },
+ { ARM::t2UXTB, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0,1,0 },
+ { ARM::t2UXTH, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0,1,0 },
+
+ // FIXME: Clean this up after splitting each Thumb load / store opcode
+ // into multiple ones.
+ { ARM::t2LDRi12,ARM::tLDRi, ARM::tLDRspi, 5, 8, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2LDRs, ARM::tLDRr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2LDRBi12,ARM::tLDRBi, 0, 5, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2LDRBs, ARM::tLDRBr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2LDRHi12,ARM::tLDRHi, 0, 5, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2LDRHs, ARM::tLDRHr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2LDRSBs,ARM::tLDRSB, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2LDRSHs,ARM::tLDRSH, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2STRi12,ARM::tSTRi, ARM::tSTRspi, 5, 8, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2STRs, ARM::tSTRr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2STRBi12,ARM::tSTRBi, 0, 5, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2STRBs, ARM::tSTRBr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2STRHi12,ARM::tSTRHi, 0, 5, 0, 1, 0, 0,0, 0,1,0 },
+ { ARM::t2STRHs, ARM::tSTRHr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
+
+ { ARM::t2LDMIA, ARM::tLDMIA, 0, 0, 0, 1, 1, 1,1, 0,1,0 },
+ { ARM::t2LDMIA_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 0,1,0 },
+ { ARM::t2LDMIA_UPD,ARM::tLDMIA_UPD,ARM::tPOP,0, 0, 1, 1, 1,1, 0,1,0 },
+ // ARM::t2STM (with no basereg writeback) has no Thumb1 equivalent
+ { ARM::t2STMIA_UPD,ARM::tSTMIA_UPD, 0, 0, 0, 1, 1, 1,1, 0,1,0 },
+ { ARM::t2STMDB_UPD, 0, ARM::tPUSH, 0, 0, 1, 1, 1,1, 0,1,0 }
};
class Thumb2SizeReduce : public MachineFunctionPass {
/// ReduceOpcodeMap - Maps wide opcode to index of entry in ReduceTable.
DenseMap<unsigned, unsigned> ReduceOpcodeMap;
- bool canAddPseudoFlagDep(MachineInstr *Def, MachineInstr *Use);
+ bool canAddPseudoFlagDep(MachineInstr *Use, bool IsSelfLoop);
bool VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
bool is2Addr, ARMCC::CondCodes Pred,
const ReduceEntry &Entry);
bool ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
- const ReduceEntry &Entry, bool LiveCPSR,
- MachineInstr *CPSRDef);
+ const ReduceEntry &Entry, bool LiveCPSR, bool IsSelfLoop);
/// ReduceTo2Addr - Reduce a 32-bit instruction to a 16-bit two-address
/// instruction.
bool ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
- const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef);
+ const ReduceEntry &Entry, bool LiveCPSR,
+ bool IsSelfLoop);
/// ReduceToNarrow - Reduce a 32-bit instruction to a 16-bit
/// non-two-address instruction.
bool ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
- const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef);
+ const ReduceEntry &Entry, bool LiveCPSR,
+ bool IsSelfLoop);
+
+ /// ReduceMI - Attempt to reduce MI, return true on success.
+ bool ReduceMI(MachineBasicBlock &MBB, MachineInstr *MI,
+ bool LiveCPSR, bool IsSelfLoop);
/// ReduceMBB - Reduce width of instructions in the specified basic block.
bool ReduceMBB(MachineBasicBlock &MBB);
+
+ bool OptimizeSize;
+ bool MinimizeSize;
+
+ // Last instruction to define CPSR in the current block.
+ MachineInstr *CPSRDef;
+ // Was CPSR last defined by a high latency instruction?
+ // When CPSRDef is null, this refers to CPSR defs in predecessors.
+ bool HighLatencyCPSR;
+
+ struct MBBInfo {
+ // The flags leaving this block have high latency.
+ bool HighLatencyCPSR;
+ // Has this block been visited yet?
+ bool Visited;
+
+ MBBInfo() : HighLatencyCPSR(false), Visited(false) {}
+ };
+
+ SmallVector<MBBInfo, 8> BlockInfo;
};
char Thumb2SizeReduce::ID = 0;
}
Thumb2SizeReduce::Thumb2SizeReduce() : MachineFunctionPass(ID) {
+ OptimizeSize = MinimizeSize = false;
for (unsigned i = 0, e = array_lengthof(ReduceTable); i != e; ++i) {
unsigned FromOpc = ReduceTable[i].WideOpc;
if (!ReduceOpcodeMap.insert(std::make_pair(FromOpc, i)).second)
}
static bool HasImplicitCPSRDef(const MCInstrDesc &MCID) {
- for (const unsigned *Regs = MCID.ImplicitDefs; *Regs; ++Regs)
+ for (const uint16_t *Regs = MCID.getImplicitDefs(); *Regs; ++Regs)
if (*Regs == ARM::CPSR)
return true;
return false;
}
+// Check for a likely high-latency flag def.
+static bool isHighLatencyCPSR(MachineInstr *Def) {
+ switch(Def->getOpcode()) {
+ case ARM::FMSTAT:
+ case ARM::tMUL:
+ return true;
+ }
+ return false;
+}
+
/// canAddPseudoFlagDep - For A9 (and other out-of-order) implementations,
/// the 's' 16-bit instruction partially update CPSR. Abort the
/// transformation to avoid adding false dependency on last CPSR setting
/// In this case it would have been ok to narrow the mul.w to muls since there
/// are indirect RAW dependency between the muls and the mul.w
bool
-Thumb2SizeReduce::canAddPseudoFlagDep(MachineInstr *Def, MachineInstr *Use) {
- if (!Def || !STI->avoidCPSRPartialUpdate())
+Thumb2SizeReduce::canAddPseudoFlagDep(MachineInstr *Use, bool FirstInSelfLoop) {
+ // Disable the check for -Oz (aka OptimizeForSizeHarder).
+ if (MinimizeSize || !STI->avoidCPSRPartialUpdate())
return false;
+ if (!CPSRDef)
+ // If this BB loops back to itself, conservatively avoid narrowing the
+ // first instruction that does partial flag update.
+ return HighLatencyCPSR || FirstInSelfLoop;
+
SmallSet<unsigned, 2> Defs;
- for (unsigned i = 0, e = Def->getNumOperands(); i != e; ++i) {
- const MachineOperand &MO = Def->getOperand(i);
+ for (unsigned i = 0, e = CPSRDef->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = CPSRDef->getOperand(i);
if (!MO.isReg() || MO.isUndef() || MO.isUse())
continue;
unsigned Reg = MO.getReg();
return false;
}
+ // If the current CPSR has high latency, try to avoid the false dependency.
+ if (HighLatencyCPSR)
+ return true;
+
+ // tMOVi8 usually doesn't start long dependency chains, and there are a lot
+ // of them, so always shrink them when CPSR doesn't have high latency.
+ if (Use->getOpcode() == ARM::t2MOVi ||
+ Use->getOpcode() == ARM::t2MOVi16)
+ return false;
+
// No read-after-write dependency. The narrowing will add false dependency.
return true;
}
// Add the 16-bit load / store instruction.
DebugLoc dl = MI->getDebugLoc();
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, TII->get(Opc));
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, TII->get(Opc));
if (!isLdStMul) {
MIB.addOperand(MI->getOperand(0));
MIB.addOperand(MI->getOperand(1));
DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
- MBB.erase(MI);
+ MBB.erase_instr(MI);
++NumLdSts;
return true;
}
bool
Thumb2SizeReduce::ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef) {
+ bool LiveCPSR, bool IsSelfLoop) {
unsigned Opc = MI->getOpcode();
if (Opc == ARM::t2ADDri) {
// If the source register is SP, try to reduce to tADDrSPi, otherwise
// it's a normal reduce.
if (MI->getOperand(1).getReg() != ARM::SP) {
- if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef))
+ if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop))
return true;
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
}
// Try to reduce to tADDrSPi.
unsigned Imm = MI->getOperand(2).getImm();
// The immediate must be in range, the destination register must be a low
- // reg, and the condition flags must not be being set.
- if (Imm & 3 || Imm > 1024)
+ // reg, the predicate must be "always" and the condition flags must not
+ // be being set.
+ if (Imm & 3 || Imm > 1020)
return false;
if (!isARMLowRegister(MI->getOperand(0).getReg()))
return false;
+ if (MI->getOperand(3).getImm() != ARMCC::AL)
+ return false;
const MCInstrDesc &MCID = MI->getDesc();
if (MCID.hasOptionalDef() &&
MI->getOperand(MCID.getNumOperands()-1).getReg() == ARM::CPSR)
return false;
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, MI->getDebugLoc(),
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, MI->getDebugLoc(),
TII->get(ARM::tADDrSPi))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1))
.addImm(Imm / 4); // The tADDrSPi has an implied scale by four.
+ AddDefaultPred(MIB);
// Transfer MI flags.
MIB.setMIFlags(MI->getFlags());
DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " <<*MIB);
- MBB.erase(MI);
+ MBB.erase_instr(MI);
++NumNarrows;
return true;
}
if (Entry.LowRegs1 && !VerifyLowRegs(MI))
return false;
- const MCInstrDesc &MCID = MI->getDesc();
- if (MCID.mayLoad() || MCID.mayStore())
+ if (MI->mayLoad() || MI->mayStore())
return ReduceLoadStore(MBB, MI, Entry);
switch (Opc) {
switch (Opc) {
default: break;
case ARM::t2ADDSri: {
- if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef))
+ if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop))
return true;
// fallthrough
}
case ARM::t2ADDSrr:
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
}
}
break;
}
case ARM::t2RSBri:
case ARM::t2RSBSri:
+ case ARM::t2SXTB:
+ case ARM::t2SXTH:
+ case ARM::t2UXTB:
+ case ARM::t2UXTH:
if (MI->getOperand(2).getImm() == 0)
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
break;
case ARM::t2MOVi16:
// Can convert only 'pure' immediate operands, not immediates obtained as
// globals' addresses.
if (MI->getOperand(1).isImm())
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
break;
case ARM::t2CMPrr: {
// Try to reduce to the lo-reg only version first. Why there are two
// are prioritized, but the table assumes a unique entry for each
// source insn opcode. So for now, we hack a local entry record to use.
static const ReduceEntry NarrowEntry =
- { ARM::t2CMPrr,ARM::tCMPr, 0, 0, 0, 1, 1,2, 0, 0,1 };
- if (ReduceToNarrow(MBB, MI, NarrowEntry, LiveCPSR, CPSRDef))
+ { ARM::t2CMPrr,ARM::tCMPr, 0, 0, 0, 1, 1,2, 0, 0,1,0 };
+ if (ReduceToNarrow(MBB, MI, NarrowEntry, LiveCPSR, IsSelfLoop))
return true;
- return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef);
+ return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
}
}
return false;
bool
Thumb2SizeReduce::ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef) {
+ bool LiveCPSR, bool IsSelfLoop) {
if (ReduceLimit2Addr != -1 && ((int)Num2Addrs >= ReduceLimit2Addr))
return false;
+ if (!MinimizeSize && !OptimizeSize && Entry.AvoidMovs &&
+ STI->avoidMOVsShifterOperand())
+ // Don't issue movs with shifter operand for some CPUs unless we
+ // are optimizing / minimizing for size.
+ return false;
+
unsigned Reg0 = MI->getOperand(0).getReg();
unsigned Reg1 = MI->getOperand(1).getReg();
- if (Reg0 != Reg1) {
+ // t2MUL is "special". The tied source operand is second, not first.
+ if (MI->getOpcode() == ARM::t2MUL) {
+ unsigned Reg2 = MI->getOperand(2).getReg();
+ // Early exit if the regs aren't all low regs.
+ if (!isARMLowRegister(Reg0) || !isARMLowRegister(Reg1)
+ || !isARMLowRegister(Reg2))
+ return false;
+ if (Reg0 != Reg2) {
+ // If the other operand also isn't the same as the destination, we
+ // can't reduce.
+ if (Reg1 != Reg0)
+ return false;
+ // Try to commute the operands to make it a 2-address instruction.
+ MachineInstr *CommutedMI = TII->commuteInstruction(MI);
+ if (!CommutedMI)
+ return false;
+ }
+ } else if (Reg0 != Reg1) {
// Try to commute the operands to make it a 2-address instruction.
unsigned CommOpIdx1, CommOpIdx2;
if (!TII->findCommutedOpIndices(MI, CommOpIdx1, CommOpIdx2) ||
// Avoid adding a false dependency on partial flag update by some 16-bit
// instructions which has the 's' bit set.
if (Entry.PartFlag && NewMCID.hasOptionalDef() && HasCC &&
- canAddPseudoFlagDep(CPSRDef, MI))
+ canAddPseudoFlagDep(MI, IsSelfLoop))
return false;
// Add the 16-bit instruction.
DebugLoc dl = MI->getDebugLoc();
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, NewMCID);
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, NewMCID);
MIB.addOperand(MI->getOperand(0));
if (NewMCID.hasOptionalDef()) {
if (HasCC)
DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
- MBB.erase(MI);
+ MBB.erase_instr(MI);
++Num2Addrs;
return true;
}
bool
Thumb2SizeReduce::ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
const ReduceEntry &Entry,
- bool LiveCPSR, MachineInstr *CPSRDef) {
+ bool LiveCPSR, bool IsSelfLoop) {
if (ReduceLimit != -1 && ((int)NumNarrows >= ReduceLimit))
return false;
+ if (!MinimizeSize && !OptimizeSize && Entry.AvoidMovs &&
+ STI->avoidMOVsShifterOperand())
+ // Don't issue movs with shifter operand for some CPUs unless we
+ // are optimizing / minimizing for size.
+ return false;
+
unsigned Limit = ~0U;
if (Entry.Imm1Limit)
Limit = (1 << Entry.Imm1Limit) - 1;
// Avoid adding a false dependency on partial flag update by some 16-bit
// instructions which has the 's' bit set.
if (Entry.PartFlag && NewMCID.hasOptionalDef() && HasCC &&
- canAddPseudoFlagDep(CPSRDef, MI))
+ canAddPseudoFlagDep(MI, IsSelfLoop))
return false;
// Add the 16-bit instruction.
DebugLoc dl = MI->getDebugLoc();
- MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, NewMCID);
+ MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, NewMCID);
MIB.addOperand(MI->getOperand(0));
if (NewMCID.hasOptionalDef()) {
if (HasCC)
if (i < NumOps && MCID.OpInfo[i].isOptionalDef())
continue;
if ((MCID.getOpcode() == ARM::t2RSBSri ||
- MCID.getOpcode() == ARM::t2RSBri) && i == 2)
+ MCID.getOpcode() == ARM::t2RSBri ||
+ MCID.getOpcode() == ARM::t2SXTB ||
+ MCID.getOpcode() == ARM::t2SXTH ||
+ MCID.getOpcode() == ARM::t2UXTB ||
+ MCID.getOpcode() == ARM::t2UXTH) && i == 2)
// Skip the zero immediate operand, it's now implicit.
continue;
bool isPred = (i < NumOps && MCID.OpInfo[i].isPredicate());
DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
- MBB.erase(MI);
+ MBB.erase_instr(MI);
++NumNarrows;
return true;
}
return LiveCPSR;
}
+bool Thumb2SizeReduce::ReduceMI(MachineBasicBlock &MBB, MachineInstr *MI,
+ bool LiveCPSR, bool IsSelfLoop) {
+ unsigned Opcode = MI->getOpcode();
+ DenseMap<unsigned, unsigned>::iterator OPI = ReduceOpcodeMap.find(Opcode);
+ if (OPI == ReduceOpcodeMap.end())
+ return false;
+ const ReduceEntry &Entry = ReduceTable[OPI->second];
+
+ // Don't attempt normal reductions on "special" cases for now.
+ if (Entry.Special)
+ return ReduceSpecial(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
+
+ // Try to transform to a 16-bit two-address instruction.
+ if (Entry.NarrowOpc2 &&
+ ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop))
+ return true;
+
+ // Try to transform to a 16-bit non-two-address instruction.
+ if (Entry.NarrowOpc1 &&
+ ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop))
+ return true;
+
+ return false;
+}
+
bool Thumb2SizeReduce::ReduceMBB(MachineBasicBlock &MBB) {
bool Modified = false;
// Yes, CPSR could be livein.
bool LiveCPSR = MBB.isLiveIn(ARM::CPSR);
- MachineInstr *CPSRDef = 0;
+ MachineInstr *BundleMI = 0;
+
+ CPSRDef = 0;
+ HighLatencyCPSR = false;
+
+ // Check predecessors for the latest CPSRDef.
+ for (MachineBasicBlock::pred_iterator
+ I = MBB.pred_begin(), E = MBB.pred_end(); I != E; ++I) {
+ const MBBInfo &PInfo = BlockInfo[(*I)->getNumber()];
+ if (!PInfo.Visited) {
+ // Since blocks are visited in RPO, this must be a back-edge.
+ continue;
+ }
+ if (PInfo.HighLatencyCPSR) {
+ HighLatencyCPSR = true;
+ break;
+ }
+ }
- MachineBasicBlock::iterator MII = MBB.begin(), E = MBB.end();
- MachineBasicBlock::iterator NextMII;
+ // If this BB loops back to itself, conservatively avoid narrowing the
+ // first instruction that does partial flag update.
+ bool IsSelfLoop = MBB.isSuccessor(&MBB);
+ MachineBasicBlock::instr_iterator MII = MBB.instr_begin(),E = MBB.instr_end();
+ MachineBasicBlock::instr_iterator NextMII;
for (; MII != E; MII = NextMII) {
- NextMII = llvm::next(MII);
+ NextMII = std::next(MII);
MachineInstr *MI = &*MII;
- LiveCPSR = UpdateCPSRUse(*MI, LiveCPSR);
+ if (MI->isBundle()) {
+ BundleMI = MI;
+ continue;
+ }
+ if (MI->isDebugValue())
+ continue;
- unsigned Opcode = MI->getOpcode();
- DenseMap<unsigned, unsigned>::iterator OPI = ReduceOpcodeMap.find(Opcode);
- if (OPI != ReduceOpcodeMap.end()) {
- const ReduceEntry &Entry = ReduceTable[OPI->second];
- // Ignore "special" cases for now.
- if (Entry.Special) {
- if (ReduceSpecial(MBB, MI, Entry, LiveCPSR, CPSRDef)) {
- Modified = true;
- MachineBasicBlock::iterator I = prior(NextMII);
- MI = &*I;
- }
- goto ProcessNext;
- }
+ LiveCPSR = UpdateCPSRUse(*MI, LiveCPSR);
- // Try to transform to a 16-bit two-address instruction.
- if (Entry.NarrowOpc2 &&
- ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, CPSRDef)) {
- Modified = true;
- MachineBasicBlock::iterator I = prior(NextMII);
- MI = &*I;
- goto ProcessNext;
- }
+ // Does NextMII belong to the same bundle as MI?
+ bool NextInSameBundle = NextMII != E && NextMII->isBundledWithPred();
+
+ if (ReduceMI(MBB, MI, LiveCPSR, IsSelfLoop)) {
+ Modified = true;
+ MachineBasicBlock::instr_iterator I = std::prev(NextMII);
+ MI = &*I;
+ // Removing and reinserting the first instruction in a bundle will break
+ // up the bundle. Fix the bundling if it was broken.
+ if (NextInSameBundle && !NextMII->isBundledWithPred())
+ NextMII->bundleWithPred();
+ }
- // Try to transform to a 16-bit non-two-address instruction.
- if (Entry.NarrowOpc1 &&
- ReduceToNarrow(MBB, MI, Entry, LiveCPSR, CPSRDef)) {
- Modified = true;
- MachineBasicBlock::iterator I = prior(NextMII);
- MI = &*I;
- }
+ if (!NextInSameBundle && MI->isInsideBundle()) {
+ // FIXME: Since post-ra scheduler operates on bundles, the CPSR kill
+ // marker is only on the BUNDLE instruction. Process the BUNDLE
+ // instruction as we finish with the bundled instruction to work around
+ // the inconsistency.
+ if (BundleMI->killsRegister(ARM::CPSR))
+ LiveCPSR = false;
+ MachineOperand *MO = BundleMI->findRegisterDefOperand(ARM::CPSR);
+ if (MO && !MO->isDead())
+ LiveCPSR = true;
+ MO = BundleMI->findRegisterUseOperand(ARM::CPSR);
+ if (MO && !MO->isKill())
+ LiveCPSR = true;
}
- ProcessNext:
bool DefCPSR = false;
LiveCPSR = UpdateCPSRDef(*MI, LiveCPSR, DefCPSR);
- if (MI->getDesc().isCall())
+ if (MI->isCall()) {
// Calls don't really set CPSR.
CPSRDef = 0;
- else if (DefCPSR)
+ HighLatencyCPSR = false;
+ IsSelfLoop = false;
+ } else if (DefCPSR) {
// This is the last CPSR defining instruction.
CPSRDef = MI;
+ HighLatencyCPSR = isHighLatencyCPSR(CPSRDef);
+ IsSelfLoop = false;
+ }
}
+ MBBInfo &Info = BlockInfo[MBB.getNumber()];
+ Info.HighLatencyCPSR = HighLatencyCPSR;
+ Info.Visited = true;
return Modified;
}
TII = static_cast<const Thumb2InstrInfo*>(TM.getInstrInfo());
STI = &TM.getSubtarget<ARMSubtarget>();
+ // Optimizing / minimizing size?
+ AttributeSet FnAttrs = MF.getFunction()->getAttributes();
+ OptimizeSize = FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
+ Attribute::OptimizeForSize);
+ MinimizeSize = STI->isMinSize();
+
+ BlockInfo.clear();
+ BlockInfo.resize(MF.getNumBlockIDs());
+
+ // Visit blocks in reverse post-order so LastCPSRDef is known for all
+ // predecessors.
+ ReversePostOrderTraversal<MachineFunction*> RPOT(&MF);
bool Modified = false;
- for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
- Modified |= ReduceMBB(*I);
+ for (ReversePostOrderTraversal<MachineFunction*>::rpo_iterator
+ I = RPOT.begin(), E = RPOT.end(); I != E; ++I)
+ Modified |= ReduceMBB(**I);
return Modified;
}
projects / oota-llvm.git / blobdiff