SparcInstr64Bit.td [plain text]
//===-- SparcInstr64Bit.td - 64-bit instructions for Sparc Target ---------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file contains instruction definitions and patterns needed for 64-bit
// code generation on SPARC v9.
//
// Some SPARC v9 instructions are defined in SparcInstrInfo.td because they can
// also be used in 32-bit code running on a SPARC v9 CPU.
//
//===----------------------------------------------------------------------===//
let Predicates = [Is64Bit] in {
// The same integer registers are used for i32 and i64 values.
// When registers hold i32 values, the high bits are don't care.
// This give us free trunc and anyext.
def : Pat<(i64 (anyext i32:$val)), (COPY_TO_REGCLASS $val, I64Regs)>;
def : Pat<(i32 (trunc i64:$val)), (COPY_TO_REGCLASS $val, IntRegs)>;
} // Predicates = [Is64Bit]
//===----------------------------------------------------------------------===//
// 64-bit Shift Instructions.
//===----------------------------------------------------------------------===//
//
// The 32-bit shift instructions are still available. The left shift srl
// instructions shift all 64 bits, but it only accepts a 5-bit shift amount.
//
// The srl instructions only shift the low 32 bits and clear the high 32 bits.
// Finally, sra shifts the low 32 bits and sign-extends to 64 bits.
let Predicates = [Is64Bit] in {
def : Pat<(i64 (zext i32:$val)), (SRLri $val, 0)>;
def : Pat<(i64 (sext i32:$val)), (SRAri $val, 0)>;
def : Pat<(i64 (and i64:$val, 0xffffffff)), (SRLri $val, 0)>;
def : Pat<(i64 (sext_inreg i64:$val, i32)), (SRAri $val, 0)>;
defm SLLX : F3_S<"sllx", 0b100101, 1, shl, i64, I64Regs>;
defm SRLX : F3_S<"srlx", 0b100110, 1, srl, i64, I64Regs>;
defm SRAX : F3_S<"srax", 0b100111, 1, sra, i64, I64Regs>;
} // Predicates = [Is64Bit]
//===----------------------------------------------------------------------===//
// 64-bit Immediates.
//===----------------------------------------------------------------------===//
//
// All 32-bit immediates can be materialized with sethi+or, but 64-bit
// immediates may require more code. There may be a point where it is
// preferable to use a constant pool load instead, depending on the
// microarchitecture.
// Single-instruction patterns.
// The ALU instructions want their simm13 operands as i32 immediates.
def as_i32imm : SDNodeXForm<imm, [{
return CurDAG->getTargetConstant(N->getSExtValue(), MVT::i32);
}]>;
def : Pat<(i64 simm13:$val), (ORri (i64 G0), (as_i32imm $val))>;
def : Pat<(i64 SETHIimm:$val), (SETHIi (HI22 $val))>;
// Double-instruction patterns.
// All unsigned i32 immediates can be handled by sethi+or.
def uimm32 : PatLeaf<(imm), [{ return isUInt<32>(N->getZExtValue()); }]>;
def : Pat<(i64 uimm32:$val), (ORri (SETHIi (HI22 $val)), (LO10 $val))>,
Requires<[Is64Bit]>;
// All negative i33 immediates can be handled by sethi+xor.
def nimm33 : PatLeaf<(imm), [{
int64_t Imm = N->getSExtValue();
return Imm < 0 && isInt<33>(Imm);
}]>;
// Bits 10-31 inverted. Same as assembler's %hix.
def HIX22 : SDNodeXForm<imm, [{
uint64_t Val = (~N->getZExtValue() >> 10) & ((1u << 22) - 1);
return CurDAG->getTargetConstant(Val, MVT::i32);
}]>;
// Bits 0-9 with ones in bits 10-31. Same as assembler's %lox.
def LOX10 : SDNodeXForm<imm, [{
return CurDAG->getTargetConstant(~(~N->getZExtValue() & 0x3ff), MVT::i32);
}]>;
def : Pat<(i64 nimm33:$val), (XORri (SETHIi (HIX22 $val)), (LOX10 $val))>,
Requires<[Is64Bit]>;
// More possible patterns:
//
// (sllx sethi, n)
// (sllx simm13, n)
//
// 3 instrs:
//
// (xor (sllx sethi), simm13)
// (sllx (xor sethi, simm13))
//
// 4 instrs:
//
// (or sethi, (sllx sethi))
// (xnor sethi, (sllx sethi))
//
// 5 instrs:
//
// (or (sllx sethi), (or sethi, simm13))
// (xnor (sllx sethi), (or sethi, simm13))
// (or (sllx sethi), (sllx sethi))
// (xnor (sllx sethi), (sllx sethi))
//
// Worst case is 6 instrs:
//
// (or (sllx (or sethi, simmm13)), (or sethi, simm13))
// Bits 42-63, same as assembler's %hh.
def HH22 : SDNodeXForm<imm, [{
uint64_t Val = (N->getZExtValue() >> 42) & ((1u << 22) - 1);
return CurDAG->getTargetConstant(Val, MVT::i32);
}]>;
// Bits 32-41, same as assembler's %hm.
def HM10 : SDNodeXForm<imm, [{
uint64_t Val = (N->getZExtValue() >> 32) & ((1u << 10) - 1);
return CurDAG->getTargetConstant(Val, MVT::i32);
}]>;
def : Pat<(i64 imm:$val),
(ORrr (SLLXri (ORri (SETHIi (HH22 $val)), (HM10 $val)), (i32 32)),
(ORri (SETHIi (HI22 $val)), (LO10 $val)))>,
Requires<[Is64Bit]>;
//===----------------------------------------------------------------------===//
// 64-bit Integer Arithmetic and Logic.
//===----------------------------------------------------------------------===//
let Predicates = [Is64Bit] in {
// Register-register instructions.
let isCodeGenOnly = 1 in {
defm ANDX : F3_12<"and", 0b000001, and, I64Regs, i64, i64imm>;
defm ORX : F3_12<"or", 0b000010, or, I64Regs, i64, i64imm>;
defm XORX : F3_12<"xor", 0b000011, xor, I64Regs, i64, i64imm>;
def ANDXNrr : F3_1<2, 0b000101,
(outs I64Regs:$dst), (ins I64Regs:$b, I64Regs:$c),
"andn $b, $c, $dst",
[(set i64:$dst, (and i64:$b, (not i64:$c)))]>;
def ORXNrr : F3_1<2, 0b000110,
(outs I64Regs:$dst), (ins I64Regs:$b, I64Regs:$c),
"orn $b, $c, $dst",
[(set i64:$dst, (or i64:$b, (not i64:$c)))]>;
def XNORXrr : F3_1<2, 0b000111,
(outs I64Regs:$dst), (ins I64Regs:$b, I64Regs:$c),
"xnor $b, $c, $dst",
[(set i64:$dst, (not (xor i64:$b, i64:$c)))]>;
defm ADDX : F3_12<"add", 0b000000, add, I64Regs, i64, i64imm>;
defm SUBX : F3_12<"sub", 0b000100, sub, I64Regs, i64, i64imm>;
def TLS_ADDXrr : F3_1<2, 0b000000, (outs I64Regs:$rd),
(ins I64Regs:$rs1, I64Regs:$rs2, TLSSym:$sym),
"add $rs1, $rs2, $rd, $sym",
[(set i64:$rd,
(tlsadd i64:$rs1, i64:$rs2, tglobaltlsaddr:$sym))]>;
// "LEA" form of add
def LEAX_ADDri : F3_2<2, 0b000000,
(outs I64Regs:$dst), (ins MEMri:$addr),
"add ${addr:arith}, $dst",
[(set iPTR:$dst, ADDRri:$addr)]>;
}
def : Pat<(SPcmpicc i64:$a, i64:$b), (CMPrr $a, $b)>;
def : Pat<(SPcmpicc i64:$a, (i64 simm13:$b)), (CMPri $a, (as_i32imm $b))>;
def : Pat<(ctpop i64:$src), (POPCrr $src)>;
} // Predicates = [Is64Bit]
//===----------------------------------------------------------------------===//
// 64-bit Integer Multiply and Divide.
//===----------------------------------------------------------------------===//
let Predicates = [Is64Bit] in {
def MULXrr : F3_1<2, 0b001001,
(outs I64Regs:$rd), (ins I64Regs:$rs1, I64Regs:$rs2),
"mulx $rs1, $rs2, $rd",
[(set i64:$rd, (mul i64:$rs1, i64:$rs2))]>;
def MULXri : F3_2<2, 0b001001,
(outs IntRegs:$rd), (ins IntRegs:$rs1, i64imm:$simm13),
"mulx $rs1, $simm13, $rd",
[(set i64:$rd, (mul i64:$rs1, (i64 simm13:$simm13)))]>;
// Division can trap.
let hasSideEffects = 1 in {
def SDIVXrr : F3_1<2, 0b101101,
(outs I64Regs:$rd), (ins I64Regs:$rs1, I64Regs:$rs2),
"sdivx $rs1, $rs2, $rd",
[(set i64:$rd, (sdiv i64:$rs1, i64:$rs2))]>;
def SDIVXri : F3_2<2, 0b101101,
(outs IntRegs:$rd), (ins IntRegs:$rs1, i64imm:$simm13),
"sdivx $rs1, $simm13, $rd",
[(set i64:$rd, (sdiv i64:$rs1, (i64 simm13:$simm13)))]>;
def UDIVXrr : F3_1<2, 0b001101,
(outs I64Regs:$rd), (ins I64Regs:$rs1, I64Regs:$rs2),
"udivx $rs1, $rs2, $rd",
[(set i64:$rd, (udiv i64:$rs1, i64:$rs2))]>;
def UDIVXri : F3_2<2, 0b001101,
(outs IntRegs:$rd), (ins IntRegs:$rs1, i64imm:$simm13),
"udivx $rs1, $simm13, $rd",
[(set i64:$rd, (udiv i64:$rs1, (i64 simm13:$simm13)))]>;
} // hasSideEffects = 1
} // Predicates = [Is64Bit]
//===----------------------------------------------------------------------===//
// 64-bit Loads and Stores.
//===----------------------------------------------------------------------===//
//
// All the 32-bit loads and stores are available. The extending loads are sign
// or zero-extending to 64 bits. The LDrr and LDri instructions load 32 bits
// zero-extended to i64. Their mnemonic is lduw in SPARC v9 (Load Unsigned
// Word).
//
// SPARC v9 adds 64-bit loads as well as a sign-extending ldsw i32 loads.
let Predicates = [Is64Bit] in {
// 64-bit loads.
defm LDX : Load<"ldx", 0b001011, load, I64Regs, i64>;
let mayLoad = 1, isCodeGenOnly = 1, isAsmParserOnly = 1 in
def TLS_LDXrr : F3_1<3, 0b001011,
(outs IntRegs:$dst), (ins MEMrr:$addr, TLSSym:$sym),
"ldx [$addr], $dst, $sym",
[(set i64:$dst,
(tlsld ADDRrr:$addr, tglobaltlsaddr:$sym))]>;
// Extending loads to i64.
def : Pat<(i64 (zextloadi1 ADDRrr:$addr)), (LDUBrr ADDRrr:$addr)>;
def : Pat<(i64 (zextloadi1 ADDRri:$addr)), (LDUBri ADDRri:$addr)>;
def : Pat<(i64 (extloadi1 ADDRrr:$addr)), (LDUBrr ADDRrr:$addr)>;
def : Pat<(i64 (extloadi1 ADDRri:$addr)), (LDUBri ADDRri:$addr)>;
def : Pat<(i64 (zextloadi8 ADDRrr:$addr)), (LDUBrr ADDRrr:$addr)>;
def : Pat<(i64 (zextloadi8 ADDRri:$addr)), (LDUBri ADDRri:$addr)>;
def : Pat<(i64 (extloadi8 ADDRrr:$addr)), (LDUBrr ADDRrr:$addr)>;
def : Pat<(i64 (extloadi8 ADDRri:$addr)), (LDUBri ADDRri:$addr)>;
def : Pat<(i64 (sextloadi8 ADDRrr:$addr)), (LDSBrr ADDRrr:$addr)>;
def : Pat<(i64 (sextloadi8 ADDRri:$addr)), (LDSBri ADDRri:$addr)>;
def : Pat<(i64 (zextloadi16 ADDRrr:$addr)), (LDUHrr ADDRrr:$addr)>;
def : Pat<(i64 (zextloadi16 ADDRri:$addr)), (LDUHri ADDRri:$addr)>;
def : Pat<(i64 (extloadi16 ADDRrr:$addr)), (LDUHrr ADDRrr:$addr)>;
def : Pat<(i64 (extloadi16 ADDRri:$addr)), (LDUHri ADDRri:$addr)>;
def : Pat<(i64 (sextloadi16 ADDRrr:$addr)), (LDSHrr ADDRrr:$addr)>;
def : Pat<(i64 (sextloadi16 ADDRri:$addr)), (LDSHri ADDRri:$addr)>;
def : Pat<(i64 (zextloadi32 ADDRrr:$addr)), (LDrr ADDRrr:$addr)>;
def : Pat<(i64 (zextloadi32 ADDRri:$addr)), (LDri ADDRri:$addr)>;
def : Pat<(i64 (extloadi32 ADDRrr:$addr)), (LDrr ADDRrr:$addr)>;
def : Pat<(i64 (extloadi32 ADDRri:$addr)), (LDri ADDRri:$addr)>;
// Sign-extending load of i32 into i64 is a new SPARC v9 instruction.
defm LDSW : Load<"ldsw", 0b001000, sextloadi32, I64Regs, i64>;
// 64-bit stores.
defm STX : Store<"stx", 0b001110, store, I64Regs, i64>;
// Truncating stores from i64 are identical to the i32 stores.
def : Pat<(truncstorei8 i64:$src, ADDRrr:$addr), (STBrr ADDRrr:$addr, $src)>;
def : Pat<(truncstorei8 i64:$src, ADDRri:$addr), (STBri ADDRri:$addr, $src)>;
def : Pat<(truncstorei16 i64:$src, ADDRrr:$addr), (STHrr ADDRrr:$addr, $src)>;
def : Pat<(truncstorei16 i64:$src, ADDRri:$addr), (STHri ADDRri:$addr, $src)>;
def : Pat<(truncstorei32 i64:$src, ADDRrr:$addr), (STrr ADDRrr:$addr, $src)>;
def : Pat<(truncstorei32 i64:$src, ADDRri:$addr), (STri ADDRri:$addr, $src)>;
// store 0, addr -> store %g0, addr
def : Pat<(store (i64 0), ADDRrr:$dst), (STXrr ADDRrr:$dst, (i64 G0))>;
def : Pat<(store (i64 0), ADDRri:$dst), (STXri ADDRri:$dst, (i64 G0))>;
} // Predicates = [Is64Bit]
//===----------------------------------------------------------------------===//
// 64-bit Conditionals.
//===----------------------------------------------------------------------===//
// Conditional branch class on %xcc:
class XBranchSP<dag ins, string asmstr, list<dag> pattern>
: F2_3<0b001, 0b10, (outs), ins, asmstr, pattern> {
let isBranch = 1;
let isTerminator = 1;
let hasDelaySlot = 1;
}
//
// Flag-setting instructions like subcc and addcc set both icc and xcc flags.
// The icc flags correspond to the 32-bit result, and the xcc are for the
// full 64-bit result.
//
// We reuse CMPICC SDNodes for compares, but use new BRXCC branch nodes for
// 64-bit compares. See LowerBR_CC.
let Predicates = [Is64Bit] in {
let Uses = [ICC] in
def BPXCC : XBranchSP<(ins brtarget:$imm19, CCOp:$cond),
"b$cond %xcc, $imm19",
[(SPbrxcc bb:$imm19, imm:$cond)]>;
// Conditional moves on %xcc.
let Uses = [ICC], Constraints = "$f = $rd" in {
let cc = 0b110 in {
def MOVXCCrr : F4_1<0b101100, (outs IntRegs:$rd),
(ins IntRegs:$rs2, IntRegs:$f, CCOp:$cond),
"mov$cond %xcc, $rs2, $rd",
[(set i32:$rd,
(SPselectxcc i32:$rs2, i32:$f, imm:$cond))]>;
def MOVXCCri : F4_2<0b101100, (outs IntRegs:$rd),
(ins i32imm:$simm11, IntRegs:$f, CCOp:$cond),
"mov$cond %xcc, $simm11, $rd",
[(set i32:$rd,
(SPselectxcc simm11:$simm11, i32:$f, imm:$cond))]>;
} // cc
let opf_cc = 0b110 in {
def FMOVS_XCC : F4_3<0b110101, 0b000001, (outs FPRegs:$rd),
(ins FPRegs:$rs2, FPRegs:$f, CCOp:$cond),
"fmovs$cond %xcc, $rs2, $rd",
[(set f32:$rd,
(SPselectxcc f32:$rs2, f32:$f, imm:$cond))]>;
def FMOVD_XCC : F4_3<0b110101, 0b000010, (outs DFPRegs:$rd),
(ins DFPRegs:$rs2, DFPRegs:$f, CCOp:$cond),
"fmovd$cond %xcc, $rs2, $rd",
[(set f64:$rd,
(SPselectxcc f64:$rs2, f64:$f, imm:$cond))]>;
def FMOVQ_XCC : F4_3<0b110101, 0b000011, (outs QFPRegs:$rd),
(ins QFPRegs:$rs2, QFPRegs:$f, CCOp:$cond),
"fmovq$cond %xcc, $rs2, $rd",
[(set f128:$rd,
(SPselectxcc f128:$rs2, f128:$f, imm:$cond))]>;
} // opf_cc
} // Uses, Constraints
//===----------------------------------------------------------------------===//
// 64-bit Floating Point Conversions.
//===----------------------------------------------------------------------===//
let Predicates = [Is64Bit] in {
def FXTOS : F3_3u<2, 0b110100, 0b010000100,
(outs FPRegs:$rd), (ins DFPRegs:$rs2),
"fxtos $rs2, $rd",
[(set FPRegs:$rd, (SPxtof DFPRegs:$rs2))]>;
def FXTOD : F3_3u<2, 0b110100, 0b010001000,
(outs DFPRegs:$rd), (ins DFPRegs:$rs2),
"fxtod $rs2, $rd",
[(set DFPRegs:$rd, (SPxtof DFPRegs:$rs2))]>;
def FXTOQ : F3_3u<2, 0b110100, 0b010001100,
(outs QFPRegs:$rd), (ins DFPRegs:$rs2),
"fxtoq $rs2, $rd",
[(set QFPRegs:$rd, (SPxtof DFPRegs:$rs2))]>,
Requires<[HasHardQuad]>;
def FSTOX : F3_3u<2, 0b110100, 0b010000001,
(outs DFPRegs:$rd), (ins FPRegs:$rs2),
"fstox $rs2, $rd",
[(set DFPRegs:$rd, (SPftox FPRegs:$rs2))]>;
def FDTOX : F3_3u<2, 0b110100, 0b010000010,
(outs DFPRegs:$rd), (ins DFPRegs:$rs2),
"fdtox $rs2, $rd",
[(set DFPRegs:$rd, (SPftox DFPRegs:$rs2))]>;
def FQTOX : F3_3u<2, 0b110100, 0b010000011,
(outs DFPRegs:$rd), (ins QFPRegs:$rs2),
"fqtox $rs2, $rd",
[(set DFPRegs:$rd, (SPftox QFPRegs:$rs2))]>,
Requires<[HasHardQuad]>;
} // Predicates = [Is64Bit]
def : Pat<(SPselectxcc i64:$t, i64:$f, imm:$cond),
(MOVXCCrr $t, $f, imm:$cond)>;
def : Pat<(SPselectxcc (i64 simm11:$t), i64:$f, imm:$cond),
(MOVXCCri (as_i32imm $t), $f, imm:$cond)>;
def : Pat<(SPselecticc i64:$t, i64:$f, imm:$cond),
(MOVICCrr $t, $f, imm:$cond)>;
def : Pat<(SPselecticc (i64 simm11:$t), i64:$f, imm:$cond),
(MOVICCri (as_i32imm $t), $f, imm:$cond)>;
def : Pat<(SPselectfcc i64:$t, i64:$f, imm:$cond),
(MOVFCCrr $t, $f, imm:$cond)>;
def : Pat<(SPselectfcc (i64 simm11:$t), i64:$f, imm:$cond),
(MOVFCCri (as_i32imm $t), $f, imm:$cond)>;
} // Predicates = [Is64Bit]
// 64 bit SETHI
let Predicates = [Is64Bit], isCodeGenOnly = 1 in {
def SETHIXi : F2_1<0b100,
(outs IntRegs:$rd), (ins i64imm:$imm22),
"sethi $imm22, $rd",
[(set i64:$rd, SETHIimm:$imm22)]>;
}
// ATOMICS.
let Predicates = [Is64Bit], Constraints = "$swap = $rd" in {
def CASXrr: F3_1<3, 0b111110,
(outs I64Regs:$rd), (ins I64Regs:$rs1, I64Regs:$rs2,
I64Regs:$swap),
"casx [$rs1], $rs2, $rd",
[(set i64:$rd,
(atomic_cmp_swap i64:$rs1, i64:$rs2, i64:$swap))]>;
} // Predicates = [Is64Bit], Constraints = ...
let Predicates = [Is64Bit] in {
def : Pat<(atomic_fence imm, imm), (MEMBARi 0xf)>;
// atomic_load_64 addr -> load addr
def : Pat<(i64 (atomic_load ADDRrr:$src)), (LDXrr ADDRrr:$src)>;
def : Pat<(i64 (atomic_load ADDRri:$src)), (LDXri ADDRri:$src)>;
// atomic_store_64 val, addr -> store val, addr
def : Pat<(atomic_store ADDRrr:$dst, i64:$val), (STXrr ADDRrr:$dst, $val)>;
def : Pat<(atomic_store ADDRri:$dst, i64:$val), (STXri ADDRri:$dst, $val)>;
} // Predicates = [Is64Bit]
// Global addresses, constant pool entries
let Predicates = [Is64Bit] in {
def : Pat<(SPhi tglobaladdr:$in), (SETHIi tglobaladdr:$in)>;
def : Pat<(SPlo tglobaladdr:$in), (ORXri (i64 G0), tglobaladdr:$in)>;
def : Pat<(SPhi tconstpool:$in), (SETHIi tconstpool:$in)>;
def : Pat<(SPlo tconstpool:$in), (ORXri (i64 G0), tconstpool:$in)>;
// GlobalTLS addresses
def : Pat<(SPhi tglobaltlsaddr:$in), (SETHIi tglobaltlsaddr:$in)>;
def : Pat<(SPlo tglobaltlsaddr:$in), (ORXri (i64 G0), tglobaltlsaddr:$in)>;
def : Pat<(add (SPhi tglobaltlsaddr:$in1), (SPlo tglobaltlsaddr:$in2)),
(ADDXri (SETHIXi tglobaltlsaddr:$in1), (tglobaltlsaddr:$in2))>;
def : Pat<(xor (SPhi tglobaltlsaddr:$in1), (SPlo tglobaltlsaddr:$in2)),
(XORXri (SETHIXi tglobaltlsaddr:$in1), (tglobaltlsaddr:$in2))>;
// Blockaddress
def : Pat<(SPhi tblockaddress:$in), (SETHIi tblockaddress:$in)>;
def : Pat<(SPlo tblockaddress:$in), (ORXri (i64 G0), tblockaddress:$in)>;
// Add reg, lo. This is used when taking the addr of a global/constpool entry.
def : Pat<(add iPTR:$r, (SPlo tglobaladdr:$in)), (ADDXri $r, tglobaladdr:$in)>;
def : Pat<(add iPTR:$r, (SPlo tconstpool:$in)), (ADDXri $r, tconstpool:$in)>;
def : Pat<(add iPTR:$r, (SPlo tblockaddress:$in)),
(ADDXri $r, tblockaddress:$in)>;
}