Source file src/cmd/compile/internal/ssa/regalloc.go

     1  // Copyright 2015 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // Register allocation.
     6  //
     7  // We use a version of a linear scan register allocator. We treat the
     8  // whole function as a single long basic block and run through
     9  // it using a greedy register allocator. Then all merge edges
    10  // (those targeting a block with len(Preds)>1) are processed to
    11  // shuffle data into the place that the target of the edge expects.
    12  //
    13  // The greedy allocator moves values into registers just before they
    14  // are used, spills registers only when necessary, and spills the
    15  // value whose next use is farthest in the future.
    16  //
    17  // The register allocator requires that a block is not scheduled until
    18  // at least one of its predecessors have been scheduled. The most recent
    19  // such predecessor provides the starting register state for a block.
    20  //
    21  // It also requires that there are no critical edges (critical =
    22  // comes from a block with >1 successor and goes to a block with >1
    23  // predecessor).  This makes it easy to add fixup code on merge edges -
    24  // the source of a merge edge has only one successor, so we can add
    25  // fixup code to the end of that block.
    26  
    27  // Spilling
    28  //
    29  // During the normal course of the allocator, we might throw a still-live
    30  // value out of all registers. When that value is subsequently used, we must
    31  // load it from a slot on the stack. We must also issue an instruction to
    32  // initialize that stack location with a copy of v.
    33  //
    34  // pre-regalloc:
    35  //   (1) v = Op ...
    36  //   (2) x = Op ...
    37  //   (3) ... = Op v ...
    38  //
    39  // post-regalloc:
    40  //   (1) v = Op ...    : AX // computes v, store result in AX
    41  //       s = StoreReg v     // spill v to a stack slot
    42  //   (2) x = Op ...    : AX // some other op uses AX
    43  //       c = LoadReg s : CX // restore v from stack slot
    44  //   (3) ... = Op c ...     // use the restored value
    45  //
    46  // Allocation occurs normally until we reach (3) and we realize we have
    47  // a use of v and it isn't in any register. At that point, we allocate
    48  // a spill (a StoreReg) for v. We can't determine the correct place for
    49  // the spill at this point, so we allocate the spill as blockless initially.
    50  // The restore is then generated to load v back into a register so it can
    51  // be used. Subsequent uses of v will use the restored value c instead.
    52  //
    53  // What remains is the question of where to schedule the spill.
    54  // During allocation, we keep track of the dominator of all restores of v.
    55  // The spill of v must dominate that block. The spill must also be issued at
    56  // a point where v is still in a register.
    57  //
    58  // To find the right place, start at b, the block which dominates all restores.
    59  //  - If b is v.Block, then issue the spill right after v.
    60  //    It is known to be in a register at that point, and dominates any restores.
    61  //  - Otherwise, if v is in a register at the start of b,
    62  //    put the spill of v at the start of b.
    63  //  - Otherwise, set b = immediate dominator of b, and repeat.
    64  //
    65  // Phi values are special, as always. We define two kinds of phis, those
    66  // where the merge happens in a register (a "register" phi) and those where
    67  // the merge happens in a stack location (a "stack" phi).
    68  //
    69  // A register phi must have the phi and all of its inputs allocated to the
    70  // same register. Register phis are spilled similarly to regular ops.
    71  //
    72  // A stack phi must have the phi and all of its inputs allocated to the same
    73  // stack location. Stack phis start out life already spilled - each phi
    74  // input must be a store (using StoreReg) at the end of the corresponding
    75  // predecessor block.
    76  //     b1: y = ... : AX        b2: z = ... : BX
    77  //         y2 = StoreReg y         z2 = StoreReg z
    78  //         goto b3                 goto b3
    79  //     b3: x = phi(y2, z2)
    80  // The stack allocator knows that StoreReg args of stack-allocated phis
    81  // must be allocated to the same stack slot as the phi that uses them.
    82  // x is now a spilled value and a restore must appear before its first use.
    83  
    84  // TODO
    85  
    86  // Use an affinity graph to mark two values which should use the
    87  // same register. This affinity graph will be used to prefer certain
    88  // registers for allocation. This affinity helps eliminate moves that
    89  // are required for phi implementations and helps generate allocations
    90  // for 2-register architectures.
    91  
    92  // Note: regalloc generates a not-quite-SSA output. If we have:
    93  //
    94  //             b1: x = ... : AX
    95  //                 x2 = StoreReg x
    96  //                 ... AX gets reused for something else ...
    97  //                 if ... goto b3 else b4
    98  //
    99  //   b3: x3 = LoadReg x2 : BX       b4: x4 = LoadReg x2 : CX
   100  //       ... use x3 ...                 ... use x4 ...
   101  //
   102  //             b2: ... use x3 ...
   103  //
   104  // If b3 is the primary predecessor of b2, then we use x3 in b2 and
   105  // add a x4:CX->BX copy at the end of b4.
   106  // But the definition of x3 doesn't dominate b2.  We should really
   107  // insert an extra phi at the start of b2 (x5=phi(x3,x4):BX) to keep
   108  // SSA form. For now, we ignore this problem as remaining in strict
   109  // SSA form isn't needed after regalloc. We'll just leave the use
   110  // of x3 not dominated by the definition of x3, and the CX->BX copy
   111  // will have no use (so don't run deadcode after regalloc!).
   112  // TODO: maybe we should introduce these extra phis?
   113  
   114  package ssa
   115  
   116  import (
   117  	"cmd/compile/internal/base"
   118  	"cmd/compile/internal/ir"
   119  	"cmd/compile/internal/types"
   120  	"cmd/internal/src"
   121  	"cmd/internal/sys"
   122  	"fmt"
   123  	"internal/buildcfg"
   124  	"math"
   125  	"math/bits"
   126  	"unsafe"
   127  )
   128  
   129  const (
   130  	moveSpills = iota
   131  	logSpills
   132  	regDebug
   133  	stackDebug
   134  )
   135  
   136  // distance is a measure of how far into the future values are used.
   137  // distance is measured in units of instructions.
   138  const (
   139  	likelyDistance   = 1
   140  	normalDistance   = 10
   141  	unlikelyDistance = 100
   142  )
   143  
   144  // regalloc performs register allocation on f. It sets f.RegAlloc
   145  // to the resulting allocation.
   146  func regalloc(f *Func) {
   147  	var s regAllocState
   148  	s.init(f)
   149  	s.regalloc(f)
   150  	s.close()
   151  }
   152  
   153  type register uint8
   154  
   155  const noRegister register = 255
   156  
   157  // For bulk initializing
   158  var noRegisters [32]register = [32]register{
   159  	noRegister, noRegister, noRegister, noRegister, noRegister, noRegister, noRegister, noRegister,
   160  	noRegister, noRegister, noRegister, noRegister, noRegister, noRegister, noRegister, noRegister,
   161  	noRegister, noRegister, noRegister, noRegister, noRegister, noRegister, noRegister, noRegister,
   162  	noRegister, noRegister, noRegister, noRegister, noRegister, noRegister, noRegister, noRegister,
   163  }
   164  
   165  // A regMask encodes a set of machine registers.
   166  // TODO: regMask -> regSet?
   167  type regMask uint64
   168  
   169  func (m regMask) String() string {
   170  	s := ""
   171  	for r := register(0); m != 0; r++ {
   172  		if m>>r&1 == 0 {
   173  			continue
   174  		}
   175  		m &^= regMask(1) << r
   176  		if s != "" {
   177  			s += " "
   178  		}
   179  		s += fmt.Sprintf("r%d", r)
   180  	}
   181  	return s
   182  }
   183  
   184  func (m regMask) contains(r register) bool {
   185  	return m>>r&1 != 0
   186  }
   187  
   188  func (s *regAllocState) RegMaskString(m regMask) string {
   189  	str := ""
   190  	for r := register(0); m != 0; r++ {
   191  		if m>>r&1 == 0 {
   192  			continue
   193  		}
   194  		m &^= regMask(1) << r
   195  		if str != "" {
   196  			str += " "
   197  		}
   198  		str += s.registers[r].String()
   199  	}
   200  	return str
   201  }
   202  
   203  // countRegs returns the number of set bits in the register mask.
   204  func countRegs(r regMask) int {
   205  	return bits.OnesCount64(uint64(r))
   206  }
   207  
   208  // pickReg picks an arbitrary register from the register mask.
   209  func pickReg(r regMask) register {
   210  	if r == 0 {
   211  		panic("can't pick a register from an empty set")
   212  	}
   213  	// pick the lowest one
   214  	return register(bits.TrailingZeros64(uint64(r)))
   215  }
   216  
   217  type use struct {
   218  	// distance from start of the block to a use of a value
   219  	//   dist == 0                 used by first instruction in block
   220  	//   dist == len(b.Values)-1   used by last instruction in block
   221  	//   dist == len(b.Values)     used by block's control value
   222  	//   dist  > len(b.Values)     used by a subsequent block
   223  	dist int32
   224  	pos  src.XPos // source position of the use
   225  	next *use     // linked list of uses of a value in nondecreasing dist order
   226  }
   227  
   228  // A valState records the register allocation state for a (pre-regalloc) value.
   229  type valState struct {
   230  	regs              regMask // the set of registers holding a Value (usually just one)
   231  	uses              *use    // list of uses in this block
   232  	spill             *Value  // spilled copy of the Value (if any)
   233  	restoreMin        int32   // minimum of all restores' blocks' sdom.entry
   234  	restoreMax        int32   // maximum of all restores' blocks' sdom.exit
   235  	needReg           bool    // cached value of !v.Type.IsMemory() && !v.Type.IsVoid() && !.v.Type.IsFlags()
   236  	rematerializeable bool    // cached value of v.rematerializeable()
   237  }
   238  
   239  type regState struct {
   240  	v *Value // Original (preregalloc) Value stored in this register.
   241  	c *Value // A Value equal to v which is currently in a register.  Might be v or a copy of it.
   242  	// If a register is unused, v==c==nil
   243  }
   244  
   245  type regAllocState struct {
   246  	f *Func
   247  
   248  	sdom        SparseTree
   249  	registers   []Register
   250  	numRegs     register
   251  	SPReg       register
   252  	SBReg       register
   253  	GReg        register
   254  	ZeroIntReg  register
   255  	allocatable regMask
   256  
   257  	// live values at the end of each block.  live[b.ID] is a list of value IDs
   258  	// which are live at the end of b, together with a count of how many instructions
   259  	// forward to the next use.
   260  	live [][]liveInfo
   261  	// desired register assignments at the end of each block.
   262  	// Note that this is a static map computed before allocation occurs. Dynamic
   263  	// register desires (from partially completed allocations) will trump
   264  	// this information.
   265  	desired []desiredState
   266  
   267  	// current state of each (preregalloc) Value
   268  	values []valState
   269  
   270  	// ID of SP, SB values
   271  	sp, sb ID
   272  
   273  	// For each Value, map from its value ID back to the
   274  	// preregalloc Value it was derived from.
   275  	orig []*Value
   276  
   277  	// current state of each register.
   278  	// Includes only registers in allocatable.
   279  	regs []regState
   280  
   281  	// registers that contain values which can't be kicked out
   282  	nospill regMask
   283  
   284  	// mask of registers currently in use
   285  	used regMask
   286  
   287  	// mask of registers used since the start of the current block
   288  	usedSinceBlockStart regMask
   289  
   290  	// mask of registers used in the current instruction
   291  	tmpused regMask
   292  
   293  	// current block we're working on
   294  	curBlock *Block
   295  
   296  	// cache of use records
   297  	freeUseRecords *use
   298  
   299  	// endRegs[blockid] is the register state at the end of each block.
   300  	// encoded as a set of endReg records.
   301  	endRegs [][]endReg
   302  
   303  	// startRegs[blockid] is the register state at the start of merge blocks.
   304  	// saved state does not include the state of phi ops in the block.
   305  	startRegs [][]startReg
   306  
   307  	// startRegsMask is a mask of the registers in startRegs[curBlock.ID].
   308  	// Registers dropped from startRegsMask are later synchronoized back to
   309  	// startRegs by dropping from there as well.
   310  	startRegsMask regMask
   311  
   312  	// spillLive[blockid] is the set of live spills at the end of each block
   313  	spillLive [][]ID
   314  
   315  	// a set of copies we generated to move things around, and
   316  	// whether it is used in shuffle. Unused copies will be deleted.
   317  	copies map[*Value]bool
   318  
   319  	loopnest *loopnest
   320  
   321  	// choose a good order in which to visit blocks for allocation purposes.
   322  	visitOrder []*Block
   323  
   324  	// blockOrder[b.ID] corresponds to the index of block b in visitOrder.
   325  	blockOrder []int32
   326  
   327  	// whether to insert instructions that clobber dead registers at call sites
   328  	doClobber bool
   329  
   330  	// For each instruction index in a basic block, the index of the next call
   331  	// at or after that instruction index.
   332  	// If there is no next call, returns maxInt32.
   333  	// nextCall for a call instruction points to itself.
   334  	// (Indexes and results are pre-regalloc.)
   335  	nextCall []int32
   336  
   337  	// Index of the instruction we're currently working on.
   338  	// Index is expressed in terms of the pre-regalloc b.Values list.
   339  	curIdx int
   340  }
   341  
   342  type endReg struct {
   343  	r register
   344  	v *Value // pre-regalloc value held in this register (TODO: can we use ID here?)
   345  	c *Value // cached version of the value
   346  }
   347  
   348  type startReg struct {
   349  	r   register
   350  	v   *Value   // pre-regalloc value needed in this register
   351  	c   *Value   // cached version of the value
   352  	pos src.XPos // source position of use of this register
   353  }
   354  
   355  // freeReg frees up register r. Any current user of r is kicked out.
   356  func (s *regAllocState) freeReg(r register) {
   357  	if !s.allocatable.contains(r) && !s.isGReg(r) {
   358  		return
   359  	}
   360  	v := s.regs[r].v
   361  	if v == nil {
   362  		s.f.Fatalf("tried to free an already free register %d\n", r)
   363  	}
   364  
   365  	// Mark r as unused.
   366  	if s.f.pass.debug > regDebug {
   367  		fmt.Printf("freeReg %s (dump %s/%s)\n", &s.registers[r], v, s.regs[r].c)
   368  	}
   369  	s.regs[r] = regState{}
   370  	s.values[v.ID].regs &^= regMask(1) << r
   371  	s.used &^= regMask(1) << r
   372  }
   373  
   374  // freeRegs frees up all registers listed in m.
   375  func (s *regAllocState) freeRegs(m regMask) {
   376  	for m&s.used != 0 {
   377  		s.freeReg(pickReg(m & s.used))
   378  	}
   379  }
   380  
   381  // clobberRegs inserts instructions that clobber registers listed in m.
   382  func (s *regAllocState) clobberRegs(m regMask) {
   383  	m &= s.allocatable & s.f.Config.gpRegMask // only integer register can contain pointers, only clobber them
   384  	for m != 0 {
   385  		r := pickReg(m)
   386  		m &^= 1 << r
   387  		x := s.curBlock.NewValue0(src.NoXPos, OpClobberReg, types.TypeVoid)
   388  		s.f.setHome(x, &s.registers[r])
   389  	}
   390  }
   391  
   392  // setOrig records that c's original value is the same as
   393  // v's original value.
   394  func (s *regAllocState) setOrig(c *Value, v *Value) {
   395  	if int(c.ID) >= cap(s.orig) {
   396  		x := s.f.Cache.allocValueSlice(int(c.ID) + 1)
   397  		copy(x, s.orig)
   398  		s.f.Cache.freeValueSlice(s.orig)
   399  		s.orig = x
   400  	}
   401  	for int(c.ID) >= len(s.orig) {
   402  		s.orig = append(s.orig, nil)
   403  	}
   404  	if s.orig[c.ID] != nil {
   405  		s.f.Fatalf("orig value set twice %s %s", c, v)
   406  	}
   407  	s.orig[c.ID] = s.orig[v.ID]
   408  }
   409  
   410  // assignReg assigns register r to hold c, a copy of v.
   411  // r must be unused.
   412  func (s *regAllocState) assignReg(r register, v *Value, c *Value) {
   413  	if s.f.pass.debug > regDebug {
   414  		fmt.Printf("assignReg %s %s/%s\n", &s.registers[r], v, c)
   415  	}
   416  	// Allocate v to r.
   417  	s.values[v.ID].regs |= regMask(1) << r
   418  	s.f.setHome(c, &s.registers[r])
   419  
   420  	// Allocate r to v.
   421  	if !s.allocatable.contains(r) && !s.isGReg(r) {
   422  		return
   423  	}
   424  	if s.regs[r].v != nil {
   425  		s.f.Fatalf("tried to assign register %d to %s/%s but it is already used by %s", r, v, c, s.regs[r].v)
   426  	}
   427  	s.regs[r] = regState{v, c}
   428  	s.used |= regMask(1) << r
   429  }
   430  
   431  // allocReg chooses a register from the set of registers in mask.
   432  // If there is no unused register, a Value will be kicked out of
   433  // a register to make room.
   434  func (s *regAllocState) allocReg(mask regMask, v *Value) register {
   435  	if v.OnWasmStack {
   436  		return noRegister
   437  	}
   438  
   439  	mask &= s.allocatable
   440  	mask &^= s.nospill
   441  	if mask == 0 {
   442  		s.f.Fatalf("no register available for %s", v.LongString())
   443  	}
   444  
   445  	// Pick an unused register if one is available.
   446  	if mask&^s.used != 0 {
   447  		r := pickReg(mask &^ s.used)
   448  		s.usedSinceBlockStart |= regMask(1) << r
   449  		return r
   450  	}
   451  
   452  	// Pick a value to spill. Spill the value with the
   453  	// farthest-in-the-future use.
   454  	// TODO: Prefer registers with already spilled Values?
   455  	// TODO: Modify preference using affinity graph.
   456  	// TODO: if a single value is in multiple registers, spill one of them
   457  	// before spilling a value in just a single register.
   458  
   459  	// Find a register to spill. We spill the register containing the value
   460  	// whose next use is as far in the future as possible.
   461  	// https://en.wikipedia.org/wiki/Page_replacement_algorithm#The_theoretically_optimal_page_replacement_algorithm
   462  	var r register
   463  	maxuse := int32(-1)
   464  	for t := register(0); t < s.numRegs; t++ {
   465  		if mask>>t&1 == 0 {
   466  			continue
   467  		}
   468  		v := s.regs[t].v
   469  		if n := s.values[v.ID].uses.dist; n > maxuse {
   470  			// v's next use is farther in the future than any value
   471  			// we've seen so far. A new best spill candidate.
   472  			r = t
   473  			maxuse = n
   474  		}
   475  	}
   476  	if maxuse == -1 {
   477  		s.f.Fatalf("couldn't find register to spill")
   478  	}
   479  
   480  	if s.f.Config.ctxt.Arch.Arch == sys.ArchWasm {
   481  		// TODO(neelance): In theory this should never happen, because all wasm registers are equal.
   482  		// So if there is still a free register, the allocation should have picked that one in the first place instead of
   483  		// trying to kick some other value out. In practice, this case does happen and it breaks the stack optimization.
   484  		s.freeReg(r)
   485  		return r
   486  	}
   487  
   488  	// Try to move it around before kicking out, if there is a free register.
   489  	// We generate a Copy and record it. It will be deleted if never used.
   490  	v2 := s.regs[r].v
   491  	m := s.compatRegs(v2.Type) &^ s.used &^ s.tmpused &^ (regMask(1) << r)
   492  	if m != 0 && !s.values[v2.ID].rematerializeable && countRegs(s.values[v2.ID].regs) == 1 {
   493  		s.usedSinceBlockStart |= regMask(1) << r
   494  		r2 := pickReg(m)
   495  		c := s.curBlock.NewValue1(v2.Pos, OpCopy, v2.Type, s.regs[r].c)
   496  		s.copies[c] = false
   497  		if s.f.pass.debug > regDebug {
   498  			fmt.Printf("copy %s to %s : %s\n", v2, c, &s.registers[r2])
   499  		}
   500  		s.setOrig(c, v2)
   501  		s.assignReg(r2, v2, c)
   502  	}
   503  
   504  	// If the evicted register isn't used between the start of the block
   505  	// and now then there is no reason to even request it on entry. We can
   506  	// drop from startRegs in that case.
   507  	if s.usedSinceBlockStart&(regMask(1)<<r) == 0 {
   508  		if s.startRegsMask&(regMask(1)<<r) == 1 {
   509  			if s.f.pass.debug > regDebug {
   510  				fmt.Printf("dropped from startRegs: %s\n", &s.registers[r])
   511  			}
   512  			s.startRegsMask &^= regMask(1) << r
   513  		}
   514  	}
   515  
   516  	s.freeReg(r)
   517  	s.usedSinceBlockStart |= regMask(1) << r
   518  	return r
   519  }
   520  
   521  // makeSpill returns a Value which represents the spilled value of v.
   522  // b is the block in which the spill is used.
   523  func (s *regAllocState) makeSpill(v *Value, b *Block) *Value {
   524  	vi := &s.values[v.ID]
   525  	if vi.spill != nil {
   526  		// Final block not known - keep track of subtree where restores reside.
   527  		vi.restoreMin = min(vi.restoreMin, s.sdom[b.ID].entry)
   528  		vi.restoreMax = max(vi.restoreMax, s.sdom[b.ID].exit)
   529  		return vi.spill
   530  	}
   531  	// Make a spill for v. We don't know where we want
   532  	// to put it yet, so we leave it blockless for now.
   533  	spill := s.f.newValueNoBlock(OpStoreReg, v.Type, v.Pos)
   534  	// We also don't know what the spill's arg will be.
   535  	// Leave it argless for now.
   536  	s.setOrig(spill, v)
   537  	vi.spill = spill
   538  	vi.restoreMin = s.sdom[b.ID].entry
   539  	vi.restoreMax = s.sdom[b.ID].exit
   540  	return spill
   541  }
   542  
   543  // allocValToReg allocates v to a register selected from regMask and
   544  // returns the register copy of v. Any previous user is kicked out and spilled
   545  // (if necessary). Load code is added at the current pc. If nospill is set the
   546  // allocated register is marked nospill so the assignment cannot be
   547  // undone until the caller allows it by clearing nospill. Returns a
   548  // *Value which is either v or a copy of v allocated to the chosen register.
   549  func (s *regAllocState) allocValToReg(v *Value, mask regMask, nospill bool, pos src.XPos) *Value {
   550  	if s.f.Config.ctxt.Arch.Arch == sys.ArchWasm && v.rematerializeable() {
   551  		c := v.copyIntoWithXPos(s.curBlock, pos)
   552  		c.OnWasmStack = true
   553  		s.setOrig(c, v)
   554  		return c
   555  	}
   556  	if v.OnWasmStack {
   557  		return v
   558  	}
   559  
   560  	vi := &s.values[v.ID]
   561  	pos = pos.WithNotStmt()
   562  	// Check if v is already in a requested register.
   563  	if mask&vi.regs != 0 {
   564  		r := pickReg(mask & vi.regs)
   565  		if !s.allocatable.contains(r) {
   566  			return v // v is in a fixed register
   567  		}
   568  		if s.regs[r].v != v || s.regs[r].c == nil {
   569  			panic("bad register state")
   570  		}
   571  		if nospill {
   572  			s.nospill |= regMask(1) << r
   573  		}
   574  		s.usedSinceBlockStart |= regMask(1) << r
   575  		return s.regs[r].c
   576  	}
   577  
   578  	var r register
   579  	// If nospill is set, the value is used immediately, so it can live on the WebAssembly stack.
   580  	onWasmStack := nospill && s.f.Config.ctxt.Arch.Arch == sys.ArchWasm
   581  	if !onWasmStack {
   582  		// Allocate a register.
   583  		r = s.allocReg(mask, v)
   584  	}
   585  
   586  	// Allocate v to the new register.
   587  	var c *Value
   588  	if vi.regs != 0 {
   589  		// Copy from a register that v is already in.
   590  		r2 := pickReg(vi.regs)
   591  		var current *Value
   592  		if !s.allocatable.contains(r2) {
   593  			current = v // v is in a fixed register
   594  		} else {
   595  			if s.regs[r2].v != v {
   596  				panic("bad register state")
   597  			}
   598  			current = s.regs[r2].c
   599  		}
   600  		s.usedSinceBlockStart |= regMask(1) << r2
   601  		c = s.curBlock.NewValue1(pos, OpCopy, v.Type, current)
   602  	} else if v.rematerializeable() {
   603  		// Rematerialize instead of loading from the spill location.
   604  		c = v.copyIntoWithXPos(s.curBlock, pos)
   605  	} else {
   606  		// Load v from its spill location.
   607  		spill := s.makeSpill(v, s.curBlock)
   608  		if s.f.pass.debug > logSpills {
   609  			s.f.Warnl(vi.spill.Pos, "load spill for %v from %v", v, spill)
   610  		}
   611  		c = s.curBlock.NewValue1(pos, OpLoadReg, v.Type, spill)
   612  	}
   613  
   614  	s.setOrig(c, v)
   615  
   616  	if onWasmStack {
   617  		c.OnWasmStack = true
   618  		return c
   619  	}
   620  
   621  	s.assignReg(r, v, c)
   622  	if c.Op == OpLoadReg && s.isGReg(r) {
   623  		s.f.Fatalf("allocValToReg.OpLoadReg targeting g: " + c.LongString())
   624  	}
   625  	if nospill {
   626  		s.nospill |= regMask(1) << r
   627  	}
   628  	return c
   629  }
   630  
   631  // isLeaf reports whether f performs any calls.
   632  func isLeaf(f *Func) bool {
   633  	for _, b := range f.Blocks {
   634  		for _, v := range b.Values {
   635  			if v.Op.IsCall() && !v.Op.IsTailCall() {
   636  				// tail call is not counted as it does not save the return PC or need a frame
   637  				return false
   638  			}
   639  		}
   640  	}
   641  	return true
   642  }
   643  
   644  // needRegister reports whether v needs a register.
   645  func (v *Value) needRegister() bool {
   646  	return !v.Type.IsMemory() && !v.Type.IsVoid() && !v.Type.IsFlags() && !v.Type.IsTuple()
   647  }
   648  
   649  func (s *regAllocState) init(f *Func) {
   650  	s.f = f
   651  	s.f.RegAlloc = s.f.Cache.locs[:0]
   652  	s.registers = f.Config.registers
   653  	if nr := len(s.registers); nr == 0 || nr > int(noRegister) || nr > int(unsafe.Sizeof(regMask(0))*8) {
   654  		s.f.Fatalf("bad number of registers: %d", nr)
   655  	} else {
   656  		s.numRegs = register(nr)
   657  	}
   658  	// Locate SP, SB, and g registers.
   659  	s.SPReg = noRegister
   660  	s.SBReg = noRegister
   661  	s.GReg = noRegister
   662  	s.ZeroIntReg = noRegister
   663  	for r := register(0); r < s.numRegs; r++ {
   664  		switch s.registers[r].String() {
   665  		case "SP":
   666  			s.SPReg = r
   667  		case "SB":
   668  			s.SBReg = r
   669  		case "g":
   670  			s.GReg = r
   671  		case "ZERO": // TODO: arch-specific?
   672  			s.ZeroIntReg = r
   673  		}
   674  	}
   675  	// Make sure we found all required registers.
   676  	switch noRegister {
   677  	case s.SPReg:
   678  		s.f.Fatalf("no SP register found")
   679  	case s.SBReg:
   680  		s.f.Fatalf("no SB register found")
   681  	case s.GReg:
   682  		if f.Config.hasGReg {
   683  			s.f.Fatalf("no g register found")
   684  		}
   685  	}
   686  
   687  	// Figure out which registers we're allowed to use.
   688  	s.allocatable = s.f.Config.gpRegMask | s.f.Config.fpRegMask | s.f.Config.specialRegMask
   689  	s.allocatable &^= 1 << s.SPReg
   690  	s.allocatable &^= 1 << s.SBReg
   691  	if s.f.Config.hasGReg {
   692  		s.allocatable &^= 1 << s.GReg
   693  	}
   694  	if s.ZeroIntReg != noRegister {
   695  		s.allocatable &^= 1 << s.ZeroIntReg
   696  	}
   697  	if buildcfg.FramePointerEnabled && s.f.Config.FPReg >= 0 {
   698  		s.allocatable &^= 1 << uint(s.f.Config.FPReg)
   699  	}
   700  	if s.f.Config.LinkReg != -1 {
   701  		if isLeaf(f) {
   702  			// Leaf functions don't save/restore the link register.
   703  			s.allocatable &^= 1 << uint(s.f.Config.LinkReg)
   704  		}
   705  	}
   706  	if s.f.Config.ctxt.Flag_dynlink {
   707  		switch s.f.Config.arch {
   708  		case "386":
   709  			// nothing to do.
   710  			// Note that for Flag_shared (position independent code)
   711  			// we do need to be careful, but that carefulness is hidden
   712  			// in the rewrite rules so we always have a free register
   713  			// available for global load/stores. See _gen/386.rules (search for Flag_shared).
   714  		case "amd64":
   715  			s.allocatable &^= 1 << 15 // R15
   716  		case "arm":
   717  			s.allocatable &^= 1 << 9 // R9
   718  		case "arm64":
   719  			// nothing to do
   720  		case "loong64": // R2 (aka TP) already reserved.
   721  			// nothing to do
   722  		case "ppc64le": // R2 already reserved.
   723  			// nothing to do
   724  		case "riscv64": // X3 (aka GP) and X4 (aka TP) already reserved.
   725  			// nothing to do
   726  		case "s390x":
   727  			s.allocatable &^= 1 << 11 // R11
   728  		default:
   729  			s.f.fe.Fatalf(src.NoXPos, "arch %s not implemented", s.f.Config.arch)
   730  		}
   731  	}
   732  
   733  	// Linear scan register allocation can be influenced by the order in which blocks appear.
   734  	// Decouple the register allocation order from the generated block order.
   735  	// This also creates an opportunity for experiments to find a better order.
   736  	s.visitOrder = layoutRegallocOrder(f)
   737  
   738  	// Compute block order. This array allows us to distinguish forward edges
   739  	// from backward edges and compute how far they go.
   740  	s.blockOrder = make([]int32, f.NumBlocks())
   741  	for i, b := range s.visitOrder {
   742  		s.blockOrder[b.ID] = int32(i)
   743  	}
   744  
   745  	s.regs = make([]regState, s.numRegs)
   746  	nv := f.NumValues()
   747  	if cap(s.f.Cache.regallocValues) >= nv {
   748  		s.f.Cache.regallocValues = s.f.Cache.regallocValues[:nv]
   749  	} else {
   750  		s.f.Cache.regallocValues = make([]valState, nv)
   751  	}
   752  	s.values = s.f.Cache.regallocValues
   753  	s.orig = s.f.Cache.allocValueSlice(nv)
   754  	s.copies = make(map[*Value]bool)
   755  	for _, b := range s.visitOrder {
   756  		for _, v := range b.Values {
   757  			if v.needRegister() {
   758  				s.values[v.ID].needReg = true
   759  				s.values[v.ID].rematerializeable = v.rematerializeable()
   760  				s.orig[v.ID] = v
   761  			}
   762  			// Note: needReg is false for values returning Tuple types.
   763  			// Instead, we mark the corresponding Selects as needReg.
   764  		}
   765  	}
   766  	s.computeLive()
   767  
   768  	s.endRegs = make([][]endReg, f.NumBlocks())
   769  	s.startRegs = make([][]startReg, f.NumBlocks())
   770  	s.spillLive = make([][]ID, f.NumBlocks())
   771  	s.sdom = f.Sdom()
   772  
   773  	// wasm: Mark instructions that can be optimized to have their values only on the WebAssembly stack.
   774  	if f.Config.ctxt.Arch.Arch == sys.ArchWasm {
   775  		canLiveOnStack := f.newSparseSet(f.NumValues())
   776  		defer f.retSparseSet(canLiveOnStack)
   777  		for _, b := range f.Blocks {
   778  			// New block. Clear candidate set.
   779  			canLiveOnStack.clear()
   780  			for _, c := range b.ControlValues() {
   781  				if c.Uses == 1 && !opcodeTable[c.Op].generic {
   782  					canLiveOnStack.add(c.ID)
   783  				}
   784  			}
   785  			// Walking backwards.
   786  			for i := len(b.Values) - 1; i >= 0; i-- {
   787  				v := b.Values[i]
   788  				if canLiveOnStack.contains(v.ID) {
   789  					v.OnWasmStack = true
   790  				} else {
   791  					// Value can not live on stack. Values are not allowed to be reordered, so clear candidate set.
   792  					canLiveOnStack.clear()
   793  				}
   794  				for _, arg := range v.Args {
   795  					// Value can live on the stack if:
   796  					// - it is only used once
   797  					// - it is used in the same basic block
   798  					// - it is not a "mem" value
   799  					// - it is a WebAssembly op
   800  					if arg.Uses == 1 && arg.Block == v.Block && !arg.Type.IsMemory() && !opcodeTable[arg.Op].generic {
   801  						canLiveOnStack.add(arg.ID)
   802  					}
   803  				}
   804  			}
   805  		}
   806  	}
   807  
   808  	// The clobberdeadreg experiment inserts code to clobber dead registers
   809  	// at call sites.
   810  	// Ignore huge functions to avoid doing too much work.
   811  	if base.Flag.ClobberDeadReg && len(s.f.Blocks) <= 10000 {
   812  		// TODO: honor GOCLOBBERDEADHASH, or maybe GOSSAHASH.
   813  		s.doClobber = true
   814  	}
   815  }
   816  
   817  func (s *regAllocState) close() {
   818  	s.f.Cache.freeValueSlice(s.orig)
   819  }
   820  
   821  // Adds a use record for id at distance dist from the start of the block.
   822  // All calls to addUse must happen with nonincreasing dist.
   823  func (s *regAllocState) addUse(id ID, dist int32, pos src.XPos) {
   824  	r := s.freeUseRecords
   825  	if r != nil {
   826  		s.freeUseRecords = r.next
   827  	} else {
   828  		r = &use{}
   829  	}
   830  	r.dist = dist
   831  	r.pos = pos
   832  	r.next = s.values[id].uses
   833  	s.values[id].uses = r
   834  	if r.next != nil && dist > r.next.dist {
   835  		s.f.Fatalf("uses added in wrong order")
   836  	}
   837  }
   838  
   839  // advanceUses advances the uses of v's args from the state before v to the state after v.
   840  // Any values which have no more uses are deallocated from registers.
   841  func (s *regAllocState) advanceUses(v *Value) {
   842  	for _, a := range v.Args {
   843  		if !s.values[a.ID].needReg {
   844  			continue
   845  		}
   846  		ai := &s.values[a.ID]
   847  		r := ai.uses
   848  		ai.uses = r.next
   849  		if r.next == nil || (!opcodeTable[a.Op].fixedReg && r.next.dist > s.nextCall[s.curIdx]) {
   850  			// Value is dead (or is not used again until after a call), free all registers that hold it.
   851  			s.freeRegs(ai.regs)
   852  		}
   853  		r.next = s.freeUseRecords
   854  		s.freeUseRecords = r
   855  	}
   856  	s.dropIfUnused(v)
   857  }
   858  
   859  // Drop v from registers if it isn't used again, or its only uses are after
   860  // a call instruction.
   861  func (s *regAllocState) dropIfUnused(v *Value) {
   862  	if !s.values[v.ID].needReg {
   863  		return
   864  	}
   865  	vi := &s.values[v.ID]
   866  	r := vi.uses
   867  	if r == nil || (!opcodeTable[v.Op].fixedReg && r.dist > s.nextCall[s.curIdx]) {
   868  		s.freeRegs(vi.regs)
   869  	}
   870  }
   871  
   872  // liveAfterCurrentInstruction reports whether v is live after
   873  // the current instruction is completed.  v must be used by the
   874  // current instruction.
   875  func (s *regAllocState) liveAfterCurrentInstruction(v *Value) bool {
   876  	u := s.values[v.ID].uses
   877  	if u == nil {
   878  		panic(fmt.Errorf("u is nil, v = %s, s.values[v.ID] = %v", v.LongString(), s.values[v.ID]))
   879  	}
   880  	d := u.dist
   881  	for u != nil && u.dist == d {
   882  		u = u.next
   883  	}
   884  	return u != nil && u.dist > d
   885  }
   886  
   887  // Sets the state of the registers to that encoded in regs.
   888  func (s *regAllocState) setState(regs []endReg) {
   889  	s.freeRegs(s.used)
   890  	for _, x := range regs {
   891  		s.assignReg(x.r, x.v, x.c)
   892  	}
   893  }
   894  
   895  // compatRegs returns the set of registers which can store a type t.
   896  func (s *regAllocState) compatRegs(t *types.Type) regMask {
   897  	var m regMask
   898  	if t.IsTuple() || t.IsFlags() {
   899  		return 0
   900  	}
   901  	if t.IsFloat() || t == types.TypeInt128 {
   902  		if t.Kind() == types.TFLOAT32 && s.f.Config.fp32RegMask != 0 {
   903  			m = s.f.Config.fp32RegMask
   904  		} else if t.Kind() == types.TFLOAT64 && s.f.Config.fp64RegMask != 0 {
   905  			m = s.f.Config.fp64RegMask
   906  		} else {
   907  			m = s.f.Config.fpRegMask
   908  		}
   909  	} else {
   910  		m = s.f.Config.gpRegMask
   911  	}
   912  	return m & s.allocatable
   913  }
   914  
   915  // regspec returns the regInfo for operation op.
   916  func (s *regAllocState) regspec(v *Value) regInfo {
   917  	op := v.Op
   918  	if op == OpConvert {
   919  		// OpConvert is a generic op, so it doesn't have a
   920  		// register set in the static table. It can use any
   921  		// allocatable integer register.
   922  		m := s.allocatable & s.f.Config.gpRegMask
   923  		return regInfo{inputs: []inputInfo{{regs: m}}, outputs: []outputInfo{{regs: m}}}
   924  	}
   925  	if op == OpArgIntReg {
   926  		reg := v.Block.Func.Config.intParamRegs[v.AuxInt8()]
   927  		return regInfo{outputs: []outputInfo{{regs: 1 << uint(reg)}}}
   928  	}
   929  	if op == OpArgFloatReg {
   930  		reg := v.Block.Func.Config.floatParamRegs[v.AuxInt8()]
   931  		return regInfo{outputs: []outputInfo{{regs: 1 << uint(reg)}}}
   932  	}
   933  	if op.IsCall() {
   934  		if ac, ok := v.Aux.(*AuxCall); ok && ac.reg != nil {
   935  			return *ac.Reg(&opcodeTable[op].reg, s.f.Config)
   936  		}
   937  	}
   938  	if op == OpMakeResult && s.f.OwnAux.reg != nil {
   939  		return *s.f.OwnAux.ResultReg(s.f.Config)
   940  	}
   941  	return opcodeTable[op].reg
   942  }
   943  
   944  func (s *regAllocState) isGReg(r register) bool {
   945  	return s.f.Config.hasGReg && s.GReg == r
   946  }
   947  
   948  // Dummy value used to represent the value being held in a temporary register.
   949  var tmpVal Value
   950  
   951  func (s *regAllocState) regalloc(f *Func) {
   952  	regValLiveSet := f.newSparseSet(f.NumValues()) // set of values that may be live in register
   953  	defer f.retSparseSet(regValLiveSet)
   954  	var oldSched []*Value
   955  	var phis []*Value
   956  	var phiRegs []register
   957  	var args []*Value
   958  
   959  	// Data structure used for computing desired registers.
   960  	var desired desiredState
   961  	desiredSecondReg := map[ID][4]register{} // desired register allocation for 2nd part of a tuple
   962  
   963  	// Desired registers for inputs & outputs for each instruction in the block.
   964  	type dentry struct {
   965  		out [4]register    // desired output registers
   966  		in  [3][4]register // desired input registers (for inputs 0,1, and 2)
   967  	}
   968  	var dinfo []dentry
   969  
   970  	if f.Entry != f.Blocks[0] {
   971  		f.Fatalf("entry block must be first")
   972  	}
   973  
   974  	for _, b := range s.visitOrder {
   975  		if s.f.pass.debug > regDebug {
   976  			fmt.Printf("Begin processing block %v\n", b)
   977  		}
   978  		s.curBlock = b
   979  		s.startRegsMask = 0
   980  		s.usedSinceBlockStart = 0
   981  		clear(desiredSecondReg)
   982  
   983  		// Initialize regValLiveSet and uses fields for this block.
   984  		// Walk backwards through the block doing liveness analysis.
   985  		regValLiveSet.clear()
   986  		for _, e := range s.live[b.ID] {
   987  			s.addUse(e.ID, int32(len(b.Values))+e.dist, e.pos) // pseudo-uses from beyond end of block
   988  			regValLiveSet.add(e.ID)
   989  		}
   990  		for _, v := range b.ControlValues() {
   991  			if s.values[v.ID].needReg {
   992  				s.addUse(v.ID, int32(len(b.Values)), b.Pos) // pseudo-use by control values
   993  				regValLiveSet.add(v.ID)
   994  			}
   995  		}
   996  		if len(s.nextCall) < len(b.Values) {
   997  			s.nextCall = append(s.nextCall, make([]int32, len(b.Values)-len(s.nextCall))...)
   998  		}
   999  		var nextCall int32 = math.MaxInt32
  1000  		for i := len(b.Values) - 1; i >= 0; i-- {
  1001  			v := b.Values[i]
  1002  			regValLiveSet.remove(v.ID)
  1003  			if v.Op == OpPhi {
  1004  				// Remove v from the live set, but don't add
  1005  				// any inputs. This is the state the len(b.Preds)>1
  1006  				// case below desires; it wants to process phis specially.
  1007  				s.nextCall[i] = nextCall
  1008  				continue
  1009  			}
  1010  			if opcodeTable[v.Op].call {
  1011  				// Function call clobbers all the registers but SP and SB.
  1012  				regValLiveSet.clear()
  1013  				if s.sp != 0 && s.values[s.sp].uses != nil {
  1014  					regValLiveSet.add(s.sp)
  1015  				}
  1016  				if s.sb != 0 && s.values[s.sb].uses != nil {
  1017  					regValLiveSet.add(s.sb)
  1018  				}
  1019  				nextCall = int32(i)
  1020  			}
  1021  			for _, a := range v.Args {
  1022  				if !s.values[a.ID].needReg {
  1023  					continue
  1024  				}
  1025  				s.addUse(a.ID, int32(i), v.Pos)
  1026  				regValLiveSet.add(a.ID)
  1027  			}
  1028  			s.nextCall[i] = nextCall
  1029  		}
  1030  		if s.f.pass.debug > regDebug {
  1031  			fmt.Printf("use distances for %s\n", b)
  1032  			for i := range s.values {
  1033  				vi := &s.values[i]
  1034  				u := vi.uses
  1035  				if u == nil {
  1036  					continue
  1037  				}
  1038  				fmt.Printf("  v%d:", i)
  1039  				for u != nil {
  1040  					fmt.Printf(" %d", u.dist)
  1041  					u = u.next
  1042  				}
  1043  				fmt.Println()
  1044  			}
  1045  		}
  1046  
  1047  		// Make a copy of the block schedule so we can generate a new one in place.
  1048  		// We make a separate copy for phis and regular values.
  1049  		nphi := 0
  1050  		for _, v := range b.Values {
  1051  			if v.Op != OpPhi {
  1052  				break
  1053  			}
  1054  			nphi++
  1055  		}
  1056  		phis = append(phis[:0], b.Values[:nphi]...)
  1057  		oldSched = append(oldSched[:0], b.Values[nphi:]...)
  1058  		b.Values = b.Values[:0]
  1059  
  1060  		// Initialize start state of block.
  1061  		if b == f.Entry {
  1062  			// Regalloc state is empty to start.
  1063  			if nphi > 0 {
  1064  				f.Fatalf("phis in entry block")
  1065  			}
  1066  		} else if len(b.Preds) == 1 {
  1067  			// Start regalloc state with the end state of the previous block.
  1068  			s.setState(s.endRegs[b.Preds[0].b.ID])
  1069  			if nphi > 0 {
  1070  				f.Fatalf("phis in single-predecessor block")
  1071  			}
  1072  			// Drop any values which are no longer live.
  1073  			// This may happen because at the end of p, a value may be
  1074  			// live but only used by some other successor of p.
  1075  			for r := register(0); r < s.numRegs; r++ {
  1076  				v := s.regs[r].v
  1077  				if v != nil && !regValLiveSet.contains(v.ID) {
  1078  					s.freeReg(r)
  1079  				}
  1080  			}
  1081  		} else {
  1082  			// This is the complicated case. We have more than one predecessor,
  1083  			// which means we may have Phi ops.
  1084  
  1085  			// Start with the final register state of the predecessor with least spill values.
  1086  			// This is based on the following points:
  1087  			// 1, The less spill value indicates that the register pressure of this path is smaller,
  1088  			//    so the values of this block are more likely to be allocated to registers.
  1089  			// 2, Avoid the predecessor that contains the function call, because the predecessor that
  1090  			//    contains the function call usually generates a lot of spills and lose the previous
  1091  			//    allocation state.
  1092  			// TODO: Improve this part. At least the size of endRegs of the predecessor also has
  1093  			// an impact on the code size and compiler speed. But it is not easy to find a simple
  1094  			// and efficient method that combines multiple factors.
  1095  			idx := -1
  1096  			for i, p := range b.Preds {
  1097  				// If the predecessor has not been visited yet, skip it because its end state
  1098  				// (redRegs and spillLive) has not been computed yet.
  1099  				pb := p.b
  1100  				if s.blockOrder[pb.ID] >= s.blockOrder[b.ID] {
  1101  					continue
  1102  				}
  1103  				if idx == -1 {
  1104  					idx = i
  1105  					continue
  1106  				}
  1107  				pSel := b.Preds[idx].b
  1108  				if len(s.spillLive[pb.ID]) < len(s.spillLive[pSel.ID]) {
  1109  					idx = i
  1110  				} else if len(s.spillLive[pb.ID]) == len(s.spillLive[pSel.ID]) {
  1111  					// Use a bit of likely information. After critical pass, pb and pSel must
  1112  					// be plain blocks, so check edge pb->pb.Preds instead of edge pb->b.
  1113  					// TODO: improve the prediction of the likely predecessor. The following
  1114  					// method is only suitable for the simplest cases. For complex cases,
  1115  					// the prediction may be inaccurate, but this does not affect the
  1116  					// correctness of the program.
  1117  					// According to the layout algorithm, the predecessor with the
  1118  					// smaller blockOrder is the true branch, and the test results show
  1119  					// that it is better to choose the predecessor with a smaller
  1120  					// blockOrder than no choice.
  1121  					if pb.likelyBranch() && !pSel.likelyBranch() || s.blockOrder[pb.ID] < s.blockOrder[pSel.ID] {
  1122  						idx = i
  1123  					}
  1124  				}
  1125  			}
  1126  			if idx < 0 {
  1127  				f.Fatalf("bad visitOrder, no predecessor of %s has been visited before it", b)
  1128  			}
  1129  			p := b.Preds[idx].b
  1130  			s.setState(s.endRegs[p.ID])
  1131  
  1132  			if s.f.pass.debug > regDebug {
  1133  				fmt.Printf("starting merge block %s with end state of %s:\n", b, p)
  1134  				for _, x := range s.endRegs[p.ID] {
  1135  					fmt.Printf("  %s: orig:%s cache:%s\n", &s.registers[x.r], x.v, x.c)
  1136  				}
  1137  			}
  1138  
  1139  			// Decide on registers for phi ops. Use the registers determined
  1140  			// by the primary predecessor if we can.
  1141  			// TODO: pick best of (already processed) predecessors?
  1142  			// Majority vote? Deepest nesting level?
  1143  			phiRegs = phiRegs[:0]
  1144  			var phiUsed regMask
  1145  
  1146  			for _, v := range phis {
  1147  				if !s.values[v.ID].needReg {
  1148  					phiRegs = append(phiRegs, noRegister)
  1149  					continue
  1150  				}
  1151  				a := v.Args[idx]
  1152  				// Some instructions target not-allocatable registers.
  1153  				// They're not suitable for further (phi-function) allocation.
  1154  				m := s.values[a.ID].regs &^ phiUsed & s.allocatable
  1155  				if m != 0 {
  1156  					r := pickReg(m)
  1157  					phiUsed |= regMask(1) << r
  1158  					phiRegs = append(phiRegs, r)
  1159  				} else {
  1160  					phiRegs = append(phiRegs, noRegister)
  1161  				}
  1162  			}
  1163  
  1164  			// Second pass - deallocate all in-register phi inputs.
  1165  			for i, v := range phis {
  1166  				if !s.values[v.ID].needReg {
  1167  					continue
  1168  				}
  1169  				a := v.Args[idx]
  1170  				r := phiRegs[i]
  1171  				if r == noRegister {
  1172  					continue
  1173  				}
  1174  				if regValLiveSet.contains(a.ID) {
  1175  					// Input value is still live (it is used by something other than Phi).
  1176  					// Try to move it around before kicking out, if there is a free register.
  1177  					// We generate a Copy in the predecessor block and record it. It will be
  1178  					// deleted later if never used.
  1179  					//
  1180  					// Pick a free register. At this point some registers used in the predecessor
  1181  					// block may have been deallocated. Those are the ones used for Phis. Exclude
  1182  					// them (and they are not going to be helpful anyway).
  1183  					m := s.compatRegs(a.Type) &^ s.used &^ phiUsed
  1184  					if m != 0 && !s.values[a.ID].rematerializeable && countRegs(s.values[a.ID].regs) == 1 {
  1185  						r2 := pickReg(m)
  1186  						c := p.NewValue1(a.Pos, OpCopy, a.Type, s.regs[r].c)
  1187  						s.copies[c] = false
  1188  						if s.f.pass.debug > regDebug {
  1189  							fmt.Printf("copy %s to %s : %s\n", a, c, &s.registers[r2])
  1190  						}
  1191  						s.setOrig(c, a)
  1192  						s.assignReg(r2, a, c)
  1193  						s.endRegs[p.ID] = append(s.endRegs[p.ID], endReg{r2, a, c})
  1194  					}
  1195  				}
  1196  				s.freeReg(r)
  1197  			}
  1198  
  1199  			// Copy phi ops into new schedule.
  1200  			b.Values = append(b.Values, phis...)
  1201  
  1202  			// Third pass - pick registers for phis whose input
  1203  			// was not in a register in the primary predecessor.
  1204  			for i, v := range phis {
  1205  				if !s.values[v.ID].needReg {
  1206  					continue
  1207  				}
  1208  				if phiRegs[i] != noRegister {
  1209  					continue
  1210  				}
  1211  				m := s.compatRegs(v.Type) &^ phiUsed &^ s.used
  1212  				// If one of the other inputs of v is in a register, and the register is available,
  1213  				// select this register, which can save some unnecessary copies.
  1214  				for i, pe := range b.Preds {
  1215  					if i == idx {
  1216  						continue
  1217  					}
  1218  					ri := noRegister
  1219  					for _, er := range s.endRegs[pe.b.ID] {
  1220  						if er.v == s.orig[v.Args[i].ID] {
  1221  							ri = er.r
  1222  							break
  1223  						}
  1224  					}
  1225  					if ri != noRegister && m>>ri&1 != 0 {
  1226  						m = regMask(1) << ri
  1227  						break
  1228  					}
  1229  				}
  1230  				if m != 0 {
  1231  					r := pickReg(m)
  1232  					phiRegs[i] = r
  1233  					phiUsed |= regMask(1) << r
  1234  				}
  1235  			}
  1236  
  1237  			// Set registers for phis. Add phi spill code.
  1238  			for i, v := range phis {
  1239  				if !s.values[v.ID].needReg {
  1240  					continue
  1241  				}
  1242  				r := phiRegs[i]
  1243  				if r == noRegister {
  1244  					// stack-based phi
  1245  					// Spills will be inserted in all the predecessors below.
  1246  					s.values[v.ID].spill = v // v starts life spilled
  1247  					continue
  1248  				}
  1249  				// register-based phi
  1250  				s.assignReg(r, v, v)
  1251  			}
  1252  
  1253  			// Deallocate any values which are no longer live. Phis are excluded.
  1254  			for r := register(0); r < s.numRegs; r++ {
  1255  				if phiUsed>>r&1 != 0 {
  1256  					continue
  1257  				}
  1258  				v := s.regs[r].v
  1259  				if v != nil && !regValLiveSet.contains(v.ID) {
  1260  					s.freeReg(r)
  1261  				}
  1262  			}
  1263  
  1264  			// Save the starting state for use by merge edges.
  1265  			// We append to a stack allocated variable that we'll
  1266  			// later copy into s.startRegs in one fell swoop, to save
  1267  			// on allocations.
  1268  			regList := make([]startReg, 0, 32)
  1269  			for r := register(0); r < s.numRegs; r++ {
  1270  				v := s.regs[r].v
  1271  				if v == nil {
  1272  					continue
  1273  				}
  1274  				if phiUsed>>r&1 != 0 {
  1275  					// Skip registers that phis used, we'll handle those
  1276  					// specially during merge edge processing.
  1277  					continue
  1278  				}
  1279  				regList = append(regList, startReg{r, v, s.regs[r].c, s.values[v.ID].uses.pos})
  1280  				s.startRegsMask |= regMask(1) << r
  1281  			}
  1282  			s.startRegs[b.ID] = make([]startReg, len(regList))
  1283  			copy(s.startRegs[b.ID], regList)
  1284  
  1285  			if s.f.pass.debug > regDebug {
  1286  				fmt.Printf("after phis\n")
  1287  				for _, x := range s.startRegs[b.ID] {
  1288  					fmt.Printf("  %s: v%d\n", &s.registers[x.r], x.v.ID)
  1289  				}
  1290  			}
  1291  		}
  1292  
  1293  		// Drop phis from registers if they immediately go dead.
  1294  		for i, v := range phis {
  1295  			s.curIdx = i
  1296  			s.dropIfUnused(v)
  1297  		}
  1298  
  1299  		// Allocate space to record the desired registers for each value.
  1300  		if l := len(oldSched); cap(dinfo) < l {
  1301  			dinfo = make([]dentry, l)
  1302  		} else {
  1303  			dinfo = dinfo[:l]
  1304  			clear(dinfo)
  1305  		}
  1306  
  1307  		// Load static desired register info at the end of the block.
  1308  		desired.copy(&s.desired[b.ID])
  1309  
  1310  		// Check actual assigned registers at the start of the next block(s).
  1311  		// Dynamically assigned registers will trump the static
  1312  		// desired registers computed during liveness analysis.
  1313  		// Note that we do this phase after startRegs is set above, so that
  1314  		// we get the right behavior for a block which branches to itself.
  1315  		for _, e := range b.Succs {
  1316  			succ := e.b
  1317  			// TODO: prioritize likely successor?
  1318  			for _, x := range s.startRegs[succ.ID] {
  1319  				desired.add(x.v.ID, x.r)
  1320  			}
  1321  			// Process phi ops in succ.
  1322  			pidx := e.i
  1323  			for _, v := range succ.Values {
  1324  				if v.Op != OpPhi {
  1325  					break
  1326  				}
  1327  				if !s.values[v.ID].needReg {
  1328  					continue
  1329  				}
  1330  				rp, ok := s.f.getHome(v.ID).(*Register)
  1331  				if !ok {
  1332  					// If v is not assigned a register, pick a register assigned to one of v's inputs.
  1333  					// Hopefully v will get assigned that register later.
  1334  					// If the inputs have allocated register information, add it to desired,
  1335  					// which may reduce spill or copy operations when the register is available.
  1336  					for _, a := range v.Args {
  1337  						rp, ok = s.f.getHome(a.ID).(*Register)
  1338  						if ok {
  1339  							break
  1340  						}
  1341  					}
  1342  					if !ok {
  1343  						continue
  1344  					}
  1345  				}
  1346  				desired.add(v.Args[pidx].ID, register(rp.num))
  1347  			}
  1348  		}
  1349  		// Walk values backwards computing desired register info.
  1350  		// See computeLive for more comments.
  1351  		for i := len(oldSched) - 1; i >= 0; i-- {
  1352  			v := oldSched[i]
  1353  			prefs := desired.remove(v.ID)
  1354  			regspec := s.regspec(v)
  1355  			desired.clobber(regspec.clobbers)
  1356  			for _, j := range regspec.inputs {
  1357  				if countRegs(j.regs) != 1 {
  1358  					continue
  1359  				}
  1360  				desired.clobber(j.regs)
  1361  				desired.add(v.Args[j.idx].ID, pickReg(j.regs))
  1362  			}
  1363  			if opcodeTable[v.Op].resultInArg0 || v.Op == OpAMD64ADDQconst || v.Op == OpAMD64ADDLconst || v.Op == OpSelect0 {
  1364  				if opcodeTable[v.Op].commutative {
  1365  					desired.addList(v.Args[1].ID, prefs)
  1366  				}
  1367  				desired.addList(v.Args[0].ID, prefs)
  1368  			}
  1369  			// Save desired registers for this value.
  1370  			dinfo[i].out = prefs
  1371  			for j, a := range v.Args {
  1372  				if j >= len(dinfo[i].in) {
  1373  					break
  1374  				}
  1375  				dinfo[i].in[j] = desired.get(a.ID)
  1376  			}
  1377  			if v.Op == OpSelect1 && prefs[0] != noRegister {
  1378  				// Save desired registers of select1 for
  1379  				// use by the tuple generating instruction.
  1380  				desiredSecondReg[v.Args[0].ID] = prefs
  1381  			}
  1382  		}
  1383  
  1384  		// Process all the non-phi values.
  1385  		for idx, v := range oldSched {
  1386  			s.curIdx = nphi + idx
  1387  			tmpReg := noRegister
  1388  			if s.f.pass.debug > regDebug {
  1389  				fmt.Printf("  processing %s\n", v.LongString())
  1390  			}
  1391  			regspec := s.regspec(v)
  1392  			if v.Op == OpPhi {
  1393  				f.Fatalf("phi %s not at start of block", v)
  1394  			}
  1395  			if opcodeTable[v.Op].fixedReg {
  1396  				switch v.Op {
  1397  				case OpSP:
  1398  					s.assignReg(s.SPReg, v, v)
  1399  					s.sp = v.ID
  1400  				case OpSB:
  1401  					s.assignReg(s.SBReg, v, v)
  1402  					s.sb = v.ID
  1403  				case OpARM64ZERO:
  1404  					s.assignReg(s.ZeroIntReg, v, v)
  1405  				default:
  1406  					f.Fatalf("unknown fixed-register op %s", v)
  1407  				}
  1408  				b.Values = append(b.Values, v)
  1409  				s.advanceUses(v)
  1410  				continue
  1411  			}
  1412  			if v.Op == OpSelect0 || v.Op == OpSelect1 || v.Op == OpSelectN {
  1413  				if s.values[v.ID].needReg {
  1414  					if v.Op == OpSelectN {
  1415  						s.assignReg(register(s.f.getHome(v.Args[0].ID).(LocResults)[int(v.AuxInt)].(*Register).num), v, v)
  1416  					} else {
  1417  						var i = 0
  1418  						if v.Op == OpSelect1 {
  1419  							i = 1
  1420  						}
  1421  						s.assignReg(register(s.f.getHome(v.Args[0].ID).(LocPair)[i].(*Register).num), v, v)
  1422  					}
  1423  				}
  1424  				b.Values = append(b.Values, v)
  1425  				s.advanceUses(v)
  1426  				continue
  1427  			}
  1428  			if v.Op == OpGetG && s.f.Config.hasGReg {
  1429  				// use hardware g register
  1430  				if s.regs[s.GReg].v != nil {
  1431  					s.freeReg(s.GReg) // kick out the old value
  1432  				}
  1433  				s.assignReg(s.GReg, v, v)
  1434  				b.Values = append(b.Values, v)
  1435  				s.advanceUses(v)
  1436  				continue
  1437  			}
  1438  			if v.Op == OpArg {
  1439  				// Args are "pre-spilled" values. We don't allocate
  1440  				// any register here. We just set up the spill pointer to
  1441  				// point at itself and any later user will restore it to use it.
  1442  				s.values[v.ID].spill = v
  1443  				b.Values = append(b.Values, v)
  1444  				s.advanceUses(v)
  1445  				continue
  1446  			}
  1447  			if v.Op == OpKeepAlive {
  1448  				// Make sure the argument to v is still live here.
  1449  				s.advanceUses(v)
  1450  				a := v.Args[0]
  1451  				vi := &s.values[a.ID]
  1452  				if vi.regs == 0 && !vi.rematerializeable {
  1453  					// Use the spill location.
  1454  					// This forces later liveness analysis to make the
  1455  					// value live at this point.
  1456  					v.SetArg(0, s.makeSpill(a, b))
  1457  				} else if _, ok := a.Aux.(*ir.Name); ok && vi.rematerializeable {
  1458  					// Rematerializeable value with a *ir.Name. This is the address of
  1459  					// a stack object (e.g. an LEAQ). Keep the object live.
  1460  					// Change it to VarLive, which is what plive expects for locals.
  1461  					v.Op = OpVarLive
  1462  					v.SetArgs1(v.Args[1])
  1463  					v.Aux = a.Aux
  1464  				} else {
  1465  					// In-register and rematerializeable values are already live.
  1466  					// These are typically rematerializeable constants like nil,
  1467  					// or values of a variable that were modified since the last call.
  1468  					v.Op = OpCopy
  1469  					v.SetArgs1(v.Args[1])
  1470  				}
  1471  				b.Values = append(b.Values, v)
  1472  				continue
  1473  			}
  1474  			if len(regspec.inputs) == 0 && len(regspec.outputs) == 0 {
  1475  				// No register allocation required (or none specified yet)
  1476  				if s.doClobber && v.Op.IsCall() {
  1477  					s.clobberRegs(regspec.clobbers)
  1478  				}
  1479  				s.freeRegs(regspec.clobbers)
  1480  				b.Values = append(b.Values, v)
  1481  				s.advanceUses(v)
  1482  				continue
  1483  			}
  1484  
  1485  			if s.values[v.ID].rematerializeable {
  1486  				// Value is rematerializeable, don't issue it here.
  1487  				// It will get issued just before each use (see
  1488  				// allocValueToReg).
  1489  				for _, a := range v.Args {
  1490  					a.Uses--
  1491  				}
  1492  				s.advanceUses(v)
  1493  				continue
  1494  			}
  1495  
  1496  			if s.f.pass.debug > regDebug {
  1497  				fmt.Printf("value %s\n", v.LongString())
  1498  				fmt.Printf("  out:")
  1499  				for _, r := range dinfo[idx].out {
  1500  					if r != noRegister {
  1501  						fmt.Printf(" %s", &s.registers[r])
  1502  					}
  1503  				}
  1504  				fmt.Println()
  1505  				for i := 0; i < len(v.Args) && i < 3; i++ {
  1506  					fmt.Printf("  in%d:", i)
  1507  					for _, r := range dinfo[idx].in[i] {
  1508  						if r != noRegister {
  1509  							fmt.Printf(" %s", &s.registers[r])
  1510  						}
  1511  					}
  1512  					fmt.Println()
  1513  				}
  1514  			}
  1515  
  1516  			// Move arguments to registers.
  1517  			// First, if an arg must be in a specific register and it is already
  1518  			// in place, keep it.
  1519  			args = append(args[:0], make([]*Value, len(v.Args))...)
  1520  			for i, a := range v.Args {
  1521  				if !s.values[a.ID].needReg {
  1522  					args[i] = a
  1523  				}
  1524  			}
  1525  			for _, i := range regspec.inputs {
  1526  				mask := i.regs
  1527  				if countRegs(mask) == 1 && mask&s.values[v.Args[i.idx].ID].regs != 0 {
  1528  					args[i.idx] = s.allocValToReg(v.Args[i.idx], mask, true, v.Pos)
  1529  				}
  1530  			}
  1531  			// Then, if an arg must be in a specific register and that
  1532  			// register is free, allocate that one. Otherwise when processing
  1533  			// another input we may kick a value into the free register, which
  1534  			// then will be kicked out again.
  1535  			// This is a common case for passing-in-register arguments for
  1536  			// function calls.
  1537  			for {
  1538  				freed := false
  1539  				for _, i := range regspec.inputs {
  1540  					if args[i.idx] != nil {
  1541  						continue // already allocated
  1542  					}
  1543  					mask := i.regs
  1544  					if countRegs(mask) == 1 && mask&^s.used != 0 {
  1545  						args[i.idx] = s.allocValToReg(v.Args[i.idx], mask, true, v.Pos)
  1546  						// If the input is in other registers that will be clobbered by v,
  1547  						// or the input is dead, free the registers. This may make room
  1548  						// for other inputs.
  1549  						oldregs := s.values[v.Args[i.idx].ID].regs
  1550  						if oldregs&^regspec.clobbers == 0 || !s.liveAfterCurrentInstruction(v.Args[i.idx]) {
  1551  							s.freeRegs(oldregs &^ mask &^ s.nospill)
  1552  							freed = true
  1553  						}
  1554  					}
  1555  				}
  1556  				if !freed {
  1557  					break
  1558  				}
  1559  			}
  1560  			// Last, allocate remaining ones, in an ordering defined
  1561  			// by the register specification (most constrained first).
  1562  			for _, i := range regspec.inputs {
  1563  				if args[i.idx] != nil {
  1564  					continue // already allocated
  1565  				}
  1566  				mask := i.regs
  1567  				if mask&s.values[v.Args[i.idx].ID].regs == 0 {
  1568  					// Need a new register for the input.
  1569  					mask &= s.allocatable
  1570  					mask &^= s.nospill
  1571  					// Used desired register if available.
  1572  					if i.idx < 3 {
  1573  						for _, r := range dinfo[idx].in[i.idx] {
  1574  							if r != noRegister && (mask&^s.used)>>r&1 != 0 {
  1575  								// Desired register is allowed and unused.
  1576  								mask = regMask(1) << r
  1577  								break
  1578  							}
  1579  						}
  1580  					}
  1581  					// Avoid registers we're saving for other values.
  1582  					if mask&^desired.avoid != 0 {
  1583  						mask &^= desired.avoid
  1584  					}
  1585  				}
  1586  				args[i.idx] = s.allocValToReg(v.Args[i.idx], mask, true, v.Pos)
  1587  			}
  1588  
  1589  			// If the output clobbers the input register, make sure we have
  1590  			// at least two copies of the input register so we don't
  1591  			// have to reload the value from the spill location.
  1592  			if opcodeTable[v.Op].resultInArg0 {
  1593  				var m regMask
  1594  				if !s.liveAfterCurrentInstruction(v.Args[0]) {
  1595  					// arg0 is dead.  We can clobber its register.
  1596  					goto ok
  1597  				}
  1598  				if opcodeTable[v.Op].commutative && !s.liveAfterCurrentInstruction(v.Args[1]) {
  1599  					args[0], args[1] = args[1], args[0]
  1600  					goto ok
  1601  				}
  1602  				if s.values[v.Args[0].ID].rematerializeable {
  1603  					// We can rematerialize the input, don't worry about clobbering it.
  1604  					goto ok
  1605  				}
  1606  				if opcodeTable[v.Op].commutative && s.values[v.Args[1].ID].rematerializeable {
  1607  					args[0], args[1] = args[1], args[0]
  1608  					goto ok
  1609  				}
  1610  				if countRegs(s.values[v.Args[0].ID].regs) >= 2 {
  1611  					// we have at least 2 copies of arg0.  We can afford to clobber one.
  1612  					goto ok
  1613  				}
  1614  				if opcodeTable[v.Op].commutative && countRegs(s.values[v.Args[1].ID].regs) >= 2 {
  1615  					args[0], args[1] = args[1], args[0]
  1616  					goto ok
  1617  				}
  1618  
  1619  				// We can't overwrite arg0 (or arg1, if commutative).  So we
  1620  				// need to make a copy of an input so we have a register we can modify.
  1621  
  1622  				// Possible new registers to copy into.
  1623  				m = s.compatRegs(v.Args[0].Type) &^ s.used
  1624  				if m == 0 {
  1625  					// No free registers.  In this case we'll just clobber
  1626  					// an input and future uses of that input must use a restore.
  1627  					// TODO(khr): We should really do this like allocReg does it,
  1628  					// spilling the value with the most distant next use.
  1629  					goto ok
  1630  				}
  1631  
  1632  				// Try to move an input to the desired output, if allowed.
  1633  				for _, r := range dinfo[idx].out {
  1634  					if r != noRegister && (m&regspec.outputs[0].regs)>>r&1 != 0 {
  1635  						m = regMask(1) << r
  1636  						args[0] = s.allocValToReg(v.Args[0], m, true, v.Pos)
  1637  						// Note: we update args[0] so the instruction will
  1638  						// use the register copy we just made.
  1639  						goto ok
  1640  					}
  1641  				}
  1642  				// Try to copy input to its desired location & use its old
  1643  				// location as the result register.
  1644  				for _, r := range dinfo[idx].in[0] {
  1645  					if r != noRegister && m>>r&1 != 0 {
  1646  						m = regMask(1) << r
  1647  						c := s.allocValToReg(v.Args[0], m, true, v.Pos)
  1648  						s.copies[c] = false
  1649  						// Note: no update to args[0] so the instruction will
  1650  						// use the original copy.
  1651  						goto ok
  1652  					}
  1653  				}
  1654  				if opcodeTable[v.Op].commutative {
  1655  					for _, r := range dinfo[idx].in[1] {
  1656  						if r != noRegister && m>>r&1 != 0 {
  1657  							m = regMask(1) << r
  1658  							c := s.allocValToReg(v.Args[1], m, true, v.Pos)
  1659  							s.copies[c] = false
  1660  							args[0], args[1] = args[1], args[0]
  1661  							goto ok
  1662  						}
  1663  					}
  1664  				}
  1665  
  1666  				// Avoid future fixed uses if we can.
  1667  				if m&^desired.avoid != 0 {
  1668  					m &^= desired.avoid
  1669  				}
  1670  				// Save input 0 to a new register so we can clobber it.
  1671  				c := s.allocValToReg(v.Args[0], m, true, v.Pos)
  1672  				s.copies[c] = false
  1673  
  1674  				// Normally we use the register of the old copy of input 0 as the target.
  1675  				// However, if input 0 is already in its desired register then we use
  1676  				// the register of the new copy instead.
  1677  				if regspec.outputs[0].regs>>s.f.getHome(c.ID).(*Register).num&1 != 0 {
  1678  					if rp, ok := s.f.getHome(args[0].ID).(*Register); ok {
  1679  						r := register(rp.num)
  1680  						for _, r2 := range dinfo[idx].in[0] {
  1681  							if r == r2 {
  1682  								args[0] = c
  1683  								break
  1684  							}
  1685  						}
  1686  					}
  1687  				}
  1688  			}
  1689  
  1690  		ok:
  1691  			// Pick a temporary register if needed.
  1692  			// It should be distinct from all the input registers, so we
  1693  			// allocate it after all the input registers, but before
  1694  			// the input registers are freed via advanceUses below.
  1695  			// (Not all instructions need that distinct part, but it is conservative.)
  1696  			// We also ensure it is not any of the single-choice output registers.
  1697  			if opcodeTable[v.Op].needIntTemp {
  1698  				m := s.allocatable & s.f.Config.gpRegMask
  1699  				for _, out := range regspec.outputs {
  1700  					if countRegs(out.regs) == 1 {
  1701  						m &^= out.regs
  1702  					}
  1703  				}
  1704  				if m&^desired.avoid&^s.nospill != 0 {
  1705  					m &^= desired.avoid
  1706  				}
  1707  				tmpReg = s.allocReg(m, &tmpVal)
  1708  				s.nospill |= regMask(1) << tmpReg
  1709  				s.tmpused |= regMask(1) << tmpReg
  1710  			}
  1711  
  1712  			// Now that all args are in regs, we're ready to issue the value itself.
  1713  			// Before we pick a register for the output value, allow input registers
  1714  			// to be deallocated. We do this here so that the output can use the
  1715  			// same register as a dying input.
  1716  			if !opcodeTable[v.Op].resultNotInArgs {
  1717  				s.tmpused = s.nospill
  1718  				s.nospill = 0
  1719  				s.advanceUses(v) // frees any registers holding args that are no longer live
  1720  			}
  1721  
  1722  			// Dump any registers which will be clobbered
  1723  			if s.doClobber && v.Op.IsCall() {
  1724  				// clobber registers that are marked as clobber in regmask, but
  1725  				// don't clobber inputs.
  1726  				s.clobberRegs(regspec.clobbers &^ s.tmpused &^ s.nospill)
  1727  			}
  1728  			s.freeRegs(regspec.clobbers)
  1729  			s.tmpused |= regspec.clobbers
  1730  
  1731  			// Pick registers for outputs.
  1732  			{
  1733  				outRegs := noRegisters // TODO if this is costly, hoist and clear incrementally below.
  1734  				maxOutIdx := -1
  1735  				var used regMask
  1736  				if tmpReg != noRegister {
  1737  					// Ensure output registers are distinct from the temporary register.
  1738  					// (Not all instructions need that distinct part, but it is conservative.)
  1739  					used |= regMask(1) << tmpReg
  1740  				}
  1741  				for _, out := range regspec.outputs {
  1742  					if out.regs == 0 {
  1743  						continue
  1744  					}
  1745  					mask := out.regs & s.allocatable &^ used
  1746  					if mask == 0 {
  1747  						s.f.Fatalf("can't find any output register %s", v.LongString())
  1748  					}
  1749  					if opcodeTable[v.Op].resultInArg0 && out.idx == 0 {
  1750  						if !opcodeTable[v.Op].commutative {
  1751  							// Output must use the same register as input 0.
  1752  							r := register(s.f.getHome(args[0].ID).(*Register).num)
  1753  							if mask>>r&1 == 0 {
  1754  								s.f.Fatalf("resultInArg0 value's input %v cannot be an output of %s", s.f.getHome(args[0].ID).(*Register), v.LongString())
  1755  							}
  1756  							mask = regMask(1) << r
  1757  						} else {
  1758  							// Output must use the same register as input 0 or 1.
  1759  							r0 := register(s.f.getHome(args[0].ID).(*Register).num)
  1760  							r1 := register(s.f.getHome(args[1].ID).(*Register).num)
  1761  							// Check r0 and r1 for desired output register.
  1762  							found := false
  1763  							for _, r := range dinfo[idx].out {
  1764  								if (r == r0 || r == r1) && (mask&^s.used)>>r&1 != 0 {
  1765  									mask = regMask(1) << r
  1766  									found = true
  1767  									if r == r1 {
  1768  										args[0], args[1] = args[1], args[0]
  1769  									}
  1770  									break
  1771  								}
  1772  							}
  1773  							if !found {
  1774  								// Neither are desired, pick r0.
  1775  								mask = regMask(1) << r0
  1776  							}
  1777  						}
  1778  					}
  1779  					if out.idx == 0 { // desired registers only apply to the first element of a tuple result
  1780  						for _, r := range dinfo[idx].out {
  1781  							if r != noRegister && (mask&^s.used)>>r&1 != 0 {
  1782  								// Desired register is allowed and unused.
  1783  								mask = regMask(1) << r
  1784  								break
  1785  							}
  1786  						}
  1787  					}
  1788  					if out.idx == 1 {
  1789  						if prefs, ok := desiredSecondReg[v.ID]; ok {
  1790  							for _, r := range prefs {
  1791  								if r != noRegister && (mask&^s.used)>>r&1 != 0 {
  1792  									// Desired register is allowed and unused.
  1793  									mask = regMask(1) << r
  1794  									break
  1795  								}
  1796  							}
  1797  						}
  1798  					}
  1799  					// Avoid registers we're saving for other values.
  1800  					if mask&^desired.avoid&^s.nospill&^s.used != 0 {
  1801  						mask &^= desired.avoid
  1802  					}
  1803  					r := s.allocReg(mask, v)
  1804  					if out.idx > maxOutIdx {
  1805  						maxOutIdx = out.idx
  1806  					}
  1807  					outRegs[out.idx] = r
  1808  					used |= regMask(1) << r
  1809  					s.tmpused |= regMask(1) << r
  1810  				}
  1811  				// Record register choices
  1812  				if v.Type.IsTuple() {
  1813  					var outLocs LocPair
  1814  					if r := outRegs[0]; r != noRegister {
  1815  						outLocs[0] = &s.registers[r]
  1816  					}
  1817  					if r := outRegs[1]; r != noRegister {
  1818  						outLocs[1] = &s.registers[r]
  1819  					}
  1820  					s.f.setHome(v, outLocs)
  1821  					// Note that subsequent SelectX instructions will do the assignReg calls.
  1822  				} else if v.Type.IsResults() {
  1823  					// preallocate outLocs to the right size, which is maxOutIdx+1
  1824  					outLocs := make(LocResults, maxOutIdx+1, maxOutIdx+1)
  1825  					for i := 0; i <= maxOutIdx; i++ {
  1826  						if r := outRegs[i]; r != noRegister {
  1827  							outLocs[i] = &s.registers[r]
  1828  						}
  1829  					}
  1830  					s.f.setHome(v, outLocs)
  1831  				} else {
  1832  					if r := outRegs[0]; r != noRegister {
  1833  						s.assignReg(r, v, v)
  1834  					}
  1835  				}
  1836  				if tmpReg != noRegister {
  1837  					// Remember the temp register allocation, if any.
  1838  					if s.f.tempRegs == nil {
  1839  						s.f.tempRegs = map[ID]*Register{}
  1840  					}
  1841  					s.f.tempRegs[v.ID] = &s.registers[tmpReg]
  1842  				}
  1843  			}
  1844  
  1845  			// deallocate dead args, if we have not done so
  1846  			if opcodeTable[v.Op].resultNotInArgs {
  1847  				s.nospill = 0
  1848  				s.advanceUses(v) // frees any registers holding args that are no longer live
  1849  			}
  1850  			s.tmpused = 0
  1851  
  1852  			// Issue the Value itself.
  1853  			for i, a := range args {
  1854  				v.SetArg(i, a) // use register version of arguments
  1855  			}
  1856  			b.Values = append(b.Values, v)
  1857  			s.dropIfUnused(v)
  1858  		}
  1859  
  1860  		// Copy the control values - we need this so we can reduce the
  1861  		// uses property of these values later.
  1862  		controls := append(make([]*Value, 0, 2), b.ControlValues()...)
  1863  
  1864  		// Load control values into registers.
  1865  		for i, v := range b.ControlValues() {
  1866  			if !s.values[v.ID].needReg {
  1867  				continue
  1868  			}
  1869  			if s.f.pass.debug > regDebug {
  1870  				fmt.Printf("  processing control %s\n", v.LongString())
  1871  			}
  1872  			// We assume that a control input can be passed in any
  1873  			// type-compatible register. If this turns out not to be true,
  1874  			// we'll need to introduce a regspec for a block's control value.
  1875  			b.ReplaceControl(i, s.allocValToReg(v, s.compatRegs(v.Type), false, b.Pos))
  1876  		}
  1877  
  1878  		// Reduce the uses of the control values once registers have been loaded.
  1879  		// This loop is equivalent to the advanceUses method.
  1880  		for _, v := range controls {
  1881  			vi := &s.values[v.ID]
  1882  			if !vi.needReg {
  1883  				continue
  1884  			}
  1885  			// Remove this use from the uses list.
  1886  			u := vi.uses
  1887  			vi.uses = u.next
  1888  			if u.next == nil {
  1889  				s.freeRegs(vi.regs) // value is dead
  1890  			}
  1891  			u.next = s.freeUseRecords
  1892  			s.freeUseRecords = u
  1893  		}
  1894  
  1895  		// If we are approaching a merge point and we are the primary
  1896  		// predecessor of it, find live values that we use soon after
  1897  		// the merge point and promote them to registers now.
  1898  		if len(b.Succs) == 1 {
  1899  			if s.f.Config.hasGReg && s.regs[s.GReg].v != nil {
  1900  				s.freeReg(s.GReg) // Spill value in G register before any merge.
  1901  			}
  1902  			if s.blockOrder[b.ID] > s.blockOrder[b.Succs[0].b.ID] {
  1903  				// No point if we've already regalloc'd the destination.
  1904  				goto badloop
  1905  			}
  1906  			// For this to be worthwhile, the loop must have no calls in it.
  1907  			top := b.Succs[0].b
  1908  			loop := s.loopnest.b2l[top.ID]
  1909  			if loop == nil || loop.header != top || loop.containsUnavoidableCall {
  1910  				goto badloop
  1911  			}
  1912  
  1913  			// TODO: sort by distance, pick the closest ones?
  1914  			for _, live := range s.live[b.ID] {
  1915  				if live.dist >= unlikelyDistance {
  1916  					// Don't preload anything live after the loop.
  1917  					continue
  1918  				}
  1919  				vid := live.ID
  1920  				vi := &s.values[vid]
  1921  				if vi.regs != 0 {
  1922  					continue
  1923  				}
  1924  				if vi.rematerializeable {
  1925  					continue
  1926  				}
  1927  				v := s.orig[vid]
  1928  				m := s.compatRegs(v.Type) &^ s.used
  1929  				// Used desired register if available.
  1930  			outerloop:
  1931  				for _, e := range desired.entries {
  1932  					if e.ID != v.ID {
  1933  						continue
  1934  					}
  1935  					for _, r := range e.regs {
  1936  						if r != noRegister && m>>r&1 != 0 {
  1937  							m = regMask(1) << r
  1938  							break outerloop
  1939  						}
  1940  					}
  1941  				}
  1942  				if m&^desired.avoid != 0 {
  1943  					m &^= desired.avoid
  1944  				}
  1945  				if m != 0 {
  1946  					s.allocValToReg(v, m, false, b.Pos)
  1947  				}
  1948  			}
  1949  		}
  1950  	badloop:
  1951  		;
  1952  
  1953  		// Save end-of-block register state.
  1954  		// First count how many, this cuts allocations in half.
  1955  		k := 0
  1956  		for r := register(0); r < s.numRegs; r++ {
  1957  			v := s.regs[r].v
  1958  			if v == nil {
  1959  				continue
  1960  			}
  1961  			k++
  1962  		}
  1963  		regList := make([]endReg, 0, k)
  1964  		for r := register(0); r < s.numRegs; r++ {
  1965  			v := s.regs[r].v
  1966  			if v == nil {
  1967  				continue
  1968  			}
  1969  			regList = append(regList, endReg{r, v, s.regs[r].c})
  1970  		}
  1971  		s.endRegs[b.ID] = regList
  1972  
  1973  		if checkEnabled {
  1974  			regValLiveSet.clear()
  1975  			for _, x := range s.live[b.ID] {
  1976  				regValLiveSet.add(x.ID)
  1977  			}
  1978  			for r := register(0); r < s.numRegs; r++ {
  1979  				v := s.regs[r].v
  1980  				if v == nil {
  1981  					continue
  1982  				}
  1983  				if !regValLiveSet.contains(v.ID) {
  1984  					s.f.Fatalf("val %s is in reg but not live at end of %s", v, b)
  1985  				}
  1986  			}
  1987  		}
  1988  
  1989  		// If a value is live at the end of the block and
  1990  		// isn't in a register, generate a use for the spill location.
  1991  		// We need to remember this information so that
  1992  		// the liveness analysis in stackalloc is correct.
  1993  		for _, e := range s.live[b.ID] {
  1994  			vi := &s.values[e.ID]
  1995  			if vi.regs != 0 {
  1996  				// in a register, we'll use that source for the merge.
  1997  				continue
  1998  			}
  1999  			if vi.rematerializeable {
  2000  				// we'll rematerialize during the merge.
  2001  				continue
  2002  			}
  2003  			if s.f.pass.debug > regDebug {
  2004  				fmt.Printf("live-at-end spill for %s at %s\n", s.orig[e.ID], b)
  2005  			}
  2006  			spill := s.makeSpill(s.orig[e.ID], b)
  2007  			s.spillLive[b.ID] = append(s.spillLive[b.ID], spill.ID)
  2008  		}
  2009  
  2010  		// Clear any final uses.
  2011  		// All that is left should be the pseudo-uses added for values which
  2012  		// are live at the end of b.
  2013  		for _, e := range s.live[b.ID] {
  2014  			u := s.values[e.ID].uses
  2015  			if u == nil {
  2016  				f.Fatalf("live at end, no uses v%d", e.ID)
  2017  			}
  2018  			if u.next != nil {
  2019  				f.Fatalf("live at end, too many uses v%d", e.ID)
  2020  			}
  2021  			s.values[e.ID].uses = nil
  2022  			u.next = s.freeUseRecords
  2023  			s.freeUseRecords = u
  2024  		}
  2025  
  2026  		// allocReg may have dropped registers from startRegsMask that
  2027  		// aren't actually needed in startRegs. Synchronize back to
  2028  		// startRegs.
  2029  		//
  2030  		// This must be done before placing spills, which will look at
  2031  		// startRegs to decide if a block is a valid block for a spill.
  2032  		if c := countRegs(s.startRegsMask); c != len(s.startRegs[b.ID]) {
  2033  			regs := make([]startReg, 0, c)
  2034  			for _, sr := range s.startRegs[b.ID] {
  2035  				if s.startRegsMask&(regMask(1)<<sr.r) == 0 {
  2036  					continue
  2037  				}
  2038  				regs = append(regs, sr)
  2039  			}
  2040  			s.startRegs[b.ID] = regs
  2041  		}
  2042  	}
  2043  
  2044  	// Decide where the spills we generated will go.
  2045  	s.placeSpills()
  2046  
  2047  	// Anything that didn't get a register gets a stack location here.
  2048  	// (StoreReg, stack-based phis, inputs, ...)
  2049  	stacklive := stackalloc(s.f, s.spillLive)
  2050  
  2051  	// Fix up all merge edges.
  2052  	s.shuffle(stacklive)
  2053  
  2054  	// Erase any copies we never used.
  2055  	// Also, an unused copy might be the only use of another copy,
  2056  	// so continue erasing until we reach a fixed point.
  2057  	for {
  2058  		progress := false
  2059  		for c, used := range s.copies {
  2060  			if !used && c.Uses == 0 {
  2061  				if s.f.pass.debug > regDebug {
  2062  					fmt.Printf("delete copied value %s\n", c.LongString())
  2063  				}
  2064  				c.resetArgs()
  2065  				f.freeValue(c)
  2066  				delete(s.copies, c)
  2067  				progress = true
  2068  			}
  2069  		}
  2070  		if !progress {
  2071  			break
  2072  		}
  2073  	}
  2074  
  2075  	for _, b := range s.visitOrder {
  2076  		i := 0
  2077  		for _, v := range b.Values {
  2078  			if v.Op == OpInvalid {
  2079  				continue
  2080  			}
  2081  			b.Values[i] = v
  2082  			i++
  2083  		}
  2084  		b.Values = b.Values[:i]
  2085  	}
  2086  }
  2087  
  2088  func (s *regAllocState) placeSpills() {
  2089  	mustBeFirst := func(op Op) bool {
  2090  		return op.isLoweredGetClosurePtr() || op == OpPhi || op == OpArgIntReg || op == OpArgFloatReg
  2091  	}
  2092  
  2093  	// Start maps block IDs to the list of spills
  2094  	// that go at the start of the block (but after any phis).
  2095  	start := map[ID][]*Value{}
  2096  	// After maps value IDs to the list of spills
  2097  	// that go immediately after that value ID.
  2098  	after := map[ID][]*Value{}
  2099  
  2100  	for i := range s.values {
  2101  		vi := s.values[i]
  2102  		spill := vi.spill
  2103  		if spill == nil {
  2104  			continue
  2105  		}
  2106  		if spill.Block != nil {
  2107  			// Some spills are already fully set up,
  2108  			// like OpArgs and stack-based phis.
  2109  			continue
  2110  		}
  2111  		v := s.orig[i]
  2112  
  2113  		// Walk down the dominator tree looking for a good place to
  2114  		// put the spill of v.  At the start "best" is the best place
  2115  		// we have found so far.
  2116  		// TODO: find a way to make this O(1) without arbitrary cutoffs.
  2117  		if v == nil {
  2118  			panic(fmt.Errorf("nil v, s.orig[%d], vi = %v, spill = %s", i, vi, spill.LongString()))
  2119  		}
  2120  		best := v.Block
  2121  		bestArg := v
  2122  		var bestDepth int16
  2123  		if l := s.loopnest.b2l[best.ID]; l != nil {
  2124  			bestDepth = l.depth
  2125  		}
  2126  		b := best
  2127  		const maxSpillSearch = 100
  2128  		for i := 0; i < maxSpillSearch; i++ {
  2129  			// Find the child of b in the dominator tree which
  2130  			// dominates all restores.
  2131  			p := b
  2132  			b = nil
  2133  			for c := s.sdom.Child(p); c != nil && i < maxSpillSearch; c, i = s.sdom.Sibling(c), i+1 {
  2134  				if s.sdom[c.ID].entry <= vi.restoreMin && s.sdom[c.ID].exit >= vi.restoreMax {
  2135  					// c also dominates all restores.  Walk down into c.
  2136  					b = c
  2137  					break
  2138  				}
  2139  			}
  2140  			if b == nil {
  2141  				// Ran out of blocks which dominate all restores.
  2142  				break
  2143  			}
  2144  
  2145  			var depth int16
  2146  			if l := s.loopnest.b2l[b.ID]; l != nil {
  2147  				depth = l.depth
  2148  			}
  2149  			if depth > bestDepth {
  2150  				// Don't push the spill into a deeper loop.
  2151  				continue
  2152  			}
  2153  
  2154  			// If v is in a register at the start of b, we can
  2155  			// place the spill here (after the phis).
  2156  			if len(b.Preds) == 1 {
  2157  				for _, e := range s.endRegs[b.Preds[0].b.ID] {
  2158  					if e.v == v {
  2159  						// Found a better spot for the spill.
  2160  						best = b
  2161  						bestArg = e.c
  2162  						bestDepth = depth
  2163  						break
  2164  					}
  2165  				}
  2166  			} else {
  2167  				for _, e := range s.startRegs[b.ID] {
  2168  					if e.v == v {
  2169  						// Found a better spot for the spill.
  2170  						best = b
  2171  						bestArg = e.c
  2172  						bestDepth = depth
  2173  						break
  2174  					}
  2175  				}
  2176  			}
  2177  		}
  2178  
  2179  		// Put the spill in the best block we found.
  2180  		spill.Block = best
  2181  		spill.AddArg(bestArg)
  2182  		if best == v.Block && !mustBeFirst(v.Op) {
  2183  			// Place immediately after v.
  2184  			after[v.ID] = append(after[v.ID], spill)
  2185  		} else {
  2186  			// Place at the start of best block.
  2187  			start[best.ID] = append(start[best.ID], spill)
  2188  		}
  2189  	}
  2190  
  2191  	// Insert spill instructions into the block schedules.
  2192  	var oldSched []*Value
  2193  	for _, b := range s.visitOrder {
  2194  		nfirst := 0
  2195  		for _, v := range b.Values {
  2196  			if !mustBeFirst(v.Op) {
  2197  				break
  2198  			}
  2199  			nfirst++
  2200  		}
  2201  		oldSched = append(oldSched[:0], b.Values[nfirst:]...)
  2202  		b.Values = b.Values[:nfirst]
  2203  		b.Values = append(b.Values, start[b.ID]...)
  2204  		for _, v := range oldSched {
  2205  			b.Values = append(b.Values, v)
  2206  			b.Values = append(b.Values, after[v.ID]...)
  2207  		}
  2208  	}
  2209  }
  2210  
  2211  // shuffle fixes up all the merge edges (those going into blocks of indegree > 1).
  2212  func (s *regAllocState) shuffle(stacklive [][]ID) {
  2213  	var e edgeState
  2214  	e.s = s
  2215  	e.cache = map[ID][]*Value{}
  2216  	e.contents = map[Location]contentRecord{}
  2217  	if s.f.pass.debug > regDebug {
  2218  		fmt.Printf("shuffle %s\n", s.f.Name)
  2219  		fmt.Println(s.f.String())
  2220  	}
  2221  
  2222  	for _, b := range s.visitOrder {
  2223  		if len(b.Preds) <= 1 {
  2224  			continue
  2225  		}
  2226  		e.b = b
  2227  		for i, edge := range b.Preds {
  2228  			p := edge.b
  2229  			e.p = p
  2230  			e.setup(i, s.endRegs[p.ID], s.startRegs[b.ID], stacklive[p.ID])
  2231  			e.process()
  2232  		}
  2233  	}
  2234  
  2235  	if s.f.pass.debug > regDebug {
  2236  		fmt.Printf("post shuffle %s\n", s.f.Name)
  2237  		fmt.Println(s.f.String())
  2238  	}
  2239  }
  2240  
  2241  type edgeState struct {
  2242  	s    *regAllocState
  2243  	p, b *Block // edge goes from p->b.
  2244  
  2245  	// for each pre-regalloc value, a list of equivalent cached values
  2246  	cache      map[ID][]*Value
  2247  	cachedVals []ID // (superset of) keys of the above map, for deterministic iteration
  2248  
  2249  	// map from location to the value it contains
  2250  	contents map[Location]contentRecord
  2251  
  2252  	// desired destination locations
  2253  	destinations []dstRecord
  2254  	extra        []dstRecord
  2255  
  2256  	usedRegs              regMask // registers currently holding something
  2257  	uniqueRegs            regMask // registers holding the only copy of a value
  2258  	finalRegs             regMask // registers holding final target
  2259  	rematerializeableRegs regMask // registers that hold rematerializeable values
  2260  }
  2261  
  2262  type contentRecord struct {
  2263  	vid   ID       // pre-regalloc value
  2264  	c     *Value   // cached value
  2265  	final bool     // this is a satisfied destination
  2266  	pos   src.XPos // source position of use of the value
  2267  }
  2268  
  2269  type dstRecord struct {
  2270  	loc    Location // register or stack slot
  2271  	vid    ID       // pre-regalloc value it should contain
  2272  	splice **Value  // place to store reference to the generating instruction
  2273  	pos    src.XPos // source position of use of this location
  2274  }
  2275  
  2276  // setup initializes the edge state for shuffling.
  2277  func (e *edgeState) setup(idx int, srcReg []endReg, dstReg []startReg, stacklive []ID) {
  2278  	if e.s.f.pass.debug > regDebug {
  2279  		fmt.Printf("edge %s->%s\n", e.p, e.b)
  2280  	}
  2281  
  2282  	// Clear state.
  2283  	clear(e.cache)
  2284  	e.cachedVals = e.cachedVals[:0]
  2285  	clear(e.contents)
  2286  	e.usedRegs = 0
  2287  	e.uniqueRegs = 0
  2288  	e.finalRegs = 0
  2289  	e.rematerializeableRegs = 0
  2290  
  2291  	// Live registers can be sources.
  2292  	for _, x := range srcReg {
  2293  		e.set(&e.s.registers[x.r], x.v.ID, x.c, false, src.NoXPos) // don't care the position of the source
  2294  	}
  2295  	// So can all of the spill locations.
  2296  	for _, spillID := range stacklive {
  2297  		v := e.s.orig[spillID]
  2298  		spill := e.s.values[v.ID].spill
  2299  		if !e.s.sdom.IsAncestorEq(spill.Block, e.p) {
  2300  			// Spills were placed that only dominate the uses found
  2301  			// during the first regalloc pass. The edge fixup code
  2302  			// can't use a spill location if the spill doesn't dominate
  2303  			// the edge.
  2304  			// We are guaranteed that if the spill doesn't dominate this edge,
  2305  			// then the value is available in a register (because we called
  2306  			// makeSpill for every value not in a register at the start
  2307  			// of an edge).
  2308  			continue
  2309  		}
  2310  		e.set(e.s.f.getHome(spillID), v.ID, spill, false, src.NoXPos) // don't care the position of the source
  2311  	}
  2312  
  2313  	// Figure out all the destinations we need.
  2314  	dsts := e.destinations[:0]
  2315  	for _, x := range dstReg {
  2316  		dsts = append(dsts, dstRecord{&e.s.registers[x.r], x.v.ID, nil, x.pos})
  2317  	}
  2318  	// Phis need their args to end up in a specific location.
  2319  	for _, v := range e.b.Values {
  2320  		if v.Op != OpPhi {
  2321  			break
  2322  		}
  2323  		loc := e.s.f.getHome(v.ID)
  2324  		if loc == nil {
  2325  			continue
  2326  		}
  2327  		dsts = append(dsts, dstRecord{loc, v.Args[idx].ID, &v.Args[idx], v.Pos})
  2328  	}
  2329  	e.destinations = dsts
  2330  
  2331  	if e.s.f.pass.debug > regDebug {
  2332  		for _, vid := range e.cachedVals {
  2333  			a := e.cache[vid]
  2334  			for _, c := range a {
  2335  				fmt.Printf("src %s: v%d cache=%s\n", e.s.f.getHome(c.ID), vid, c)
  2336  			}
  2337  		}
  2338  		for _, d := range e.destinations {
  2339  			fmt.Printf("dst %s: v%d\n", d.loc, d.vid)
  2340  		}
  2341  	}
  2342  }
  2343  
  2344  // process generates code to move all the values to the right destination locations.
  2345  func (e *edgeState) process() {
  2346  	dsts := e.destinations
  2347  
  2348  	// Process the destinations until they are all satisfied.
  2349  	for len(dsts) > 0 {
  2350  		i := 0
  2351  		for _, d := range dsts {
  2352  			if !e.processDest(d.loc, d.vid, d.splice, d.pos) {
  2353  				// Failed - save for next iteration.
  2354  				dsts[i] = d
  2355  				i++
  2356  			}
  2357  		}
  2358  		if i < len(dsts) {
  2359  			// Made some progress. Go around again.
  2360  			dsts = dsts[:i]
  2361  
  2362  			// Append any extras destinations we generated.
  2363  			dsts = append(dsts, e.extra...)
  2364  			e.extra = e.extra[:0]
  2365  			continue
  2366  		}
  2367  
  2368  		// We made no progress. That means that any
  2369  		// remaining unsatisfied moves are in simple cycles.
  2370  		// For example, A -> B -> C -> D -> A.
  2371  		//   A ----> B
  2372  		//   ^       |
  2373  		//   |       |
  2374  		//   |       v
  2375  		//   D <---- C
  2376  
  2377  		// To break the cycle, we pick an unused register, say R,
  2378  		// and put a copy of B there.
  2379  		//   A ----> B
  2380  		//   ^       |
  2381  		//   |       |
  2382  		//   |       v
  2383  		//   D <---- C <---- R=copyofB
  2384  		// When we resume the outer loop, the A->B move can now proceed,
  2385  		// and eventually the whole cycle completes.
  2386  
  2387  		// Copy any cycle location to a temp register. This duplicates
  2388  		// one of the cycle entries, allowing the just duplicated value
  2389  		// to be overwritten and the cycle to proceed.
  2390  		d := dsts[0]
  2391  		loc := d.loc
  2392  		vid := e.contents[loc].vid
  2393  		c := e.contents[loc].c
  2394  		r := e.findRegFor(c.Type)
  2395  		if e.s.f.pass.debug > regDebug {
  2396  			fmt.Printf("breaking cycle with v%d in %s:%s\n", vid, loc, c)
  2397  		}
  2398  		e.erase(r)
  2399  		pos := d.pos.WithNotStmt()
  2400  		if _, isReg := loc.(*Register); isReg {
  2401  			c = e.p.NewValue1(pos, OpCopy, c.Type, c)
  2402  		} else {
  2403  			c = e.p.NewValue1(pos, OpLoadReg, c.Type, c)
  2404  		}
  2405  		e.set(r, vid, c, false, pos)
  2406  		if c.Op == OpLoadReg && e.s.isGReg(register(r.(*Register).num)) {
  2407  			e.s.f.Fatalf("process.OpLoadReg targeting g: " + c.LongString())
  2408  		}
  2409  	}
  2410  }
  2411  
  2412  // processDest generates code to put value vid into location loc. Returns true
  2413  // if progress was made.
  2414  func (e *edgeState) processDest(loc Location, vid ID, splice **Value, pos src.XPos) bool {
  2415  	pos = pos.WithNotStmt()
  2416  	occupant := e.contents[loc]
  2417  	if occupant.vid == vid {
  2418  		// Value is already in the correct place.
  2419  		e.contents[loc] = contentRecord{vid, occupant.c, true, pos}
  2420  		if splice != nil {
  2421  			(*splice).Uses--
  2422  			*splice = occupant.c
  2423  			occupant.c.Uses++
  2424  		}
  2425  		// Note: if splice==nil then c will appear dead. This is
  2426  		// non-SSA formed code, so be careful after this pass not to run
  2427  		// deadcode elimination.
  2428  		if _, ok := e.s.copies[occupant.c]; ok {
  2429  			// The copy at occupant.c was used to avoid spill.
  2430  			e.s.copies[occupant.c] = true
  2431  		}
  2432  		return true
  2433  	}
  2434  
  2435  	// Check if we're allowed to clobber the destination location.
  2436  	if len(e.cache[occupant.vid]) == 1 && !e.s.values[occupant.vid].rematerializeable {
  2437  		// We can't overwrite the last copy
  2438  		// of a value that needs to survive.
  2439  		return false
  2440  	}
  2441  
  2442  	// Copy from a source of v, register preferred.
  2443  	v := e.s.orig[vid]
  2444  	var c *Value
  2445  	var src Location
  2446  	if e.s.f.pass.debug > regDebug {
  2447  		fmt.Printf("moving v%d to %s\n", vid, loc)
  2448  		fmt.Printf("sources of v%d:", vid)
  2449  	}
  2450  	if opcodeTable[v.Op].fixedReg {
  2451  		c = v
  2452  		src = e.s.f.getHome(v.ID)
  2453  	} else {
  2454  		for _, w := range e.cache[vid] {
  2455  			h := e.s.f.getHome(w.ID)
  2456  			if e.s.f.pass.debug > regDebug {
  2457  				fmt.Printf(" %s:%s", h, w)
  2458  			}
  2459  			_, isreg := h.(*Register)
  2460  			if src == nil || isreg {
  2461  				c = w
  2462  				src = h
  2463  			}
  2464  		}
  2465  	}
  2466  	if e.s.f.pass.debug > regDebug {
  2467  		if src != nil {
  2468  			fmt.Printf(" [use %s]\n", src)
  2469  		} else {
  2470  			fmt.Printf(" [no source]\n")
  2471  		}
  2472  	}
  2473  	_, dstReg := loc.(*Register)
  2474  
  2475  	// Pre-clobber destination. This avoids the
  2476  	// following situation:
  2477  	//   - v is currently held in R0 and stacktmp0.
  2478  	//   - We want to copy stacktmp1 to stacktmp0.
  2479  	//   - We choose R0 as the temporary register.
  2480  	// During the copy, both R0 and stacktmp0 are
  2481  	// clobbered, losing both copies of v. Oops!
  2482  	// Erasing the destination early means R0 will not
  2483  	// be chosen as the temp register, as it will then
  2484  	// be the last copy of v.
  2485  	e.erase(loc)
  2486  	var x *Value
  2487  	if c == nil || e.s.values[vid].rematerializeable {
  2488  		if !e.s.values[vid].rematerializeable {
  2489  			e.s.f.Fatalf("can't find source for %s->%s: %s\n", e.p, e.b, v.LongString())
  2490  		}
  2491  		if dstReg {
  2492  			x = v.copyInto(e.p)
  2493  		} else {
  2494  			// Rematerialize into stack slot. Need a free
  2495  			// register to accomplish this.
  2496  			r := e.findRegFor(v.Type)
  2497  			e.erase(r)
  2498  			x = v.copyIntoWithXPos(e.p, pos)
  2499  			e.set(r, vid, x, false, pos)
  2500  			// Make sure we spill with the size of the slot, not the
  2501  			// size of x (which might be wider due to our dropping
  2502  			// of narrowing conversions).
  2503  			x = e.p.NewValue1(pos, OpStoreReg, loc.(LocalSlot).Type, x)
  2504  		}
  2505  	} else {
  2506  		// Emit move from src to dst.
  2507  		_, srcReg := src.(*Register)
  2508  		if srcReg {
  2509  			if dstReg {
  2510  				x = e.p.NewValue1(pos, OpCopy, c.Type, c)
  2511  			} else {
  2512  				x = e.p.NewValue1(pos, OpStoreReg, loc.(LocalSlot).Type, c)
  2513  			}
  2514  		} else {
  2515  			if dstReg {
  2516  				x = e.p.NewValue1(pos, OpLoadReg, c.Type, c)
  2517  			} else {
  2518  				// mem->mem. Use temp register.
  2519  				r := e.findRegFor(c.Type)
  2520  				e.erase(r)
  2521  				t := e.p.NewValue1(pos, OpLoadReg, c.Type, c)
  2522  				e.set(r, vid, t, false, pos)
  2523  				x = e.p.NewValue1(pos, OpStoreReg, loc.(LocalSlot).Type, t)
  2524  			}
  2525  		}
  2526  	}
  2527  	e.set(loc, vid, x, true, pos)
  2528  	if x.Op == OpLoadReg && e.s.isGReg(register(loc.(*Register).num)) {
  2529  		e.s.f.Fatalf("processDest.OpLoadReg targeting g: " + x.LongString())
  2530  	}
  2531  	if splice != nil {
  2532  		(*splice).Uses--
  2533  		*splice = x
  2534  		x.Uses++
  2535  	}
  2536  	return true
  2537  }
  2538  
  2539  // set changes the contents of location loc to hold the given value and its cached representative.
  2540  func (e *edgeState) set(loc Location, vid ID, c *Value, final bool, pos src.XPos) {
  2541  	e.s.f.setHome(c, loc)
  2542  	e.contents[loc] = contentRecord{vid, c, final, pos}
  2543  	a := e.cache[vid]
  2544  	if len(a) == 0 {
  2545  		e.cachedVals = append(e.cachedVals, vid)
  2546  	}
  2547  	a = append(a, c)
  2548  	e.cache[vid] = a
  2549  	if r, ok := loc.(*Register); ok {
  2550  		if e.usedRegs&(regMask(1)<<uint(r.num)) != 0 {
  2551  			e.s.f.Fatalf("%v is already set (v%d/%v)", r, vid, c)
  2552  		}
  2553  		e.usedRegs |= regMask(1) << uint(r.num)
  2554  		if final {
  2555  			e.finalRegs |= regMask(1) << uint(r.num)
  2556  		}
  2557  		if len(a) == 1 {
  2558  			e.uniqueRegs |= regMask(1) << uint(r.num)
  2559  		}
  2560  		if len(a) == 2 {
  2561  			if t, ok := e.s.f.getHome(a[0].ID).(*Register); ok {
  2562  				e.uniqueRegs &^= regMask(1) << uint(t.num)
  2563  			}
  2564  		}
  2565  		if e.s.values[vid].rematerializeable {
  2566  			e.rematerializeableRegs |= regMask(1) << uint(r.num)
  2567  		}
  2568  	}
  2569  	if e.s.f.pass.debug > regDebug {
  2570  		fmt.Printf("%s\n", c.LongString())
  2571  		fmt.Printf("v%d now available in %s:%s\n", vid, loc, c)
  2572  	}
  2573  }
  2574  
  2575  // erase removes any user of loc.
  2576  func (e *edgeState) erase(loc Location) {
  2577  	cr := e.contents[loc]
  2578  	if cr.c == nil {
  2579  		return
  2580  	}
  2581  	vid := cr.vid
  2582  
  2583  	if cr.final {
  2584  		// Add a destination to move this value back into place.
  2585  		// Make sure it gets added to the tail of the destination queue
  2586  		// so we make progress on other moves first.
  2587  		e.extra = append(e.extra, dstRecord{loc, cr.vid, nil, cr.pos})
  2588  	}
  2589  
  2590  	// Remove c from the list of cached values.
  2591  	a := e.cache[vid]
  2592  	for i, c := range a {
  2593  		if e.s.f.getHome(c.ID) == loc {
  2594  			if e.s.f.pass.debug > regDebug {
  2595  				fmt.Printf("v%d no longer available in %s:%s\n", vid, loc, c)
  2596  			}
  2597  			a[i], a = a[len(a)-1], a[:len(a)-1]
  2598  			break
  2599  		}
  2600  	}
  2601  	e.cache[vid] = a
  2602  
  2603  	// Update register masks.
  2604  	if r, ok := loc.(*Register); ok {
  2605  		e.usedRegs &^= regMask(1) << uint(r.num)
  2606  		if cr.final {
  2607  			e.finalRegs &^= regMask(1) << uint(r.num)
  2608  		}
  2609  		e.rematerializeableRegs &^= regMask(1) << uint(r.num)
  2610  	}
  2611  	if len(a) == 1 {
  2612  		if r, ok := e.s.f.getHome(a[0].ID).(*Register); ok {
  2613  			e.uniqueRegs |= regMask(1) << uint(r.num)
  2614  		}
  2615  	}
  2616  }
  2617  
  2618  // findRegFor finds a register we can use to make a temp copy of type typ.
  2619  func (e *edgeState) findRegFor(typ *types.Type) Location {
  2620  	// Which registers are possibilities.
  2621  	types := &e.s.f.Config.Types
  2622  	m := e.s.compatRegs(typ)
  2623  
  2624  	// Pick a register. In priority order:
  2625  	// 1) an unused register
  2626  	// 2) a non-unique register not holding a final value
  2627  	// 3) a non-unique register
  2628  	// 4) a register holding a rematerializeable value
  2629  	x := m &^ e.usedRegs
  2630  	if x != 0 {
  2631  		return &e.s.registers[pickReg(x)]
  2632  	}
  2633  	x = m &^ e.uniqueRegs &^ e.finalRegs
  2634  	if x != 0 {
  2635  		return &e.s.registers[pickReg(x)]
  2636  	}
  2637  	x = m &^ e.uniqueRegs
  2638  	if x != 0 {
  2639  		return &e.s.registers[pickReg(x)]
  2640  	}
  2641  	x = m & e.rematerializeableRegs
  2642  	if x != 0 {
  2643  		return &e.s.registers[pickReg(x)]
  2644  	}
  2645  
  2646  	// No register is available.
  2647  	// Pick a register to spill.
  2648  	for _, vid := range e.cachedVals {
  2649  		a := e.cache[vid]
  2650  		for _, c := range a {
  2651  			if r, ok := e.s.f.getHome(c.ID).(*Register); ok && m>>uint(r.num)&1 != 0 {
  2652  				if !c.rematerializeable() {
  2653  					x := e.p.NewValue1(c.Pos, OpStoreReg, c.Type, c)
  2654  					// Allocate a temp location to spill a register to.
  2655  					// The type of the slot is immaterial - it will not be live across
  2656  					// any safepoint. Just use a type big enough to hold any register.
  2657  					t := LocalSlot{N: e.s.f.NewLocal(c.Pos, types.Int64), Type: types.Int64}
  2658  					// TODO: reuse these slots. They'll need to be erased first.
  2659  					e.set(t, vid, x, false, c.Pos)
  2660  					if e.s.f.pass.debug > regDebug {
  2661  						fmt.Printf("  SPILL %s->%s %s\n", r, t, x.LongString())
  2662  					}
  2663  				}
  2664  				// r will now be overwritten by the caller. At some point
  2665  				// later, the newly saved value will be moved back to its
  2666  				// final destination in processDest.
  2667  				return r
  2668  			}
  2669  		}
  2670  	}
  2671  
  2672  	fmt.Printf("m:%d unique:%d final:%d rematerializable:%d\n", m, e.uniqueRegs, e.finalRegs, e.rematerializeableRegs)
  2673  	for _, vid := range e.cachedVals {
  2674  		a := e.cache[vid]
  2675  		for _, c := range a {
  2676  			fmt.Printf("v%d: %s %s\n", vid, c, e.s.f.getHome(c.ID))
  2677  		}
  2678  	}
  2679  	e.s.f.Fatalf("can't find empty register on edge %s->%s", e.p, e.b)
  2680  	return nil
  2681  }
  2682  
  2683  // rematerializeable reports whether the register allocator should recompute
  2684  // a value instead of spilling/restoring it.
  2685  func (v *Value) rematerializeable() bool {
  2686  	if !opcodeTable[v.Op].rematerializeable {
  2687  		return false
  2688  	}
  2689  	for _, a := range v.Args {
  2690  		// Fixed-register allocations (SP, SB, etc.) are always available.
  2691  		// Any other argument of an opcode makes it not rematerializeable.
  2692  		if !opcodeTable[a.Op].fixedReg {
  2693  			return false
  2694  		}
  2695  	}
  2696  	return true
  2697  }
  2698  
  2699  type liveInfo struct {
  2700  	ID   ID       // ID of value
  2701  	dist int32    // # of instructions before next use
  2702  	pos  src.XPos // source position of next use
  2703  }
  2704  
  2705  // computeLive computes a map from block ID to a list of value IDs live at the end
  2706  // of that block. Together with the value ID is a count of how many instructions
  2707  // to the next use of that value. The resulting map is stored in s.live.
  2708  // computeLive also computes the desired register information at the end of each block.
  2709  // This desired register information is stored in s.desired.
  2710  // TODO: this could be quadratic if lots of variables are live across lots of
  2711  // basic blocks. Figure out a way to make this function (or, more precisely, the user
  2712  // of this function) require only linear size & time.
  2713  func (s *regAllocState) computeLive() {
  2714  	f := s.f
  2715  	s.live = make([][]liveInfo, f.NumBlocks())
  2716  	s.desired = make([]desiredState, f.NumBlocks())
  2717  	var phis []*Value
  2718  
  2719  	live := f.newSparseMapPos(f.NumValues())
  2720  	defer f.retSparseMapPos(live)
  2721  	t := f.newSparseMapPos(f.NumValues())
  2722  	defer f.retSparseMapPos(t)
  2723  
  2724  	// Keep track of which value we want in each register.
  2725  	var desired desiredState
  2726  
  2727  	// Instead of iterating over f.Blocks, iterate over their postordering.
  2728  	// Liveness information flows backward, so starting at the end
  2729  	// increases the probability that we will stabilize quickly.
  2730  	// TODO: Do a better job yet. Here's one possibility:
  2731  	// Calculate the dominator tree and locate all strongly connected components.
  2732  	// If a value is live in one block of an SCC, it is live in all.
  2733  	// Walk the dominator tree from end to beginning, just once, treating SCC
  2734  	// components as single blocks, duplicated calculated liveness information
  2735  	// out to all of them.
  2736  	po := f.postorder()
  2737  	s.loopnest = f.loopnest()
  2738  	s.loopnest.calculateDepths()
  2739  	for {
  2740  		changed := false
  2741  
  2742  		for _, b := range po {
  2743  			// Start with known live values at the end of the block.
  2744  			// Add len(b.Values) to adjust from end-of-block distance
  2745  			// to beginning-of-block distance.
  2746  			live.clear()
  2747  			for _, e := range s.live[b.ID] {
  2748  				live.set(e.ID, e.dist+int32(len(b.Values)), e.pos)
  2749  			}
  2750  
  2751  			// Mark control values as live
  2752  			for _, c := range b.ControlValues() {
  2753  				if s.values[c.ID].needReg {
  2754  					live.set(c.ID, int32(len(b.Values)), b.Pos)
  2755  				}
  2756  			}
  2757  
  2758  			// Propagate backwards to the start of the block
  2759  			// Assumes Values have been scheduled.
  2760  			phis = phis[:0]
  2761  			for i := len(b.Values) - 1; i >= 0; i-- {
  2762  				v := b.Values[i]
  2763  				live.remove(v.ID)
  2764  				if v.Op == OpPhi {
  2765  					// save phi ops for later
  2766  					phis = append(phis, v)
  2767  					continue
  2768  				}
  2769  				if opcodeTable[v.Op].call {
  2770  					c := live.contents()
  2771  					for i := range c {
  2772  						c[i].val += unlikelyDistance
  2773  					}
  2774  				}
  2775  				for _, a := range v.Args {
  2776  					if s.values[a.ID].needReg {
  2777  						live.set(a.ID, int32(i), v.Pos)
  2778  					}
  2779  				}
  2780  			}
  2781  			// Propagate desired registers backwards.
  2782  			desired.copy(&s.desired[b.ID])
  2783  			for i := len(b.Values) - 1; i >= 0; i-- {
  2784  				v := b.Values[i]
  2785  				prefs := desired.remove(v.ID)
  2786  				if v.Op == OpPhi {
  2787  					// TODO: if v is a phi, save desired register for phi inputs.
  2788  					// For now, we just drop it and don't propagate
  2789  					// desired registers back though phi nodes.
  2790  					continue
  2791  				}
  2792  				regspec := s.regspec(v)
  2793  				// Cancel desired registers if they get clobbered.
  2794  				desired.clobber(regspec.clobbers)
  2795  				// Update desired registers if there are any fixed register inputs.
  2796  				for _, j := range regspec.inputs {
  2797  					if countRegs(j.regs) != 1 {
  2798  						continue
  2799  					}
  2800  					desired.clobber(j.regs)
  2801  					desired.add(v.Args[j.idx].ID, pickReg(j.regs))
  2802  				}
  2803  				// Set desired register of input 0 if this is a 2-operand instruction.
  2804  				if opcodeTable[v.Op].resultInArg0 || v.Op == OpAMD64ADDQconst || v.Op == OpAMD64ADDLconst || v.Op == OpSelect0 {
  2805  					// ADDQconst is added here because we want to treat it as resultInArg0 for
  2806  					// the purposes of desired registers, even though it is not an absolute requirement.
  2807  					// This is because we'd rather implement it as ADDQ instead of LEAQ.
  2808  					// Same for ADDLconst
  2809  					// Select0 is added here to propagate the desired register to the tuple-generating instruction.
  2810  					if opcodeTable[v.Op].commutative {
  2811  						desired.addList(v.Args[1].ID, prefs)
  2812  					}
  2813  					desired.addList(v.Args[0].ID, prefs)
  2814  				}
  2815  			}
  2816  
  2817  			// For each predecessor of b, expand its list of live-at-end values.
  2818  			// invariant: live contains the values live at the start of b (excluding phi inputs)
  2819  			for i, e := range b.Preds {
  2820  				p := e.b
  2821  				// Compute additional distance for the edge.
  2822  				// Note: delta must be at least 1 to distinguish the control
  2823  				// value use from the first user in a successor block.
  2824  				delta := int32(normalDistance)
  2825  				if len(p.Succs) == 2 {
  2826  					if p.Succs[0].b == b && p.Likely == BranchLikely ||
  2827  						p.Succs[1].b == b && p.Likely == BranchUnlikely {
  2828  						delta = likelyDistance
  2829  					}
  2830  					if p.Succs[0].b == b && p.Likely == BranchUnlikely ||
  2831  						p.Succs[1].b == b && p.Likely == BranchLikely {
  2832  						delta = unlikelyDistance
  2833  					}
  2834  				}
  2835  
  2836  				// Update any desired registers at the end of p.
  2837  				s.desired[p.ID].merge(&desired)
  2838  
  2839  				// Start t off with the previously known live values at the end of p.
  2840  				t.clear()
  2841  				for _, e := range s.live[p.ID] {
  2842  					t.set(e.ID, e.dist, e.pos)
  2843  				}
  2844  				update := false
  2845  
  2846  				// Add new live values from scanning this block.
  2847  				for _, e := range live.contents() {
  2848  					d := e.val + delta
  2849  					if !t.contains(e.key) || d < t.get(e.key) {
  2850  						update = true
  2851  						t.set(e.key, d, e.pos)
  2852  					}
  2853  				}
  2854  				// Also add the correct arg from the saved phi values.
  2855  				// All phis are at distance delta (we consider them
  2856  				// simultaneously happening at the start of the block).
  2857  				for _, v := range phis {
  2858  					id := v.Args[i].ID
  2859  					if s.values[id].needReg && (!t.contains(id) || delta < t.get(id)) {
  2860  						update = true
  2861  						t.set(id, delta, v.Pos)
  2862  					}
  2863  				}
  2864  
  2865  				if !update {
  2866  					continue
  2867  				}
  2868  				// The live set has changed, update it.
  2869  				l := s.live[p.ID][:0]
  2870  				if cap(l) < t.size() {
  2871  					l = make([]liveInfo, 0, t.size())
  2872  				}
  2873  				for _, e := range t.contents() {
  2874  					l = append(l, liveInfo{e.key, e.val, e.pos})
  2875  				}
  2876  				s.live[p.ID] = l
  2877  				changed = true
  2878  			}
  2879  		}
  2880  
  2881  		if !changed {
  2882  			break
  2883  		}
  2884  	}
  2885  	if f.pass.debug > regDebug {
  2886  		fmt.Println("live values at end of each block")
  2887  		for _, b := range f.Blocks {
  2888  			fmt.Printf("  %s:", b)
  2889  			for _, x := range s.live[b.ID] {
  2890  				fmt.Printf(" v%d(%d)", x.ID, x.dist)
  2891  				for _, e := range s.desired[b.ID].entries {
  2892  					if e.ID != x.ID {
  2893  						continue
  2894  					}
  2895  					fmt.Printf("[")
  2896  					first := true
  2897  					for _, r := range e.regs {
  2898  						if r == noRegister {
  2899  							continue
  2900  						}
  2901  						if !first {
  2902  							fmt.Printf(",")
  2903  						}
  2904  						fmt.Print(&s.registers[r])
  2905  						first = false
  2906  					}
  2907  					fmt.Printf("]")
  2908  				}
  2909  			}
  2910  			if avoid := s.desired[b.ID].avoid; avoid != 0 {
  2911  				fmt.Printf(" avoid=%v", s.RegMaskString(avoid))
  2912  			}
  2913  			fmt.Println()
  2914  		}
  2915  	}
  2916  }
  2917  
  2918  // A desiredState represents desired register assignments.
  2919  type desiredState struct {
  2920  	// Desired assignments will be small, so we just use a list
  2921  	// of valueID+registers entries.
  2922  	entries []desiredStateEntry
  2923  	// Registers that other values want to be in.  This value will
  2924  	// contain at least the union of the regs fields of entries, but
  2925  	// may contain additional entries for values that were once in
  2926  	// this data structure but are no longer.
  2927  	avoid regMask
  2928  }
  2929  type desiredStateEntry struct {
  2930  	// (pre-regalloc) value
  2931  	ID ID
  2932  	// Registers it would like to be in, in priority order.
  2933  	// Unused slots are filled with noRegister.
  2934  	// For opcodes that return tuples, we track desired registers only
  2935  	// for the first element of the tuple (see desiredSecondReg for
  2936  	// tracking the desired register for second part of a tuple).
  2937  	regs [4]register
  2938  }
  2939  
  2940  func (d *desiredState) clear() {
  2941  	d.entries = d.entries[:0]
  2942  	d.avoid = 0
  2943  }
  2944  
  2945  // get returns a list of desired registers for value vid.
  2946  func (d *desiredState) get(vid ID) [4]register {
  2947  	for _, e := range d.entries {
  2948  		if e.ID == vid {
  2949  			return e.regs
  2950  		}
  2951  	}
  2952  	return [4]register{noRegister, noRegister, noRegister, noRegister}
  2953  }
  2954  
  2955  // add records that we'd like value vid to be in register r.
  2956  func (d *desiredState) add(vid ID, r register) {
  2957  	d.avoid |= regMask(1) << r
  2958  	for i := range d.entries {
  2959  		e := &d.entries[i]
  2960  		if e.ID != vid {
  2961  			continue
  2962  		}
  2963  		if e.regs[0] == r {
  2964  			// Already known and highest priority
  2965  			return
  2966  		}
  2967  		for j := 1; j < len(e.regs); j++ {
  2968  			if e.regs[j] == r {
  2969  				// Move from lower priority to top priority
  2970  				copy(e.regs[1:], e.regs[:j])
  2971  				e.regs[0] = r
  2972  				return
  2973  			}
  2974  		}
  2975  		copy(e.regs[1:], e.regs[:])
  2976  		e.regs[0] = r
  2977  		return
  2978  	}
  2979  	d.entries = append(d.entries, desiredStateEntry{vid, [4]register{r, noRegister, noRegister, noRegister}})
  2980  }
  2981  
  2982  func (d *desiredState) addList(vid ID, regs [4]register) {
  2983  	// regs is in priority order, so iterate in reverse order.
  2984  	for i := len(regs) - 1; i >= 0; i-- {
  2985  		r := regs[i]
  2986  		if r != noRegister {
  2987  			d.add(vid, r)
  2988  		}
  2989  	}
  2990  }
  2991  
  2992  // clobber erases any desired registers in the set m.
  2993  func (d *desiredState) clobber(m regMask) {
  2994  	for i := 0; i < len(d.entries); {
  2995  		e := &d.entries[i]
  2996  		j := 0
  2997  		for _, r := range e.regs {
  2998  			if r != noRegister && m>>r&1 == 0 {
  2999  				e.regs[j] = r
  3000  				j++
  3001  			}
  3002  		}
  3003  		if j == 0 {
  3004  			// No more desired registers for this value.
  3005  			d.entries[i] = d.entries[len(d.entries)-1]
  3006  			d.entries = d.entries[:len(d.entries)-1]
  3007  			continue
  3008  		}
  3009  		for ; j < len(e.regs); j++ {
  3010  			e.regs[j] = noRegister
  3011  		}
  3012  		i++
  3013  	}
  3014  	d.avoid &^= m
  3015  }
  3016  
  3017  // copy copies a desired state from another desiredState x.
  3018  func (d *desiredState) copy(x *desiredState) {
  3019  	d.entries = append(d.entries[:0], x.entries...)
  3020  	d.avoid = x.avoid
  3021  }
  3022  
  3023  // remove removes the desired registers for vid and returns them.
  3024  func (d *desiredState) remove(vid ID) [4]register {
  3025  	for i := range d.entries {
  3026  		if d.entries[i].ID == vid {
  3027  			regs := d.entries[i].regs
  3028  			d.entries[i] = d.entries[len(d.entries)-1]
  3029  			d.entries = d.entries[:len(d.entries)-1]
  3030  			return regs
  3031  		}
  3032  	}
  3033  	return [4]register{noRegister, noRegister, noRegister, noRegister}
  3034  }
  3035  
  3036  // merge merges another desired state x into d.
  3037  func (d *desiredState) merge(x *desiredState) {
  3038  	d.avoid |= x.avoid
  3039  	// There should only be a few desired registers, so
  3040  	// linear insert is ok.
  3041  	for _, e := range x.entries {
  3042  		d.addList(e.ID, e.regs)
  3043  	}
  3044  }
  3045  

View as plain text