Source file src/cmd/internal/obj/mips/obj0.go

     1  // cmd/9l/noop.c, cmd/9l/pass.c, cmd/9l/span.c from Vita Nuova.
     2  //
     3  //	Copyright © 1994-1999 Lucent Technologies Inc.  All rights reserved.
     4  //	Portions Copyright © 1995-1997 C H Forsyth (forsyth@terzarima.net)
     5  //	Portions Copyright © 1997-1999 Vita Nuova Limited
     6  //	Portions Copyright © 2000-2008 Vita Nuova Holdings Limited (www.vitanuova.com)
     7  //	Portions Copyright © 2004,2006 Bruce Ellis
     8  //	Portions Copyright © 2005-2007 C H Forsyth (forsyth@terzarima.net)
     9  //	Revisions Copyright © 2000-2008 Lucent Technologies Inc. and others
    10  //	Portions Copyright © 2009 The Go Authors. All rights reserved.
    11  //
    12  // Permission is hereby granted, free of charge, to any person obtaining a copy
    13  // of this software and associated documentation files (the "Software"), to deal
    14  // in the Software without restriction, including without limitation the rights
    15  // to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
    16  // copies of the Software, and to permit persons to whom the Software is
    17  // furnished to do so, subject to the following conditions:
    18  //
    19  // The above copyright notice and this permission notice shall be included in
    20  // all copies or substantial portions of the Software.
    21  //
    22  // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
    23  // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
    24  // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
    25  // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
    26  // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
    27  // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
    28  // THE SOFTWARE.
    29  
    30  package mips
    31  
    32  import (
    33  	"cmd/internal/obj"
    34  	"cmd/internal/sys"
    35  	"encoding/binary"
    36  	"fmt"
    37  	"internal/abi"
    38  	"log"
    39  	"math"
    40  )
    41  
    42  func progedit(ctxt *obj.Link, p *obj.Prog, newprog obj.ProgAlloc) {
    43  	c := ctxt0{ctxt: ctxt}
    44  
    45  	p.From.Class = 0
    46  	p.To.Class = 0
    47  
    48  	// Rewrite JMP/JAL to symbol as TYPE_BRANCH.
    49  	switch p.As {
    50  	case AJMP,
    51  		AJAL,
    52  		ARET,
    53  		obj.ADUFFZERO,
    54  		obj.ADUFFCOPY:
    55  		if p.To.Sym != nil {
    56  			p.To.Type = obj.TYPE_BRANCH
    57  		}
    58  	}
    59  
    60  	// Rewrite float constants to values stored in memory.
    61  	switch p.As {
    62  	case AMOVF:
    63  		if p.From.Type == obj.TYPE_FCONST {
    64  			f32 := float32(p.From.Val.(float64))
    65  			if math.Float32bits(f32) == 0 {
    66  				p.As = AMOVW
    67  				p.From.Type = obj.TYPE_REG
    68  				p.From.Reg = REGZERO
    69  				break
    70  			}
    71  			p.From.Type = obj.TYPE_MEM
    72  			p.From.Sym = ctxt.Float32Sym(f32)
    73  			p.From.Name = obj.NAME_EXTERN
    74  			p.From.Offset = 0
    75  		}
    76  
    77  	case AMOVD:
    78  		if p.From.Type == obj.TYPE_FCONST {
    79  			f64 := p.From.Val.(float64)
    80  			if math.Float64bits(f64) == 0 && c.ctxt.Arch.Family == sys.MIPS64 {
    81  				p.As = AMOVV
    82  				p.From.Type = obj.TYPE_REG
    83  				p.From.Reg = REGZERO
    84  				break
    85  			}
    86  			p.From.Type = obj.TYPE_MEM
    87  			p.From.Sym = ctxt.Float64Sym(f64)
    88  			p.From.Name = obj.NAME_EXTERN
    89  			p.From.Offset = 0
    90  		}
    91  
    92  		// Put >32-bit constants in memory and load them
    93  	case AMOVV:
    94  		if p.From.Type == obj.TYPE_CONST && p.From.Name == obj.NAME_NONE && p.From.Reg == 0 && int64(int32(p.From.Offset)) != p.From.Offset {
    95  			p.From.Type = obj.TYPE_MEM
    96  			p.From.Sym = ctxt.Int64Sym(p.From.Offset)
    97  			p.From.Name = obj.NAME_EXTERN
    98  			p.From.Offset = 0
    99  		}
   100  	}
   101  
   102  	// Rewrite SUB constants into ADD.
   103  	switch p.As {
   104  	case ASUB:
   105  		if p.From.Type == obj.TYPE_CONST {
   106  			p.From.Offset = -p.From.Offset
   107  			p.As = AADD
   108  		}
   109  
   110  	case ASUBU:
   111  		if p.From.Type == obj.TYPE_CONST {
   112  			p.From.Offset = -p.From.Offset
   113  			p.As = AADDU
   114  		}
   115  
   116  	case ASUBV:
   117  		if p.From.Type == obj.TYPE_CONST {
   118  			p.From.Offset = -p.From.Offset
   119  			p.As = AADDV
   120  		}
   121  
   122  	case ASUBVU:
   123  		if p.From.Type == obj.TYPE_CONST {
   124  			p.From.Offset = -p.From.Offset
   125  			p.As = AADDVU
   126  		}
   127  	}
   128  }
   129  
   130  func preprocess(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
   131  	// TODO(minux): add morestack short-cuts with small fixed frame-size.
   132  	c := ctxt0{ctxt: ctxt, newprog: newprog, cursym: cursym}
   133  
   134  	// a switch for enabling/disabling instruction scheduling
   135  	nosched := true
   136  
   137  	if c.cursym.Func().Text == nil || c.cursym.Func().Text.Link == nil {
   138  		return
   139  	}
   140  
   141  	p := c.cursym.Func().Text
   142  	textstksiz := p.To.Offset
   143  	if textstksiz == -ctxt.Arch.FixedFrameSize {
   144  		// Historical way to mark NOFRAME.
   145  		p.From.Sym.Set(obj.AttrNoFrame, true)
   146  		textstksiz = 0
   147  	}
   148  	if textstksiz < 0 {
   149  		c.ctxt.Diag("negative frame size %d - did you mean NOFRAME?", textstksiz)
   150  	}
   151  	if p.From.Sym.NoFrame() {
   152  		if textstksiz != 0 {
   153  			c.ctxt.Diag("NOFRAME functions must have a frame size of 0, not %d", textstksiz)
   154  		}
   155  	}
   156  
   157  	c.cursym.Func().Args = p.To.Val.(int32)
   158  	c.cursym.Func().Locals = int32(textstksiz)
   159  
   160  	/*
   161  	 * find leaf subroutines
   162  	 * expand RET
   163  	 * expand BECOME pseudo
   164  	 */
   165  
   166  	for p := c.cursym.Func().Text; p != nil; p = p.Link {
   167  		switch p.As {
   168  		/* too hard, just leave alone */
   169  		case obj.ATEXT:
   170  			p.Mark |= LABEL | LEAF | SYNC
   171  			if p.Link != nil {
   172  				p.Link.Mark |= LABEL
   173  			}
   174  
   175  		/* too hard, just leave alone */
   176  		case AMOVW,
   177  			AMOVV:
   178  			if p.To.Type == obj.TYPE_REG && p.To.Reg >= REG_SPECIAL {
   179  				p.Mark |= LABEL | SYNC
   180  				break
   181  			}
   182  			if p.From.Type == obj.TYPE_REG && p.From.Reg >= REG_SPECIAL {
   183  				p.Mark |= LABEL | SYNC
   184  			}
   185  
   186  		/* too hard, just leave alone */
   187  		case ASYSCALL,
   188  			AWORD,
   189  			ATLBWR,
   190  			ATLBWI,
   191  			ATLBP,
   192  			ATLBR:
   193  			p.Mark |= LABEL | SYNC
   194  
   195  		case ANOR:
   196  			if p.To.Type == obj.TYPE_REG {
   197  				if p.To.Reg == REGZERO {
   198  					p.Mark |= LABEL | SYNC
   199  				}
   200  			}
   201  
   202  		case ABGEZAL,
   203  			ABLTZAL,
   204  			AJAL,
   205  			obj.ADUFFZERO,
   206  			obj.ADUFFCOPY:
   207  			c.cursym.Func().Text.Mark &^= LEAF
   208  			fallthrough
   209  
   210  		case AJMP,
   211  			ABEQ,
   212  			ABGEZ,
   213  			ABGTZ,
   214  			ABLEZ,
   215  			ABLTZ,
   216  			ABNE,
   217  			ABFPT, ABFPF:
   218  			if p.As == ABFPT || p.As == ABFPF {
   219  				// We don't treat ABFPT and ABFPF as branches here,
   220  				// so that we will always fill nop (0x0) in their
   221  				// delay slot during assembly.
   222  				// This is to workaround a kernel FPU emulator bug
   223  				// where it uses the user stack to simulate the
   224  				// instruction in the delay slot if it's not 0x0,
   225  				// and somehow that leads to SIGSEGV when the kernel
   226  				// jump to the stack.
   227  				p.Mark |= SYNC
   228  			} else {
   229  				p.Mark |= BRANCH
   230  			}
   231  			q1 := p.To.Target()
   232  			if q1 != nil {
   233  				for q1.As == obj.ANOP {
   234  					q1 = q1.Link
   235  					p.To.SetTarget(q1)
   236  				}
   237  
   238  				if q1.Mark&LEAF == 0 {
   239  					q1.Mark |= LABEL
   240  				}
   241  			}
   242  			//else {
   243  			//	p.Mark |= LABEL
   244  			//}
   245  			q1 = p.Link
   246  			if q1 != nil {
   247  				q1.Mark |= LABEL
   248  			}
   249  
   250  		case ARET:
   251  			if p.Link != nil {
   252  				p.Link.Mark |= LABEL
   253  			}
   254  		}
   255  	}
   256  
   257  	var mov, add obj.As
   258  	if c.ctxt.Arch.Family == sys.MIPS64 {
   259  		add = AADDV
   260  		mov = AMOVV
   261  	} else {
   262  		add = AADDU
   263  		mov = AMOVW
   264  	}
   265  
   266  	var q *obj.Prog
   267  	var q1 *obj.Prog
   268  	autosize := int32(0)
   269  	var p1 *obj.Prog
   270  	var p2 *obj.Prog
   271  	for p := c.cursym.Func().Text; p != nil; p = p.Link {
   272  		o := p.As
   273  		switch o {
   274  		case obj.ATEXT:
   275  			autosize = int32(textstksiz)
   276  
   277  			if p.Mark&LEAF != 0 && autosize == 0 {
   278  				// A leaf function with no locals has no frame.
   279  				p.From.Sym.Set(obj.AttrNoFrame, true)
   280  			}
   281  
   282  			if !p.From.Sym.NoFrame() {
   283  				// If there is a stack frame at all, it includes
   284  				// space to save the LR.
   285  				autosize += int32(c.ctxt.Arch.FixedFrameSize)
   286  			}
   287  
   288  			if autosize&4 != 0 && c.ctxt.Arch.Family == sys.MIPS64 {
   289  				autosize += 4
   290  			}
   291  
   292  			if autosize == 0 && c.cursym.Func().Text.Mark&LEAF == 0 {
   293  				if c.cursym.Func().Text.From.Sym.NoSplit() {
   294  					if ctxt.Debugvlog {
   295  						ctxt.Logf("save suppressed in: %s\n", c.cursym.Name)
   296  					}
   297  
   298  					c.cursym.Func().Text.Mark |= LEAF
   299  				}
   300  			}
   301  
   302  			p.To.Offset = int64(autosize) - ctxt.Arch.FixedFrameSize
   303  
   304  			if c.cursym.Func().Text.Mark&LEAF != 0 {
   305  				c.cursym.Set(obj.AttrLeaf, true)
   306  				if p.From.Sym.NoFrame() {
   307  					break
   308  				}
   309  			}
   310  
   311  			if !p.From.Sym.NoSplit() {
   312  				p = c.stacksplit(p, autosize) // emit split check
   313  			}
   314  
   315  			q = p
   316  
   317  			if autosize != 0 {
   318  				// Make sure to save link register for non-empty frame, even if
   319  				// it is a leaf function, so that traceback works.
   320  				// Store link register before decrement SP, so if a signal comes
   321  				// during the execution of the function prologue, the traceback
   322  				// code will not see a half-updated stack frame.
   323  				// This sequence is not async preemptible, as if we open a frame
   324  				// at the current SP, it will clobber the saved LR.
   325  				q = c.ctxt.StartUnsafePoint(q, c.newprog)
   326  
   327  				q = obj.Appendp(q, newprog)
   328  				q.As = mov
   329  				q.Pos = p.Pos
   330  				q.From.Type = obj.TYPE_REG
   331  				q.From.Reg = REGLINK
   332  				q.To.Type = obj.TYPE_MEM
   333  				q.To.Offset = int64(-autosize)
   334  				q.To.Reg = REGSP
   335  
   336  				q = obj.Appendp(q, newprog)
   337  				q.As = add
   338  				q.Pos = p.Pos
   339  				q.From.Type = obj.TYPE_CONST
   340  				q.From.Offset = int64(-autosize)
   341  				q.To.Type = obj.TYPE_REG
   342  				q.To.Reg = REGSP
   343  				q.Spadj = +autosize
   344  
   345  				q = c.ctxt.EndUnsafePoint(q, c.newprog, -1)
   346  
   347  				// On Linux, in a cgo binary we may get a SIGSETXID signal early on
   348  				// before the signal stack is set, as glibc doesn't allow us to block
   349  				// SIGSETXID. So a signal may land on the current stack and clobber
   350  				// the content below the SP. We store the LR again after the SP is
   351  				// decremented.
   352  				q = obj.Appendp(q, newprog)
   353  				q.As = mov
   354  				q.Pos = p.Pos
   355  				q.From.Type = obj.TYPE_REG
   356  				q.From.Reg = REGLINK
   357  				q.To.Type = obj.TYPE_MEM
   358  				q.To.Offset = 0
   359  				q.To.Reg = REGSP
   360  			}
   361  
   362  			if c.cursym.Func().Text.From.Sym.Wrapper() && c.cursym.Func().Text.Mark&LEAF == 0 {
   363  				// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
   364  				//
   365  				//	MOV	g_panic(g), R1
   366  				//	BEQ	R1, end
   367  				//	MOV	panic_argp(R1), R2
   368  				//	ADD	$(autosize+FIXED_FRAME), R29, R3
   369  				//	BNE	R2, R3, end
   370  				//	ADD	$FIXED_FRAME, R29, R2
   371  				//	MOV	R2, panic_argp(R1)
   372  				// end:
   373  				//	NOP
   374  				//
   375  				// The NOP is needed to give the jumps somewhere to land.
   376  				// It is a liblink NOP, not an mips NOP: it encodes to 0 instruction bytes.
   377  				//
   378  				// We don't generate this for leafs because that means the wrapped
   379  				// function was inlined into the wrapper.
   380  
   381  				q = obj.Appendp(q, newprog)
   382  
   383  				q.As = mov
   384  				q.From.Type = obj.TYPE_MEM
   385  				q.From.Reg = REGG
   386  				q.From.Offset = 4 * int64(c.ctxt.Arch.PtrSize) // G.panic
   387  				q.To.Type = obj.TYPE_REG
   388  				q.To.Reg = REG_R1
   389  
   390  				q = obj.Appendp(q, newprog)
   391  				q.As = ABEQ
   392  				q.From.Type = obj.TYPE_REG
   393  				q.From.Reg = REG_R1
   394  				q.To.Type = obj.TYPE_BRANCH
   395  				q.Mark |= BRANCH
   396  				p1 = q
   397  
   398  				q = obj.Appendp(q, newprog)
   399  				q.As = mov
   400  				q.From.Type = obj.TYPE_MEM
   401  				q.From.Reg = REG_R1
   402  				q.From.Offset = 0 // Panic.argp
   403  				q.To.Type = obj.TYPE_REG
   404  				q.To.Reg = REG_R2
   405  
   406  				q = obj.Appendp(q, newprog)
   407  				q.As = add
   408  				q.From.Type = obj.TYPE_CONST
   409  				q.From.Offset = int64(autosize) + ctxt.Arch.FixedFrameSize
   410  				q.Reg = REGSP
   411  				q.To.Type = obj.TYPE_REG
   412  				q.To.Reg = REG_R3
   413  
   414  				q = obj.Appendp(q, newprog)
   415  				q.As = ABNE
   416  				q.From.Type = obj.TYPE_REG
   417  				q.From.Reg = REG_R2
   418  				q.Reg = REG_R3
   419  				q.To.Type = obj.TYPE_BRANCH
   420  				q.Mark |= BRANCH
   421  				p2 = q
   422  
   423  				q = obj.Appendp(q, newprog)
   424  				q.As = add
   425  				q.From.Type = obj.TYPE_CONST
   426  				q.From.Offset = ctxt.Arch.FixedFrameSize
   427  				q.Reg = REGSP
   428  				q.To.Type = obj.TYPE_REG
   429  				q.To.Reg = REG_R2
   430  
   431  				q = obj.Appendp(q, newprog)
   432  				q.As = mov
   433  				q.From.Type = obj.TYPE_REG
   434  				q.From.Reg = REG_R2
   435  				q.To.Type = obj.TYPE_MEM
   436  				q.To.Reg = REG_R1
   437  				q.To.Offset = 0 // Panic.argp
   438  
   439  				q = obj.Appendp(q, newprog)
   440  
   441  				q.As = obj.ANOP
   442  				p1.To.SetTarget(q)
   443  				p2.To.SetTarget(q)
   444  			}
   445  
   446  		case ARET:
   447  			if p.From.Type == obj.TYPE_CONST {
   448  				ctxt.Diag("using BECOME (%v) is not supported!", p)
   449  				break
   450  			}
   451  
   452  			retSym := p.To.Sym
   453  			p.To.Name = obj.NAME_NONE // clear fields as we may modify p to other instruction
   454  			p.To.Sym = nil
   455  
   456  			if c.cursym.Func().Text.Mark&LEAF != 0 {
   457  				if autosize == 0 {
   458  					p.As = AJMP
   459  					p.From = obj.Addr{}
   460  					if retSym != nil { // retjmp
   461  						p.To.Type = obj.TYPE_BRANCH
   462  						p.To.Name = obj.NAME_EXTERN
   463  						p.To.Sym = retSym
   464  					} else {
   465  						p.To.Type = obj.TYPE_MEM
   466  						p.To.Reg = REGLINK
   467  						p.To.Offset = 0
   468  					}
   469  					p.Mark |= BRANCH
   470  					break
   471  				}
   472  
   473  				p.As = add
   474  				p.From.Type = obj.TYPE_CONST
   475  				p.From.Offset = int64(autosize)
   476  				p.To.Type = obj.TYPE_REG
   477  				p.To.Reg = REGSP
   478  				p.Spadj = -autosize
   479  
   480  				q = c.newprog()
   481  				q.As = AJMP
   482  				q.Pos = p.Pos
   483  				if retSym != nil { // retjmp
   484  					q.To.Type = obj.TYPE_BRANCH
   485  					q.To.Name = obj.NAME_EXTERN
   486  					q.To.Sym = retSym
   487  				} else {
   488  					q.To.Type = obj.TYPE_MEM
   489  					q.To.Reg = REGLINK
   490  					q.To.Offset = 0
   491  				}
   492  				q.Mark |= BRANCH
   493  				q.Spadj = +autosize
   494  
   495  				q.Link = p.Link
   496  				p.Link = q
   497  				break
   498  			}
   499  
   500  			p.As = mov
   501  			p.From.Type = obj.TYPE_MEM
   502  			p.From.Offset = 0
   503  			p.From.Reg = REGSP
   504  			p.To.Type = obj.TYPE_REG
   505  			p.To.Reg = REGLINK
   506  
   507  			if autosize != 0 {
   508  				q = c.newprog()
   509  				q.As = add
   510  				q.Pos = p.Pos
   511  				q.From.Type = obj.TYPE_CONST
   512  				q.From.Offset = int64(autosize)
   513  				q.To.Type = obj.TYPE_REG
   514  				q.To.Reg = REGSP
   515  				q.Spadj = -autosize
   516  
   517  				q.Link = p.Link
   518  				p.Link = q
   519  			}
   520  
   521  			q1 = c.newprog()
   522  			q1.As = AJMP
   523  			q1.Pos = p.Pos
   524  			if retSym != nil { // retjmp
   525  				q1.To.Type = obj.TYPE_BRANCH
   526  				q1.To.Name = obj.NAME_EXTERN
   527  				q1.To.Sym = retSym
   528  			} else {
   529  				q1.To.Type = obj.TYPE_MEM
   530  				q1.To.Offset = 0
   531  				q1.To.Reg = REGLINK
   532  			}
   533  			q1.Mark |= BRANCH
   534  			q1.Spadj = +autosize
   535  
   536  			q1.Link = q.Link
   537  			q.Link = q1
   538  
   539  		case AADD,
   540  			AADDU,
   541  			AADDV,
   542  			AADDVU:
   543  			if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.From.Type == obj.TYPE_CONST {
   544  				p.Spadj = int32(-p.From.Offset)
   545  			}
   546  
   547  		case obj.AGETCALLERPC:
   548  			if cursym.Leaf() {
   549  				/* MOV LR, Rd */
   550  				p.As = mov
   551  				p.From.Type = obj.TYPE_REG
   552  				p.From.Reg = REGLINK
   553  			} else {
   554  				/* MOV (RSP), Rd */
   555  				p.As = mov
   556  				p.From.Type = obj.TYPE_MEM
   557  				p.From.Reg = REGSP
   558  			}
   559  		}
   560  
   561  		if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.Spadj == 0 {
   562  			f := c.cursym.Func()
   563  			if f.FuncFlag&abi.FuncFlagSPWrite == 0 {
   564  				c.cursym.Func().FuncFlag |= abi.FuncFlagSPWrite
   565  				if ctxt.Debugvlog || !ctxt.IsAsm {
   566  					ctxt.Logf("auto-SPWRITE: %s %v\n", c.cursym.Name, p)
   567  					if !ctxt.IsAsm {
   568  						ctxt.Diag("invalid auto-SPWRITE in non-assembly")
   569  						ctxt.DiagFlush()
   570  						log.Fatalf("bad SPWRITE")
   571  					}
   572  				}
   573  			}
   574  		}
   575  	}
   576  
   577  	if c.ctxt.Arch.Family == sys.MIPS {
   578  		// rewrite MOVD into two MOVF in 32-bit mode to avoid unaligned memory access
   579  		for p = c.cursym.Func().Text; p != nil; p = p1 {
   580  			p1 = p.Link
   581  
   582  			if p.As != AMOVD {
   583  				continue
   584  			}
   585  			if p.From.Type != obj.TYPE_MEM && p.To.Type != obj.TYPE_MEM {
   586  				continue
   587  			}
   588  
   589  			p.As = AMOVF
   590  			q = c.newprog()
   591  			*q = *p
   592  			q.Link = p.Link
   593  			p.Link = q
   594  			p1 = q.Link
   595  
   596  			var addrOff int64
   597  			if c.ctxt.Arch.ByteOrder == binary.BigEndian {
   598  				addrOff = 4 // swap load/save order
   599  			}
   600  			if p.From.Type == obj.TYPE_MEM {
   601  				reg := REG_F0 + (p.To.Reg-REG_F0)&^1
   602  				p.To.Reg = reg
   603  				q.To.Reg = reg + 1
   604  				p.From.Offset += addrOff
   605  				q.From.Offset += 4 - addrOff
   606  			} else if p.To.Type == obj.TYPE_MEM {
   607  				reg := REG_F0 + (p.From.Reg-REG_F0)&^1
   608  				p.From.Reg = reg
   609  				q.From.Reg = reg + 1
   610  				p.To.Offset += addrOff
   611  				q.To.Offset += 4 - addrOff
   612  			}
   613  		}
   614  	}
   615  
   616  	if nosched {
   617  		// if we don't do instruction scheduling, simply add
   618  		// NOP after each branch instruction.
   619  		for p = c.cursym.Func().Text; p != nil; p = p.Link {
   620  			if p.Mark&BRANCH != 0 {
   621  				c.addnop(p)
   622  			}
   623  		}
   624  		return
   625  	}
   626  
   627  	// instruction scheduling
   628  	q = nil                   // p - 1
   629  	q1 = c.cursym.Func().Text // top of block
   630  	o := 0                    // count of instructions
   631  	for p = c.cursym.Func().Text; p != nil; p = p1 {
   632  		p1 = p.Link
   633  		o++
   634  		if p.Mark&NOSCHED != 0 {
   635  			if q1 != p {
   636  				c.sched(q1, q)
   637  			}
   638  			for ; p != nil; p = p.Link {
   639  				if p.Mark&NOSCHED == 0 {
   640  					break
   641  				}
   642  				q = p
   643  			}
   644  			p1 = p
   645  			q1 = p
   646  			o = 0
   647  			continue
   648  		}
   649  		if p.Mark&(LABEL|SYNC) != 0 {
   650  			if q1 != p {
   651  				c.sched(q1, q)
   652  			}
   653  			q1 = p
   654  			o = 1
   655  		}
   656  		if p.Mark&(BRANCH|SYNC) != 0 {
   657  			c.sched(q1, p)
   658  			q1 = p1
   659  			o = 0
   660  		}
   661  		if o >= NSCHED {
   662  			c.sched(q1, p)
   663  			q1 = p1
   664  			o = 0
   665  		}
   666  		q = p
   667  	}
   668  }
   669  
   670  func (c *ctxt0) stacksplit(p *obj.Prog, framesize int32) *obj.Prog {
   671  	var mov, add obj.As
   672  
   673  	if c.ctxt.Arch.Family == sys.MIPS64 {
   674  		add = AADDV
   675  		mov = AMOVV
   676  	} else {
   677  		add = AADDU
   678  		mov = AMOVW
   679  	}
   680  
   681  	if c.ctxt.Flag_maymorestack != "" {
   682  		// Save LR and REGCTXT.
   683  		frameSize := 2 * c.ctxt.Arch.PtrSize
   684  
   685  		p = c.ctxt.StartUnsafePoint(p, c.newprog)
   686  
   687  		// MOV	REGLINK, -8/-16(SP)
   688  		p = obj.Appendp(p, c.newprog)
   689  		p.As = mov
   690  		p.From.Type = obj.TYPE_REG
   691  		p.From.Reg = REGLINK
   692  		p.To.Type = obj.TYPE_MEM
   693  		p.To.Offset = int64(-frameSize)
   694  		p.To.Reg = REGSP
   695  
   696  		// MOV	REGCTXT, -4/-8(SP)
   697  		p = obj.Appendp(p, c.newprog)
   698  		p.As = mov
   699  		p.From.Type = obj.TYPE_REG
   700  		p.From.Reg = REGCTXT
   701  		p.To.Type = obj.TYPE_MEM
   702  		p.To.Offset = -int64(c.ctxt.Arch.PtrSize)
   703  		p.To.Reg = REGSP
   704  
   705  		// ADD	$-8/$-16, SP
   706  		p = obj.Appendp(p, c.newprog)
   707  		p.As = add
   708  		p.From.Type = obj.TYPE_CONST
   709  		p.From.Offset = int64(-frameSize)
   710  		p.To.Type = obj.TYPE_REG
   711  		p.To.Reg = REGSP
   712  		p.Spadj = int32(frameSize)
   713  
   714  		// JAL	maymorestack
   715  		p = obj.Appendp(p, c.newprog)
   716  		p.As = AJAL
   717  		p.To.Type = obj.TYPE_BRANCH
   718  		// See ../x86/obj6.go
   719  		p.To.Sym = c.ctxt.LookupABI(c.ctxt.Flag_maymorestack, c.cursym.ABI())
   720  		p.Mark |= BRANCH
   721  
   722  		// Restore LR and REGCTXT.
   723  
   724  		// MOV	0(SP), REGLINK
   725  		p = obj.Appendp(p, c.newprog)
   726  		p.As = mov
   727  		p.From.Type = obj.TYPE_MEM
   728  		p.From.Offset = 0
   729  		p.From.Reg = REGSP
   730  		p.To.Type = obj.TYPE_REG
   731  		p.To.Reg = REGLINK
   732  
   733  		// MOV	4/8(SP), REGCTXT
   734  		p = obj.Appendp(p, c.newprog)
   735  		p.As = mov
   736  		p.From.Type = obj.TYPE_MEM
   737  		p.From.Offset = int64(c.ctxt.Arch.PtrSize)
   738  		p.From.Reg = REGSP
   739  		p.To.Type = obj.TYPE_REG
   740  		p.To.Reg = REGCTXT
   741  
   742  		// ADD	$8/$16, SP
   743  		p = obj.Appendp(p, c.newprog)
   744  		p.As = add
   745  		p.From.Type = obj.TYPE_CONST
   746  		p.From.Offset = int64(frameSize)
   747  		p.To.Type = obj.TYPE_REG
   748  		p.To.Reg = REGSP
   749  		p.Spadj = int32(-frameSize)
   750  
   751  		p = c.ctxt.EndUnsafePoint(p, c.newprog, -1)
   752  	}
   753  
   754  	// Jump back to here after morestack returns.
   755  	startPred := p
   756  
   757  	// MOV	g_stackguard(g), R1
   758  	p = obj.Appendp(p, c.newprog)
   759  
   760  	p.As = mov
   761  	p.From.Type = obj.TYPE_MEM
   762  	p.From.Reg = REGG
   763  	p.From.Offset = 2 * int64(c.ctxt.Arch.PtrSize) // G.stackguard0
   764  	if c.cursym.CFunc() {
   765  		p.From.Offset = 3 * int64(c.ctxt.Arch.PtrSize) // G.stackguard1
   766  	}
   767  	p.To.Type = obj.TYPE_REG
   768  	p.To.Reg = REG_R1
   769  
   770  	// Mark the stack bound check and morestack call async nonpreemptible.
   771  	// If we get preempted here, when resumed the preemption request is
   772  	// cleared, but we'll still call morestack, which will double the stack
   773  	// unnecessarily. See issue #35470.
   774  	p = c.ctxt.StartUnsafePoint(p, c.newprog)
   775  
   776  	var q *obj.Prog
   777  	if framesize <= abi.StackSmall {
   778  		// small stack: SP < stackguard
   779  		//	AGTU	SP, stackguard, R1
   780  		p = obj.Appendp(p, c.newprog)
   781  
   782  		p.As = ASGTU
   783  		p.From.Type = obj.TYPE_REG
   784  		p.From.Reg = REGSP
   785  		p.Reg = REG_R1
   786  		p.To.Type = obj.TYPE_REG
   787  		p.To.Reg = REG_R1
   788  	} else {
   789  		// large stack: SP-framesize < stackguard-StackSmall
   790  		offset := int64(framesize) - abi.StackSmall
   791  		if framesize > abi.StackBig {
   792  			// Such a large stack we need to protect against underflow.
   793  			// The runtime guarantees SP > objabi.StackBig, but
   794  			// framesize is large enough that SP-framesize may
   795  			// underflow, causing a direct comparison with the
   796  			// stack guard to incorrectly succeed. We explicitly
   797  			// guard against underflow.
   798  			//
   799  			//	SGTU	$(framesize-StackSmall), SP, R2
   800  			//	BNE	R2, label-of-call-to-morestack
   801  
   802  			p = obj.Appendp(p, c.newprog)
   803  			p.As = ASGTU
   804  			p.From.Type = obj.TYPE_CONST
   805  			p.From.Offset = offset
   806  			p.Reg = REGSP
   807  			p.To.Type = obj.TYPE_REG
   808  			p.To.Reg = REG_R2
   809  
   810  			p = obj.Appendp(p, c.newprog)
   811  			q = p
   812  			p.As = ABNE
   813  			p.From.Type = obj.TYPE_REG
   814  			p.From.Reg = REG_R2
   815  			p.To.Type = obj.TYPE_BRANCH
   816  			p.Mark |= BRANCH
   817  		}
   818  
   819  		// Check against the stack guard. We've ensured this won't underflow.
   820  		//	ADD	$-(framesize-StackSmall), SP, R2
   821  		//	SGTU	R2, stackguard, R1
   822  		p = obj.Appendp(p, c.newprog)
   823  
   824  		p.As = add
   825  		p.From.Type = obj.TYPE_CONST
   826  		p.From.Offset = -offset
   827  		p.Reg = REGSP
   828  		p.To.Type = obj.TYPE_REG
   829  		p.To.Reg = REG_R2
   830  
   831  		p = obj.Appendp(p, c.newprog)
   832  		p.As = ASGTU
   833  		p.From.Type = obj.TYPE_REG
   834  		p.From.Reg = REG_R2
   835  		p.Reg = REG_R1
   836  		p.To.Type = obj.TYPE_REG
   837  		p.To.Reg = REG_R1
   838  	}
   839  
   840  	// q1: BNE	R1, done
   841  	p = obj.Appendp(p, c.newprog)
   842  	q1 := p
   843  
   844  	p.As = ABNE
   845  	p.From.Type = obj.TYPE_REG
   846  	p.From.Reg = REG_R1
   847  	p.To.Type = obj.TYPE_BRANCH
   848  	p.Mark |= BRANCH
   849  
   850  	// MOV	LINK, R3
   851  	p = obj.Appendp(p, c.newprog)
   852  
   853  	p.As = mov
   854  	p.From.Type = obj.TYPE_REG
   855  	p.From.Reg = REGLINK
   856  	p.To.Type = obj.TYPE_REG
   857  	p.To.Reg = REG_R3
   858  	if q != nil {
   859  		q.To.SetTarget(p)
   860  		p.Mark |= LABEL
   861  	}
   862  
   863  	p = c.ctxt.EmitEntryStackMap(c.cursym, p, c.newprog)
   864  
   865  	// JAL	runtime.morestack(SB)
   866  	p = obj.Appendp(p, c.newprog)
   867  
   868  	p.As = AJAL
   869  	p.To.Type = obj.TYPE_BRANCH
   870  	if c.cursym.CFunc() {
   871  		p.To.Sym = c.ctxt.Lookup("runtime.morestackc")
   872  	} else if !c.cursym.Func().Text.From.Sym.NeedCtxt() {
   873  		p.To.Sym = c.ctxt.Lookup("runtime.morestack_noctxt")
   874  	} else {
   875  		p.To.Sym = c.ctxt.Lookup("runtime.morestack")
   876  	}
   877  	p.Mark |= BRANCH
   878  
   879  	p = c.ctxt.EndUnsafePoint(p, c.newprog, -1)
   880  
   881  	// JMP	start
   882  	p = obj.Appendp(p, c.newprog)
   883  
   884  	p.As = AJMP
   885  	p.To.Type = obj.TYPE_BRANCH
   886  	p.To.SetTarget(startPred.Link)
   887  	startPred.Link.Mark |= LABEL
   888  	p.Mark |= BRANCH
   889  
   890  	// placeholder for q1's jump target
   891  	p = obj.Appendp(p, c.newprog)
   892  
   893  	p.As = obj.ANOP // zero-width place holder
   894  	q1.To.SetTarget(p)
   895  
   896  	return p
   897  }
   898  
   899  func (c *ctxt0) addnop(p *obj.Prog) {
   900  	q := c.newprog()
   901  	q.As = ANOOP
   902  	q.Pos = p.Pos
   903  	q.Link = p.Link
   904  	p.Link = q
   905  }
   906  
   907  const (
   908  	E_HILO  = 1 << 0
   909  	E_FCR   = 1 << 1
   910  	E_MCR   = 1 << 2
   911  	E_MEM   = 1 << 3
   912  	E_MEMSP = 1 << 4 /* uses offset and size */
   913  	E_MEMSB = 1 << 5 /* uses offset and size */
   914  	ANYMEM  = E_MEM | E_MEMSP | E_MEMSB
   915  	//DELAY = LOAD|BRANCH|FCMP
   916  	DELAY = BRANCH /* only schedule branch */
   917  )
   918  
   919  type Dep struct {
   920  	ireg uint32
   921  	freg uint32
   922  	cc   uint32
   923  }
   924  
   925  type Sch struct {
   926  	p       obj.Prog
   927  	set     Dep
   928  	used    Dep
   929  	soffset int32
   930  	size    uint8
   931  	nop     uint8
   932  	comp    bool
   933  }
   934  
   935  func (c *ctxt0) sched(p0, pe *obj.Prog) {
   936  	var sch [NSCHED]Sch
   937  
   938  	/*
   939  	 * build side structure
   940  	 */
   941  	s := sch[:]
   942  	for p := p0; ; p = p.Link {
   943  		s[0].p = *p
   944  		c.markregused(&s[0])
   945  		if p == pe {
   946  			break
   947  		}
   948  		s = s[1:]
   949  	}
   950  	se := s
   951  
   952  	for i := cap(sch) - cap(se); i >= 0; i-- {
   953  		s = sch[i:]
   954  		if s[0].p.Mark&DELAY == 0 {
   955  			continue
   956  		}
   957  		if -cap(s) < -cap(se) {
   958  			if !conflict(&s[0], &s[1]) {
   959  				continue
   960  			}
   961  		}
   962  
   963  		var t []Sch
   964  		var j int
   965  		for j = cap(sch) - cap(s) - 1; j >= 0; j-- {
   966  			t = sch[j:]
   967  			if t[0].comp {
   968  				if s[0].p.Mark&BRANCH != 0 {
   969  					continue
   970  				}
   971  			}
   972  			if t[0].p.Mark&DELAY != 0 {
   973  				if -cap(s) >= -cap(se) || conflict(&t[0], &s[1]) {
   974  					continue
   975  				}
   976  			}
   977  			for u := t[1:]; -cap(u) <= -cap(s); u = u[1:] {
   978  				if c.depend(&u[0], &t[0]) {
   979  					continue
   980  				}
   981  			}
   982  			goto out2
   983  		}
   984  
   985  		if s[0].p.Mark&BRANCH != 0 {
   986  			s[0].nop = 1
   987  		}
   988  		continue
   989  
   990  	out2:
   991  		// t[0] is the instruction being moved to fill the delay
   992  		stmp := t[0]
   993  		copy(t[:i-j], t[1:i-j+1])
   994  		s[0] = stmp
   995  
   996  		if t[i-j-1].p.Mark&BRANCH != 0 {
   997  			// t[i-j] is being put into a branch delay slot
   998  			// combine its Spadj with the branch instruction
   999  			t[i-j-1].p.Spadj += t[i-j].p.Spadj
  1000  			t[i-j].p.Spadj = 0
  1001  		}
  1002  
  1003  		i--
  1004  	}
  1005  
  1006  	/*
  1007  	 * put it all back
  1008  	 */
  1009  	var p *obj.Prog
  1010  	var q *obj.Prog
  1011  	for s, p = sch[:], p0; -cap(s) <= -cap(se); s, p = s[1:], q {
  1012  		q = p.Link
  1013  		if q != s[0].p.Link {
  1014  			*p = s[0].p
  1015  			p.Link = q
  1016  		}
  1017  		for s[0].nop != 0 {
  1018  			s[0].nop--
  1019  			c.addnop(p)
  1020  		}
  1021  	}
  1022  }
  1023  
  1024  func (c *ctxt0) markregused(s *Sch) {
  1025  	p := &s.p
  1026  	s.comp = c.compound(p)
  1027  	s.nop = 0
  1028  	if s.comp {
  1029  		s.set.ireg |= 1 << (REGTMP - REG_R0)
  1030  		s.used.ireg |= 1 << (REGTMP - REG_R0)
  1031  	}
  1032  
  1033  	ar := 0  /* dest is really reference */
  1034  	ad := 0  /* source/dest is really address */
  1035  	ld := 0  /* opcode is load instruction */
  1036  	sz := 20 /* size of load/store for overlap computation */
  1037  
  1038  	/*
  1039  	 * flags based on opcode
  1040  	 */
  1041  	switch p.As {
  1042  	case obj.ATEXT:
  1043  		c.autosize = int32(p.To.Offset + 8)
  1044  		ad = 1
  1045  
  1046  	case AJAL:
  1047  		r := p.Reg
  1048  		if r == 0 {
  1049  			r = REGLINK
  1050  		}
  1051  		s.set.ireg |= 1 << uint(r-REG_R0)
  1052  		ar = 1
  1053  		ad = 1
  1054  
  1055  	case ABGEZAL,
  1056  		ABLTZAL:
  1057  		s.set.ireg |= 1 << (REGLINK - REG_R0)
  1058  		fallthrough
  1059  	case ABEQ,
  1060  		ABGEZ,
  1061  		ABGTZ,
  1062  		ABLEZ,
  1063  		ABLTZ,
  1064  		ABNE:
  1065  		ar = 1
  1066  		ad = 1
  1067  
  1068  	case ABFPT,
  1069  		ABFPF:
  1070  		ad = 1
  1071  		s.used.cc |= E_FCR
  1072  
  1073  	case ACMPEQD,
  1074  		ACMPEQF,
  1075  		ACMPGED,
  1076  		ACMPGEF,
  1077  		ACMPGTD,
  1078  		ACMPGTF:
  1079  		ar = 1
  1080  		s.set.cc |= E_FCR
  1081  		p.Mark |= FCMP
  1082  
  1083  	case AJMP:
  1084  		ar = 1
  1085  		ad = 1
  1086  
  1087  	case AMOVB,
  1088  		AMOVBU:
  1089  		sz = 1
  1090  		ld = 1
  1091  
  1092  	case AMOVH,
  1093  		AMOVHU:
  1094  		sz = 2
  1095  		ld = 1
  1096  
  1097  	case AMOVF,
  1098  		AMOVW,
  1099  		AMOVWL,
  1100  		AMOVWR:
  1101  		sz = 4
  1102  		ld = 1
  1103  
  1104  	case AMOVD,
  1105  		AMOVV,
  1106  		AMOVVL,
  1107  		AMOVVR:
  1108  		sz = 8
  1109  		ld = 1
  1110  
  1111  	case ADIV,
  1112  		ADIVU,
  1113  		AMUL,
  1114  		AMULU,
  1115  		AREM,
  1116  		AREMU,
  1117  		ADIVV,
  1118  		ADIVVU,
  1119  		AMULV,
  1120  		AMULVU,
  1121  		AREMV,
  1122  		AREMVU:
  1123  		s.set.cc = E_HILO
  1124  		fallthrough
  1125  	case AADD,
  1126  		AADDU,
  1127  		AADDV,
  1128  		AADDVU,
  1129  		AAND,
  1130  		ANOR,
  1131  		AOR,
  1132  		ASGT,
  1133  		ASGTU,
  1134  		ASLL,
  1135  		ASRA,
  1136  		ASRL,
  1137  		ASLLV,
  1138  		ASRAV,
  1139  		ASRLV,
  1140  		ASUB,
  1141  		ASUBU,
  1142  		ASUBV,
  1143  		ASUBVU,
  1144  		AXOR,
  1145  
  1146  		AADDD,
  1147  		AADDF,
  1148  		AADDW,
  1149  		ASUBD,
  1150  		ASUBF,
  1151  		ASUBW,
  1152  		AMULF,
  1153  		AMULD,
  1154  		AMULW,
  1155  		ADIVF,
  1156  		ADIVD,
  1157  		ADIVW:
  1158  		if p.Reg == 0 {
  1159  			if p.To.Type == obj.TYPE_REG {
  1160  				p.Reg = p.To.Reg
  1161  			}
  1162  			//if(p->reg == NREG)
  1163  			//	print("botch %P\n", p);
  1164  		}
  1165  	}
  1166  
  1167  	/*
  1168  	 * flags based on 'to' field
  1169  	 */
  1170  	cls := int(p.To.Class)
  1171  	if cls == 0 {
  1172  		cls = c.aclass(&p.To) + 1
  1173  		p.To.Class = int8(cls)
  1174  	}
  1175  	cls--
  1176  	switch cls {
  1177  	default:
  1178  		fmt.Printf("unknown class %d %v\n", cls, p)
  1179  
  1180  	case C_ZCON,
  1181  		C_SCON,
  1182  		C_ADD0CON,
  1183  		C_AND0CON,
  1184  		C_ADDCON,
  1185  		C_ANDCON,
  1186  		C_UCON,
  1187  		C_LCON,
  1188  		C_NONE,
  1189  		C_SBRA,
  1190  		C_LBRA,
  1191  		C_ADDR,
  1192  		C_TEXTSIZE:
  1193  		break
  1194  
  1195  	case C_HI,
  1196  		C_LO:
  1197  		s.set.cc |= E_HILO
  1198  
  1199  	case C_FCREG:
  1200  		s.set.cc |= E_FCR
  1201  
  1202  	case C_MREG:
  1203  		s.set.cc |= E_MCR
  1204  
  1205  	case C_ZOREG,
  1206  		C_SOREG,
  1207  		C_LOREG:
  1208  		cls = int(p.To.Reg)
  1209  		s.used.ireg |= 1 << uint(cls-REG_R0)
  1210  		if ad != 0 {
  1211  			break
  1212  		}
  1213  		s.size = uint8(sz)
  1214  		s.soffset = c.regoff(&p.To)
  1215  
  1216  		m := uint32(ANYMEM)
  1217  		if cls == REGSB {
  1218  			m = E_MEMSB
  1219  		}
  1220  		if cls == REGSP {
  1221  			m = E_MEMSP
  1222  		}
  1223  
  1224  		if ar != 0 {
  1225  			s.used.cc |= m
  1226  		} else {
  1227  			s.set.cc |= m
  1228  		}
  1229  
  1230  	case C_SACON,
  1231  		C_LACON:
  1232  		s.used.ireg |= 1 << (REGSP - REG_R0)
  1233  
  1234  	case C_SECON,
  1235  		C_LECON:
  1236  		s.used.ireg |= 1 << (REGSB - REG_R0)
  1237  
  1238  	case C_REG:
  1239  		if ar != 0 {
  1240  			s.used.ireg |= 1 << uint(p.To.Reg-REG_R0)
  1241  		} else {
  1242  			s.set.ireg |= 1 << uint(p.To.Reg-REG_R0)
  1243  		}
  1244  
  1245  	case C_FREG:
  1246  		if ar != 0 {
  1247  			s.used.freg |= 1 << uint(p.To.Reg-REG_F0)
  1248  		} else {
  1249  			s.set.freg |= 1 << uint(p.To.Reg-REG_F0)
  1250  		}
  1251  		if ld != 0 && p.From.Type == obj.TYPE_REG {
  1252  			p.Mark |= LOAD
  1253  		}
  1254  
  1255  	case C_SAUTO,
  1256  		C_LAUTO:
  1257  		s.used.ireg |= 1 << (REGSP - REG_R0)
  1258  		if ad != 0 {
  1259  			break
  1260  		}
  1261  		s.size = uint8(sz)
  1262  		s.soffset = c.regoff(&p.To)
  1263  
  1264  		if ar != 0 {
  1265  			s.used.cc |= E_MEMSP
  1266  		} else {
  1267  			s.set.cc |= E_MEMSP
  1268  		}
  1269  
  1270  	case C_SEXT,
  1271  		C_LEXT:
  1272  		s.used.ireg |= 1 << (REGSB - REG_R0)
  1273  		if ad != 0 {
  1274  			break
  1275  		}
  1276  		s.size = uint8(sz)
  1277  		s.soffset = c.regoff(&p.To)
  1278  
  1279  		if ar != 0 {
  1280  			s.used.cc |= E_MEMSB
  1281  		} else {
  1282  			s.set.cc |= E_MEMSB
  1283  		}
  1284  	}
  1285  
  1286  	/*
  1287  	 * flags based on 'from' field
  1288  	 */
  1289  	cls = int(p.From.Class)
  1290  	if cls == 0 {
  1291  		cls = c.aclass(&p.From) + 1
  1292  		p.From.Class = int8(cls)
  1293  	}
  1294  	cls--
  1295  	switch cls {
  1296  	default:
  1297  		fmt.Printf("unknown class %d %v\n", cls, p)
  1298  
  1299  	case C_ZCON,
  1300  		C_SCON,
  1301  		C_ADD0CON,
  1302  		C_AND0CON,
  1303  		C_ADDCON,
  1304  		C_ANDCON,
  1305  		C_UCON,
  1306  		C_LCON,
  1307  		C_NONE,
  1308  		C_SBRA,
  1309  		C_LBRA,
  1310  		C_ADDR,
  1311  		C_TEXTSIZE:
  1312  		break
  1313  
  1314  	case C_HI,
  1315  		C_LO:
  1316  		s.used.cc |= E_HILO
  1317  
  1318  	case C_FCREG:
  1319  		s.used.cc |= E_FCR
  1320  
  1321  	case C_MREG:
  1322  		s.used.cc |= E_MCR
  1323  
  1324  	case C_ZOREG,
  1325  		C_SOREG,
  1326  		C_LOREG:
  1327  		cls = int(p.From.Reg)
  1328  		s.used.ireg |= 1 << uint(cls-REG_R0)
  1329  		if ld != 0 {
  1330  			p.Mark |= LOAD
  1331  		}
  1332  		s.size = uint8(sz)
  1333  		s.soffset = c.regoff(&p.From)
  1334  
  1335  		m := uint32(ANYMEM)
  1336  		if cls == REGSB {
  1337  			m = E_MEMSB
  1338  		}
  1339  		if cls == REGSP {
  1340  			m = E_MEMSP
  1341  		}
  1342  
  1343  		s.used.cc |= m
  1344  
  1345  	case C_SACON,
  1346  		C_LACON:
  1347  		cls = int(p.From.Reg)
  1348  		if cls == 0 {
  1349  			cls = REGSP
  1350  		}
  1351  		s.used.ireg |= 1 << uint(cls-REG_R0)
  1352  
  1353  	case C_SECON,
  1354  		C_LECON:
  1355  		s.used.ireg |= 1 << (REGSB - REG_R0)
  1356  
  1357  	case C_REG:
  1358  		s.used.ireg |= 1 << uint(p.From.Reg-REG_R0)
  1359  
  1360  	case C_FREG:
  1361  		s.used.freg |= 1 << uint(p.From.Reg-REG_F0)
  1362  		if ld != 0 && p.To.Type == obj.TYPE_REG {
  1363  			p.Mark |= LOAD
  1364  		}
  1365  
  1366  	case C_SAUTO,
  1367  		C_LAUTO:
  1368  		s.used.ireg |= 1 << (REGSP - REG_R0)
  1369  		if ld != 0 {
  1370  			p.Mark |= LOAD
  1371  		}
  1372  		if ad != 0 {
  1373  			break
  1374  		}
  1375  		s.size = uint8(sz)
  1376  		s.soffset = c.regoff(&p.From)
  1377  
  1378  		s.used.cc |= E_MEMSP
  1379  
  1380  	case C_SEXT:
  1381  	case C_LEXT:
  1382  		s.used.ireg |= 1 << (REGSB - REG_R0)
  1383  		if ld != 0 {
  1384  			p.Mark |= LOAD
  1385  		}
  1386  		if ad != 0 {
  1387  			break
  1388  		}
  1389  		s.size = uint8(sz)
  1390  		s.soffset = c.regoff(&p.From)
  1391  
  1392  		s.used.cc |= E_MEMSB
  1393  	}
  1394  
  1395  	cls = int(p.Reg)
  1396  	if cls != 0 {
  1397  		if REG_F0 <= cls && cls <= REG_F31 {
  1398  			s.used.freg |= 1 << uint(cls-REG_F0)
  1399  		} else {
  1400  			s.used.ireg |= 1 << uint(cls-REG_R0)
  1401  		}
  1402  	}
  1403  	s.set.ireg &^= (1 << (REGZERO - REG_R0)) /* R0 can't be set */
  1404  }
  1405  
  1406  /*
  1407   * test to see if two instructions can be
  1408   * interchanged without changing semantics
  1409   */
  1410  func (c *ctxt0) depend(sa, sb *Sch) bool {
  1411  	if sa.set.ireg&(sb.set.ireg|sb.used.ireg) != 0 {
  1412  		return true
  1413  	}
  1414  	if sb.set.ireg&sa.used.ireg != 0 {
  1415  		return true
  1416  	}
  1417  
  1418  	if sa.set.freg&(sb.set.freg|sb.used.freg) != 0 {
  1419  		return true
  1420  	}
  1421  	if sb.set.freg&sa.used.freg != 0 {
  1422  		return true
  1423  	}
  1424  
  1425  	/*
  1426  	 * special case.
  1427  	 * loads from same address cannot pass.
  1428  	 * this is for hardware fifo's and the like
  1429  	 */
  1430  	if sa.used.cc&sb.used.cc&E_MEM != 0 {
  1431  		if sa.p.Reg == sb.p.Reg {
  1432  			if c.regoff(&sa.p.From) == c.regoff(&sb.p.From) {
  1433  				return true
  1434  			}
  1435  		}
  1436  	}
  1437  
  1438  	x := (sa.set.cc & (sb.set.cc | sb.used.cc)) | (sb.set.cc & sa.used.cc)
  1439  	if x != 0 {
  1440  		/*
  1441  		 * allow SB and SP to pass each other.
  1442  		 * allow SB to pass SB iff doffsets are ok
  1443  		 * anything else conflicts
  1444  		 */
  1445  		if x != E_MEMSP && x != E_MEMSB {
  1446  			return true
  1447  		}
  1448  		x = sa.set.cc | sb.set.cc | sa.used.cc | sb.used.cc
  1449  		if x&E_MEM != 0 {
  1450  			return true
  1451  		}
  1452  		if offoverlap(sa, sb) {
  1453  			return true
  1454  		}
  1455  	}
  1456  
  1457  	return false
  1458  }
  1459  
  1460  func offoverlap(sa, sb *Sch) bool {
  1461  	if sa.soffset < sb.soffset {
  1462  		if sa.soffset+int32(sa.size) > sb.soffset {
  1463  			return true
  1464  		}
  1465  		return false
  1466  	}
  1467  	if sb.soffset+int32(sb.size) > sa.soffset {
  1468  		return true
  1469  	}
  1470  	return false
  1471  }
  1472  
  1473  /*
  1474   * test 2 adjacent instructions
  1475   * and find out if inserted instructions
  1476   * are desired to prevent stalls.
  1477   */
  1478  func conflict(sa, sb *Sch) bool {
  1479  	if sa.set.ireg&sb.used.ireg != 0 {
  1480  		return true
  1481  	}
  1482  	if sa.set.freg&sb.used.freg != 0 {
  1483  		return true
  1484  	}
  1485  	if sa.set.cc&sb.used.cc != 0 {
  1486  		return true
  1487  	}
  1488  	return false
  1489  }
  1490  
  1491  func (c *ctxt0) compound(p *obj.Prog) bool {
  1492  	o := c.oplook(p)
  1493  	if o.size != 4 {
  1494  		return true
  1495  	}
  1496  	if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSB {
  1497  		return true
  1498  	}
  1499  	return false
  1500  }
  1501  
  1502  var Linkmips64 = obj.LinkArch{
  1503  	Arch:           sys.ArchMIPS64,
  1504  	Init:           buildop,
  1505  	Preprocess:     preprocess,
  1506  	Assemble:       span0,
  1507  	Progedit:       progedit,
  1508  	DWARFRegisters: MIPSDWARFRegisters,
  1509  }
  1510  
  1511  var Linkmips64le = obj.LinkArch{
  1512  	Arch:           sys.ArchMIPS64LE,
  1513  	Init:           buildop,
  1514  	Preprocess:     preprocess,
  1515  	Assemble:       span0,
  1516  	Progedit:       progedit,
  1517  	DWARFRegisters: MIPSDWARFRegisters,
  1518  }
  1519  
  1520  var Linkmips = obj.LinkArch{
  1521  	Arch:           sys.ArchMIPS,
  1522  	Init:           buildop,
  1523  	Preprocess:     preprocess,
  1524  	Assemble:       span0,
  1525  	Progedit:       progedit,
  1526  	DWARFRegisters: MIPSDWARFRegisters,
  1527  }
  1528  
  1529  var Linkmipsle = obj.LinkArch{
  1530  	Arch:           sys.ArchMIPSLE,
  1531  	Init:           buildop,
  1532  	Preprocess:     preprocess,
  1533  	Assemble:       span0,
  1534  	Progedit:       progedit,
  1535  	DWARFRegisters: MIPSDWARFRegisters,
  1536  }
  1537  

View as plain text