ggen.go

     1  // Copyright 2009 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  package amd64
     6  
     7  import (
     8  	"cmd/compile/internal/ir"
     9  	"cmd/compile/internal/objw"
    10  	"cmd/compile/internal/types"
    11  	"cmd/internal/obj"
    12  	"cmd/internal/obj/x86"
    13  )
    14  
    15  // DUFFZERO consists of repeated blocks of 4 MOVUPSs + LEAQ,
    16  // See runtime/mkduff.go.
    17  const (
    18  	dzBlocks    = 16 // number of MOV/ADD blocks
    19  	dzBlockLen  = 4  // number of clears per block
    20  	dzBlockSize = 23 // size of instructions in a single block
    21  	dzMovSize   = 5  // size of single MOV instruction w/ offset
    22  	dzLeaqSize  = 4  // size of single LEAQ instruction
    23  	dzClearStep = 16 // number of bytes cleared by each MOV instruction
    24  
    25  	dzClearLen = dzClearStep * dzBlockLen // bytes cleared by one block
    26  	dzSize     = dzBlocks * dzBlockSize
    27  )
    28  
    29  // dzOff returns the offset for a jump into DUFFZERO.
    30  // b is the number of bytes to zero.
    31  func dzOff(b int64) int64 {
    32  	off := int64(dzSize)
    33  	off -= b / dzClearLen * dzBlockSize
    34  	tailLen := b % dzClearLen
    35  	if tailLen >= dzClearStep {
    36  		off -= dzLeaqSize + dzMovSize*(tailLen/dzClearStep)
    37  	}
    38  	return off
    39  }
    40  
    41  // duffzeroDI returns the pre-adjustment to DI for a call to DUFFZERO.
    42  // b is the number of bytes to zero.
    43  func dzDI(b int64) int64 {
    44  	tailLen := b % dzClearLen
    45  	if tailLen < dzClearStep {
    46  		return 0
    47  	}
    48  	tailSteps := tailLen / dzClearStep
    49  	return -dzClearStep * (dzBlockLen - tailSteps)
    50  }
    51  
    52  func zerorange(pp *objw.Progs, p *obj.Prog, off, cnt int64, state *uint32) *obj.Prog {
    53  	const (
    54  		r13 = 1 << iota // if R13 is already zeroed.
    55  	)
    56  
    57  	if cnt == 0 {
    58  		return p
    59  	}
    60  
    61  	if cnt == 8 {
    62  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_X15, 0, obj.TYPE_MEM, x86.REG_SP, off)
    63  	} else if cnt <= int64(8*types.RegSize) {
    64  		for i := int64(0); i < cnt/16; i++ {
    65  			p = pp.Append(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X15, 0, obj.TYPE_MEM, x86.REG_SP, off+i*16)
    66  		}
    67  
    68  		if cnt%16 != 0 {
    69  			p = pp.Append(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X15, 0, obj.TYPE_MEM, x86.REG_SP, off+cnt-int64(16))
    70  		}
    71  	} else if cnt <= int64(128*types.RegSize) {
    72  		// Save DI to r12. With the amd64 Go register abi, DI can contain
    73  		// an incoming parameter, whereas R12 is always scratch.
    74  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_DI, 0, obj.TYPE_REG, x86.REG_R12, 0)
    75  		// Emit duffzero call
    76  		p = pp.Append(p, leaptr, obj.TYPE_MEM, x86.REG_SP, off+dzDI(cnt), obj.TYPE_REG, x86.REG_DI, 0)
    77  		p = pp.Append(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_ADDR, 0, dzOff(cnt))
    78  		p.To.Sym = ir.Syms.Duffzero
    79  		if cnt%16 != 0 {
    80  			p = pp.Append(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X15, 0, obj.TYPE_MEM, x86.REG_DI, -int64(8))
    81  		}
    82  		// Restore DI from r12
    83  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_R12, 0, obj.TYPE_REG, x86.REG_DI, 0)
    84  
    85  	} else {
    86  		// When the register ABI is in effect, at this point in the
    87  		// prolog we may have live values in all of RAX,RDI,RCX. Save
    88  		// them off to registers before the REPSTOSQ below, then
    89  		// restore. Note that R12 and R13 are always available as
    90  		// scratch regs; here we also use R15 (this is safe to do
    91  		// since there won't be any globals accessed in the prolog).
    92  		// See rewriteToUseGot() in obj6.go for more on r15 use.
    93  
    94  		// Save rax/rdi/rcx
    95  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_DI, 0, obj.TYPE_REG, x86.REG_R12, 0)
    96  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_REG, x86.REG_R13, 0)
    97  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_CX, 0, obj.TYPE_REG, x86.REG_R15, 0)
    98  
    99  		// Set up the REPSTOSQ and kick it off.
   100  		p = pp.Append(p, x86.AXORL, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_REG, x86.REG_AX, 0)
   101  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_CONST, 0, cnt/int64(types.RegSize), obj.TYPE_REG, x86.REG_CX, 0)
   102  		p = pp.Append(p, leaptr, obj.TYPE_MEM, x86.REG_SP, off, obj.TYPE_REG, x86.REG_DI, 0)
   103  		p = pp.Append(p, x86.AREP, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0)
   104  		p = pp.Append(p, x86.ASTOSQ, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0)
   105  
   106  		// Restore rax/rdi/rcx
   107  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_R12, 0, obj.TYPE_REG, x86.REG_DI, 0)
   108  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_R13, 0, obj.TYPE_REG, x86.REG_AX, 0)
   109  		p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_R15, 0, obj.TYPE_REG, x86.REG_CX, 0)
   110  
   111  		// Record the fact that r13 is no longer zero.
   112  		*state &= ^uint32(r13)
   113  	}
   114  
   115  	return p
   116  }
   117  
   118  func ginsnop(pp *objw.Progs) *obj.Prog {
   119  	// This is a hardware nop (1-byte 0x90) instruction,
   120  	// even though we describe it as an explicit XCHGL here.
   121  	// Particularly, this does not zero the high 32 bits
   122  	// like typical *L opcodes.
   123  	// (gas assembles "xchg %eax,%eax" to 0x87 0xc0, which
   124  	// does zero the high 32 bits.)
   125  	p := pp.Prog(x86.AXCHGL)
   126  	p.From.Type = obj.TYPE_REG
   127  	p.From.Reg = x86.REG_AX
   128  	p.To.Type = obj.TYPE_REG
   129  	p.To.Reg = x86.REG_AX
   130  	return p
   131  }
   132
View as plain text