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 5package amd64 6 7import ( 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 "internal/buildcfg" 14) 15 16// no floating point in note handlers on Plan 9 17var isPlan9 = buildcfg.GOOS == "plan9" 18 19// DUFFZERO consists of repeated blocks of 4 MOVUPSs + LEAQ, 20// See runtime/mkduff.go. 21const ( 22 dzBlocks = 16 // number of MOV/ADD blocks 23 dzBlockLen = 4 // number of clears per block 24 dzBlockSize = 23 // size of instructions in a single block 25 dzMovSize = 5 // size of single MOV instruction w/ offset 26 dzLeaqSize = 4 // size of single LEAQ instruction 27 dzClearStep = 16 // number of bytes cleared by each MOV instruction 28 29 dzClearLen = dzClearStep * dzBlockLen // bytes cleared by one block 30 dzSize = dzBlocks * dzBlockSize 31) 32 33// dzOff returns the offset for a jump into DUFFZERO. 34// b is the number of bytes to zero. 35func dzOff(b int64) int64 { 36 off := int64(dzSize) 37 off -= b / dzClearLen * dzBlockSize 38 tailLen := b % dzClearLen 39 if tailLen >= dzClearStep { 40 off -= dzLeaqSize + dzMovSize*(tailLen/dzClearStep) 41 } 42 return off 43} 44 45// duffzeroDI returns the pre-adjustment to DI for a call to DUFFZERO. 46// b is the number of bytes to zero. 47func dzDI(b int64) int64 { 48 tailLen := b % dzClearLen 49 if tailLen < dzClearStep { 50 return 0 51 } 52 tailSteps := tailLen / dzClearStep 53 return -dzClearStep * (dzBlockLen - tailSteps) 54} 55 56func zerorange(pp *objw.Progs, p *obj.Prog, off, cnt int64, state *uint32) *obj.Prog { 57 const ( 58 r13 = 1 << iota // if R13 is already zeroed. 59 ) 60 61 if cnt == 0 { 62 return p 63 } 64 65 if cnt == 8 { 66 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_X15, 0, obj.TYPE_MEM, x86.REG_SP, off) 67 } else if !isPlan9 && cnt <= int64(8*types.RegSize) { 68 for i := int64(0); i < cnt/16; i++ { 69 p = pp.Append(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X15, 0, obj.TYPE_MEM, x86.REG_SP, off+i*16) 70 } 71 72 if cnt%16 != 0 { 73 p = pp.Append(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X15, 0, obj.TYPE_MEM, x86.REG_SP, off+cnt-int64(16)) 74 } 75 } else if !isPlan9 && (cnt <= int64(128*types.RegSize)) { 76 // Save DI to r12. With the amd64 Go register abi, DI can contain 77 // an incoming parameter, whereas R12 is always scratch. 78 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_DI, 0, obj.TYPE_REG, x86.REG_R12, 0) 79 // Emit duffzero call 80 p = pp.Append(p, leaptr, obj.TYPE_MEM, x86.REG_SP, off+dzDI(cnt), obj.TYPE_REG, x86.REG_DI, 0) 81 p = pp.Append(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_ADDR, 0, dzOff(cnt)) 82 p.To.Sym = ir.Syms.Duffzero 83 if cnt%16 != 0 { 84 p = pp.Append(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X15, 0, obj.TYPE_MEM, x86.REG_DI, -int64(8)) 85 } 86 // Restore DI from r12 87 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_R12, 0, obj.TYPE_REG, x86.REG_DI, 0) 88 89 } else { 90 // When the register ABI is in effect, at this point in the 91 // prolog we may have live values in all of RAX,RDI,RCX. Save 92 // them off to registers before the REPSTOSQ below, then 93 // restore. Note that R12 and R13 are always available as 94 // scratch regs; here we also use R15 (this is safe to do 95 // since there won't be any globals accessed in the prolog). 96 // See rewriteToUseGot() in obj6.go for more on r15 use. 97 98 // Save rax/rdi/rcx 99 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_DI, 0, obj.TYPE_REG, x86.REG_R12, 0) 100 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_REG, x86.REG_R13, 0) 101 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_CX, 0, obj.TYPE_REG, x86.REG_R15, 0) 102 103 // Set up the REPSTOSQ and kick it off. 104 p = pp.Append(p, x86.AXORL, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_REG, x86.REG_AX, 0) 105 p = pp.Append(p, x86.AMOVQ, obj.TYPE_CONST, 0, cnt/int64(types.RegSize), obj.TYPE_REG, x86.REG_CX, 0) 106 p = pp.Append(p, leaptr, obj.TYPE_MEM, x86.REG_SP, off, obj.TYPE_REG, x86.REG_DI, 0) 107 p = pp.Append(p, x86.AREP, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0) 108 p = pp.Append(p, x86.ASTOSQ, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0) 109 110 // Restore rax/rdi/rcx 111 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_R12, 0, obj.TYPE_REG, x86.REG_DI, 0) 112 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_R13, 0, obj.TYPE_REG, x86.REG_AX, 0) 113 p = pp.Append(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_R15, 0, obj.TYPE_REG, x86.REG_CX, 0) 114 115 // Record the fact that r13 is no longer zero. 116 *state &= ^uint32(r13) 117 } 118 119 return p 120} 121 122func ginsnop(pp *objw.Progs) *obj.Prog { 123 // This is a hardware nop (1-byte 0x90) instruction, 124 // even though we describe it as an explicit XCHGL here. 125 // Particularly, this does not zero the high 32 bits 126 // like typical *L opcodes. 127 // (gas assembles "xchg %eax,%eax" to 0x87 0xc0, which 128 // does zero the high 32 bits.) 129 p := pp.Prog(x86.AXCHGL) 130 p.From.Type = obj.TYPE_REG 131 p.From.Reg = x86.REG_AX 132 p.To.Type = obj.TYPE_REG 133 p.To.Reg = x86.REG_AX 134 return p 135} 136