mkduff.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  //go:build ignore
     6  
     7  // runtime·duffzero is a Duff's device for zeroing memory.
     8  // The compiler jumps to computed addresses within
     9  // the routine to zero chunks of memory.
    10  // Do not change duffzero without also
    11  // changing the uses in cmd/compile/internal/*/*.go.
    12  
    13  // runtime·duffcopy is a Duff's device for copying memory.
    14  // The compiler jumps to computed addresses within
    15  // the routine to copy chunks of memory.
    16  // Source and destination must not overlap.
    17  // Do not change duffcopy without also
    18  // changing the uses in cmd/compile/internal/*/*.go.
    19  
    20  // See the zero* and copy* generators below
    21  // for architecture-specific comments.
    22  
    23  // mkduff generates duff_*.s.
    24  package main
    25  
    26  import (
    27  	"bytes"
    28  	"fmt"
    29  	"io"
    30  	"log"
    31  	"os"
    32  )
    33  
    34  func main() {
    35  	gen("amd64", notags, zeroAMD64, copyAMD64)
    36  	gen("386", notags, zero386, copy386)
    37  	gen("arm", notags, zeroARM, copyARM)
    38  	gen("arm64", notags, zeroARM64, copyARM64)
    39  	gen("loong64", notags, zeroLOONG64, copyLOONG64)
    40  	gen("ppc64x", tagsPPC64x, zeroPPC64x, copyPPC64x)
    41  	gen("mips64x", tagsMIPS64x, zeroMIPS64x, copyMIPS64x)
    42  	gen("riscv64", notags, zeroRISCV64, copyRISCV64)
    43  }
    44  
    45  func gen(arch string, tags, zero, copy func(io.Writer)) {
    46  	var buf bytes.Buffer
    47  
    48  	fmt.Fprintln(&buf, "// Code generated by mkduff.go; DO NOT EDIT.")
    49  	fmt.Fprintln(&buf, "// Run go generate from src/runtime to update.")
    50  	fmt.Fprintln(&buf, "// See mkduff.go for comments.")
    51  	tags(&buf)
    52  	fmt.Fprintln(&buf, "#include \"textflag.h\"")
    53  	fmt.Fprintln(&buf)
    54  	zero(&buf)
    55  	fmt.Fprintln(&buf)
    56  	copy(&buf)
    57  
    58  	if err := os.WriteFile("duff_"+arch+".s", buf.Bytes(), 0644); err != nil {
    59  		log.Fatalln(err)
    60  	}
    61  }
    62  
    63  func notags(w io.Writer) { fmt.Fprintln(w) }
    64  
    65  func zeroAMD64(w io.Writer) {
    66  	// X15: zero
    67  	// DI: ptr to memory to be zeroed
    68  	// DI is updated as a side effect.
    69  	fmt.Fprintln(w, "TEXT runtime·duffzero<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
    70  	for i := 0; i < 16; i++ {
    71  		fmt.Fprintln(w, "\tMOVUPS\tX15,(DI)")
    72  		fmt.Fprintln(w, "\tMOVUPS\tX15,16(DI)")
    73  		fmt.Fprintln(w, "\tMOVUPS\tX15,32(DI)")
    74  		fmt.Fprintln(w, "\tMOVUPS\tX15,48(DI)")
    75  		fmt.Fprintln(w, "\tLEAQ\t64(DI),DI") // We use lea instead of add, to avoid clobbering flags
    76  		fmt.Fprintln(w)
    77  	}
    78  	fmt.Fprintln(w, "\tRET")
    79  }
    80  
    81  func copyAMD64(w io.Writer) {
    82  	// SI: ptr to source memory
    83  	// DI: ptr to destination memory
    84  	// SI and DI are updated as a side effect.
    85  	//
    86  	// This is equivalent to a sequence of MOVSQ but
    87  	// for some reason that is 3.5x slower than this code.
    88  	fmt.Fprintln(w, "TEXT runtime·duffcopy<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
    89  	for i := 0; i < 64; i++ {
    90  		fmt.Fprintln(w, "\tMOVUPS\t(SI), X0")
    91  		fmt.Fprintln(w, "\tADDQ\t$16, SI")
    92  		fmt.Fprintln(w, "\tMOVUPS\tX0, (DI)")
    93  		fmt.Fprintln(w, "\tADDQ\t$16, DI")
    94  		fmt.Fprintln(w)
    95  	}
    96  	fmt.Fprintln(w, "\tRET")
    97  }
    98  
    99  func zero386(w io.Writer) {
   100  	// AX: zero
   101  	// DI: ptr to memory to be zeroed
   102  	// DI is updated as a side effect.
   103  	fmt.Fprintln(w, "TEXT runtime·duffzero(SB), NOSPLIT, $0-0")
   104  	for i := 0; i < 128; i++ {
   105  		fmt.Fprintln(w, "\tSTOSL")
   106  	}
   107  	fmt.Fprintln(w, "\tRET")
   108  }
   109  
   110  func copy386(w io.Writer) {
   111  	// SI: ptr to source memory
   112  	// DI: ptr to destination memory
   113  	// SI and DI are updated as a side effect.
   114  	//
   115  	// This is equivalent to a sequence of MOVSL but
   116  	// for some reason MOVSL is really slow.
   117  	fmt.Fprintln(w, "TEXT runtime·duffcopy(SB), NOSPLIT, $0-0")
   118  	for i := 0; i < 128; i++ {
   119  		fmt.Fprintln(w, "\tMOVL\t(SI), CX")
   120  		fmt.Fprintln(w, "\tADDL\t$4, SI")
   121  		fmt.Fprintln(w, "\tMOVL\tCX, (DI)")
   122  		fmt.Fprintln(w, "\tADDL\t$4, DI")
   123  		fmt.Fprintln(w)
   124  	}
   125  	fmt.Fprintln(w, "\tRET")
   126  }
   127  
   128  func zeroARM(w io.Writer) {
   129  	// R0: zero
   130  	// R1: ptr to memory to be zeroed
   131  	// R1 is updated as a side effect.
   132  	fmt.Fprintln(w, "TEXT runtime·duffzero(SB), NOSPLIT, $0-0")
   133  	for i := 0; i < 128; i++ {
   134  		fmt.Fprintln(w, "\tMOVW.P\tR0, 4(R1)")
   135  	}
   136  	fmt.Fprintln(w, "\tRET")
   137  }
   138  
   139  func copyARM(w io.Writer) {
   140  	// R0: scratch space
   141  	// R1: ptr to source memory
   142  	// R2: ptr to destination memory
   143  	// R1 and R2 are updated as a side effect
   144  	fmt.Fprintln(w, "TEXT runtime·duffcopy(SB), NOSPLIT, $0-0")
   145  	for i := 0; i < 128; i++ {
   146  		fmt.Fprintln(w, "\tMOVW.P\t4(R1), R0")
   147  		fmt.Fprintln(w, "\tMOVW.P\tR0, 4(R2)")
   148  		fmt.Fprintln(w)
   149  	}
   150  	fmt.Fprintln(w, "\tRET")
   151  }
   152  
   153  func zeroARM64(w io.Writer) {
   154  	// ZR: always zero
   155  	// R20: ptr to memory to be zeroed
   156  	// On return, R20 points to the last zeroed dword.
   157  	fmt.Fprintln(w, "TEXT runtime·duffzero<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   158  	for i := 0; i < 63; i++ {
   159  		fmt.Fprintln(w, "\tSTP.P\t(ZR, ZR), 16(R20)")
   160  	}
   161  	fmt.Fprintln(w, "\tSTP\t(ZR, ZR), (R20)")
   162  	fmt.Fprintln(w, "\tRET")
   163  }
   164  
   165  func copyARM64(w io.Writer) {
   166  	// R20: ptr to source memory
   167  	// R21: ptr to destination memory
   168  	// R26, R27 (aka REGTMP): scratch space
   169  	// R20 and R21 are updated as a side effect
   170  	fmt.Fprintln(w, "TEXT runtime·duffcopy<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   171  
   172  	for i := 0; i < 64; i++ {
   173  		fmt.Fprintln(w, "\tLDP.P\t16(R20), (R26, R27)")
   174  		fmt.Fprintln(w, "\tSTP.P\t(R26, R27), 16(R21)")
   175  		fmt.Fprintln(w)
   176  	}
   177  	fmt.Fprintln(w, "\tRET")
   178  }
   179  
   180  func zeroLOONG64(w io.Writer) {
   181  	// R0: always zero
   182  	// R19 (aka REGRT1): ptr to memory to be zeroed
   183  	// On return, R19 points to the last zeroed dword.
   184  	fmt.Fprintln(w, "TEXT runtime·duffzero<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   185  	for i := 0; i < 128; i++ {
   186  		fmt.Fprintln(w, "\tMOVV\tR0, (R20)")
   187  		fmt.Fprintln(w, "\tADDV\t$8, R20")
   188  	}
   189  	fmt.Fprintln(w, "\tRET")
   190  }
   191  
   192  func copyLOONG64(w io.Writer) {
   193  	fmt.Fprintln(w, "TEXT runtime·duffcopy<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   194  	for i := 0; i < 128; i++ {
   195  		fmt.Fprintln(w, "\tMOVV\t(R20), R30")
   196  		fmt.Fprintln(w, "\tADDV\t$8, R20")
   197  		fmt.Fprintln(w, "\tMOVV\tR30, (R21)")
   198  		fmt.Fprintln(w, "\tADDV\t$8, R21")
   199  		fmt.Fprintln(w)
   200  	}
   201  	fmt.Fprintln(w, "\tRET")
   202  }
   203  
   204  func tagsPPC64x(w io.Writer) {
   205  	fmt.Fprintln(w)
   206  	fmt.Fprintln(w, "//go:build ppc64 || ppc64le")
   207  	fmt.Fprintln(w)
   208  }
   209  
   210  func zeroPPC64x(w io.Writer) {
   211  	// R0: always zero
   212  	// R3 (aka REGRT1): ptr to memory to be zeroed - 8
   213  	// On return, R3 points to the last zeroed dword.
   214  	fmt.Fprintln(w, "TEXT runtime·duffzero<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   215  	for i := 0; i < 128; i++ {
   216  		fmt.Fprintln(w, "\tMOVDU\tR0, 8(R20)")
   217  	}
   218  	fmt.Fprintln(w, "\tRET")
   219  }
   220  
   221  func copyPPC64x(w io.Writer) {
   222  	// duffcopy is not used on PPC64.
   223  	fmt.Fprintln(w, "TEXT runtime·duffcopy<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   224  	for i := 0; i < 128; i++ {
   225  		fmt.Fprintln(w, "\tMOVDU\t8(R20), R5")
   226  		fmt.Fprintln(w, "\tMOVDU\tR5, 8(R21)")
   227  	}
   228  	fmt.Fprintln(w, "\tRET")
   229  }
   230  
   231  func tagsMIPS64x(w io.Writer) {
   232  	fmt.Fprintln(w)
   233  	fmt.Fprintln(w, "//go:build mips64 || mips64le")
   234  	fmt.Fprintln(w)
   235  }
   236  
   237  func zeroMIPS64x(w io.Writer) {
   238  	// R0: always zero
   239  	// R1 (aka REGRT1): ptr to memory to be zeroed - 8
   240  	// On return, R1 points to the last zeroed dword.
   241  	fmt.Fprintln(w, "TEXT runtime·duffzero(SB), NOSPLIT|NOFRAME, $0-0")
   242  	for i := 0; i < 128; i++ {
   243  		fmt.Fprintln(w, "\tMOVV\tR0, 8(R1)")
   244  		fmt.Fprintln(w, "\tADDV\t$8, R1")
   245  	}
   246  	fmt.Fprintln(w, "\tRET")
   247  }
   248  
   249  func copyMIPS64x(w io.Writer) {
   250  	fmt.Fprintln(w, "TEXT runtime·duffcopy(SB), NOSPLIT|NOFRAME, $0-0")
   251  	for i := 0; i < 128; i++ {
   252  		fmt.Fprintln(w, "\tMOVV\t(R1), R23")
   253  		fmt.Fprintln(w, "\tADDV\t$8, R1")
   254  		fmt.Fprintln(w, "\tMOVV\tR23, (R2)")
   255  		fmt.Fprintln(w, "\tADDV\t$8, R2")
   256  		fmt.Fprintln(w)
   257  	}
   258  	fmt.Fprintln(w, "\tRET")
   259  }
   260  
   261  func zeroRISCV64(w io.Writer) {
   262  	// ZERO: always zero
   263  	// X25: ptr to memory to be zeroed
   264  	// X25 is updated as a side effect.
   265  	fmt.Fprintln(w, "TEXT runtime·duffzero<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   266  	for i := 0; i < 128; i++ {
   267  		fmt.Fprintln(w, "\tMOV\tZERO, (X25)")
   268  		fmt.Fprintln(w, "\tADD\t$8, X25")
   269  	}
   270  	fmt.Fprintln(w, "\tRET")
   271  }
   272  
   273  func copyRISCV64(w io.Writer) {
   274  	// X24: ptr to source memory
   275  	// X25: ptr to destination memory
   276  	// X24 and X25 are updated as a side effect
   277  	fmt.Fprintln(w, "TEXT runtime·duffcopy<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   278  	for i := 0; i < 128; i++ {
   279  		fmt.Fprintln(w, "\tMOV\t(X24), X31")
   280  		fmt.Fprintln(w, "\tADD\t$8, X24")
   281  		fmt.Fprintln(w, "\tMOV\tX31, (X25)")
   282  		fmt.Fprintln(w, "\tADD\t$8, X25")
   283  		fmt.Fprintln(w)
   284  	}
   285  	fmt.Fprintln(w, "\tRET")
   286  }
   287
View as plain text