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("ppc64x", tagsPPC64x, zeroPPC64x, copyPPC64x)
    40  	gen("mips64x", tagsMIPS64x, zeroMIPS64x, copyMIPS64x)
    41  	gen("riscv64", notags, zeroRISCV64, copyRISCV64)
    42  }
    43  
    44  func gen(arch string, tags, zero, copy func(io.Writer)) {
    45  	var buf bytes.Buffer
    46  
    47  	fmt.Fprintln(&buf, "// Code generated by mkduff.go; DO NOT EDIT.")
    48  	fmt.Fprintln(&buf, "// Run go generate from src/runtime to update.")
    49  	fmt.Fprintln(&buf, "// See mkduff.go for comments.")
    50  	tags(&buf)
    51  	fmt.Fprintln(&buf, "#include \"textflag.h\"")
    52  	fmt.Fprintln(&buf)
    53  	zero(&buf)
    54  	fmt.Fprintln(&buf)
    55  	copy(&buf)
    56  
    57  	if err := os.WriteFile("duff_"+arch+".s", buf.Bytes(), 0644); err != nil {
    58  		log.Fatalln(err)
    59  	}
    60  }
    61  
    62  func notags(w io.Writer) { fmt.Fprintln(w) }
    63  
    64  func zeroAMD64(w io.Writer) {
    65  	// X15: zero
    66  	// DI: ptr to memory to be zeroed
    67  	// DI is updated as a side effect.
    68  	fmt.Fprintln(w, "TEXT runtime·duffzero<ABIInternal>(SB), NOSPLIT, $0-0")
    69  	for i := 0; i < 16; i++ {
    70  		fmt.Fprintln(w, "\tMOVUPS\tX15,(DI)")
    71  		fmt.Fprintln(w, "\tMOVUPS\tX15,16(DI)")
    72  		fmt.Fprintln(w, "\tMOVUPS\tX15,32(DI)")
    73  		fmt.Fprintln(w, "\tMOVUPS\tX15,48(DI)")
    74  		fmt.Fprintln(w, "\tLEAQ\t64(DI),DI") // We use lea instead of add, to avoid clobbering flags
    75  		fmt.Fprintln(w)
    76  	}
    77  	fmt.Fprintln(w, "\tRET")
    78  }
    79  
    80  func copyAMD64(w io.Writer) {
    81  	// SI: ptr to source memory
    82  	// DI: ptr to destination memory
    83  	// SI and DI are updated as a side effect.
    84  	//
    85  	// This is equivalent to a sequence of MOVSQ but
    86  	// for some reason that is 3.5x slower than this code.
    87  	fmt.Fprintln(w, "TEXT runtime·duffcopy<ABIInternal>(SB), NOSPLIT, $0-0")
    88  	for i := 0; i < 64; i++ {
    89  		fmt.Fprintln(w, "\tMOVUPS\t(SI), X0")
    90  		fmt.Fprintln(w, "\tADDQ\t$16, SI")
    91  		fmt.Fprintln(w, "\tMOVUPS\tX0, (DI)")
    92  		fmt.Fprintln(w, "\tADDQ\t$16, DI")
    93  		fmt.Fprintln(w)
    94  	}
    95  	fmt.Fprintln(w, "\tRET")
    96  }
    97  
    98  func zero386(w io.Writer) {
    99  	// AX: zero
   100  	// DI: ptr to memory to be zeroed
   101  	// DI is updated as a side effect.
   102  	fmt.Fprintln(w, "TEXT runtime·duffzero(SB), NOSPLIT, $0-0")
   103  	for i := 0; i < 128; i++ {
   104  		fmt.Fprintln(w, "\tSTOSL")
   105  	}
   106  	fmt.Fprintln(w, "\tRET")
   107  }
   108  
   109  func copy386(w io.Writer) {
   110  	// SI: ptr to source memory
   111  	// DI: ptr to destination memory
   112  	// SI and DI are updated as a side effect.
   113  	//
   114  	// This is equivalent to a sequence of MOVSL but
   115  	// for some reason MOVSL is really slow.
   116  	fmt.Fprintln(w, "TEXT runtime·duffcopy(SB), NOSPLIT, $0-0")
   117  	for i := 0; i < 128; i++ {
   118  		fmt.Fprintln(w, "\tMOVL\t(SI), CX")
   119  		fmt.Fprintln(w, "\tADDL\t$4, SI")
   120  		fmt.Fprintln(w, "\tMOVL\tCX, (DI)")
   121  		fmt.Fprintln(w, "\tADDL\t$4, DI")
   122  		fmt.Fprintln(w)
   123  	}
   124  	fmt.Fprintln(w, "\tRET")
   125  }
   126  
   127  func zeroARM(w io.Writer) {
   128  	// R0: zero
   129  	// R1: ptr to memory to be zeroed
   130  	// R1 is updated as a side effect.
   131  	fmt.Fprintln(w, "TEXT runtime·duffzero(SB), NOSPLIT, $0-0")
   132  	for i := 0; i < 128; i++ {
   133  		fmt.Fprintln(w, "\tMOVW.P\tR0, 4(R1)")
   134  	}
   135  	fmt.Fprintln(w, "\tRET")
   136  }
   137  
   138  func copyARM(w io.Writer) {
   139  	// R0: scratch space
   140  	// R1: ptr to source memory
   141  	// R2: ptr to destination memory
   142  	// R1 and R2 are updated as a side effect
   143  	fmt.Fprintln(w, "TEXT runtime·duffcopy(SB), NOSPLIT, $0-0")
   144  	for i := 0; i < 128; i++ {
   145  		fmt.Fprintln(w, "\tMOVW.P\t4(R1), R0")
   146  		fmt.Fprintln(w, "\tMOVW.P\tR0, 4(R2)")
   147  		fmt.Fprintln(w)
   148  	}
   149  	fmt.Fprintln(w, "\tRET")
   150  }
   151  
   152  func zeroARM64(w io.Writer) {
   153  	// ZR: always zero
   154  	// R20: ptr to memory to be zeroed
   155  	// On return, R20 points to the last zeroed dword.
   156  	fmt.Fprintln(w, "TEXT runtime·duffzero<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   157  	for i := 0; i < 63; i++ {
   158  		fmt.Fprintln(w, "\tSTP.P\t(ZR, ZR), 16(R20)")
   159  	}
   160  	fmt.Fprintln(w, "\tSTP\t(ZR, ZR), (R20)")
   161  	fmt.Fprintln(w, "\tRET")
   162  }
   163  
   164  func copyARM64(w io.Writer) {
   165  	// R20: ptr to source memory
   166  	// R21: ptr to destination memory
   167  	// R26, R27 (aka REGTMP): scratch space
   168  	// R20 and R21 are updated as a side effect
   169  	fmt.Fprintln(w, "TEXT runtime·duffcopy<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   170  
   171  	for i := 0; i < 64; i++ {
   172  		fmt.Fprintln(w, "\tLDP.P\t16(R20), (R26, R27)")
   173  		fmt.Fprintln(w, "\tSTP.P\t(R26, R27), 16(R21)")
   174  		fmt.Fprintln(w)
   175  	}
   176  	fmt.Fprintln(w, "\tRET")
   177  }
   178  
   179  func tagsPPC64x(w io.Writer) {
   180  	fmt.Fprintln(w)
   181  	fmt.Fprintln(w, "//go:build ppc64 || ppc64le")
   182  	fmt.Fprintln(w)
   183  }
   184  
   185  func zeroPPC64x(w io.Writer) {
   186  	// R0: always zero
   187  	// R3 (aka REGRT1): ptr to memory to be zeroed - 8
   188  	// On return, R3 points to the last zeroed dword.
   189  	fmt.Fprintln(w, "TEXT runtime·duffzero<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   190  	for i := 0; i < 128; i++ {
   191  		fmt.Fprintln(w, "\tMOVDU\tR0, 8(R20)")
   192  	}
   193  	fmt.Fprintln(w, "\tRET")
   194  }
   195  
   196  func copyPPC64x(w io.Writer) {
   197  	// duffcopy is not used on PPC64.
   198  	fmt.Fprintln(w, "TEXT runtime·duffcopy<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-0")
   199  	for i := 0; i < 128; i++ {
   200  		fmt.Fprintln(w, "\tMOVDU\t8(R20), R5")
   201  		fmt.Fprintln(w, "\tMOVDU\tR5, 8(R21)")
   202  	}
   203  	fmt.Fprintln(w, "\tRET")
   204  }
   205  
   206  func tagsMIPS64x(w io.Writer) {
   207  	fmt.Fprintln(w)
   208  	fmt.Fprintln(w, "//go:build mips64 || mips64le")
   209  	fmt.Fprintln(w)
   210  }
   211  
   212  func zeroMIPS64x(w io.Writer) {
   213  	// R0: always zero
   214  	// R1 (aka REGRT1): ptr to memory to be zeroed - 8
   215  	// On return, R1 points to the last zeroed dword.
   216  	fmt.Fprintln(w, "TEXT runtime·duffzero(SB), NOSPLIT|NOFRAME, $0-0")
   217  	for i := 0; i < 128; i++ {
   218  		fmt.Fprintln(w, "\tMOVV\tR0, 8(R1)")
   219  		fmt.Fprintln(w, "\tADDV\t$8, R1")
   220  	}
   221  	fmt.Fprintln(w, "\tRET")
   222  }
   223  
   224  func copyMIPS64x(w io.Writer) {
   225  	fmt.Fprintln(w, "TEXT runtime·duffcopy(SB), NOSPLIT|NOFRAME, $0-0")
   226  	for i := 0; i < 128; i++ {
   227  		fmt.Fprintln(w, "\tMOVV\t(R1), R23")
   228  		fmt.Fprintln(w, "\tADDV\t$8, R1")
   229  		fmt.Fprintln(w, "\tMOVV\tR23, (R2)")
   230  		fmt.Fprintln(w, "\tADDV\t$8, R2")
   231  		fmt.Fprintln(w)
   232  	}
   233  	fmt.Fprintln(w, "\tRET")
   234  }
   235  
   236  func zeroRISCV64(w io.Writer) {
   237  	// ZERO: always zero
   238  	// X10: ptr to memory to be zeroed
   239  	// X10 is updated as a side effect.
   240  	fmt.Fprintln(w, "TEXT runtime·duffzero(SB), NOSPLIT|NOFRAME, $0-0")
   241  	for i := 0; i < 128; i++ {
   242  		fmt.Fprintln(w, "\tMOV\tZERO, (X10)")
   243  		fmt.Fprintln(w, "\tADD\t$8, X10")
   244  	}
   245  	fmt.Fprintln(w, "\tRET")
   246  }
   247  
   248  func copyRISCV64(w io.Writer) {
   249  	// X10: ptr to source memory
   250  	// X11: ptr to destination memory
   251  	// X10 and X11 are updated as a side effect
   252  	fmt.Fprintln(w, "TEXT runtime·duffcopy(SB), NOSPLIT|NOFRAME, $0-0")
   253  	for i := 0; i < 128; i++ {
   254  		fmt.Fprintln(w, "\tMOV\t(X10), X31")
   255  		fmt.Fprintln(w, "\tADD\t$8, X10")
   256  		fmt.Fprintln(w, "\tMOV\tX31, (X11)")
   257  		fmt.Fprintln(w, "\tADD\t$8, X11")
   258  		fmt.Fprintln(w)
   259  	}
   260  	fmt.Fprintln(w, "\tRET")
   261  }
   262
View as plain text