math.go

     1  // asmcheck
     2  
     3  // Copyright 2018 The Go Authors. All rights reserved.
     4  // Use of this source code is governed by a BSD-style
     5  // license that can be found in the LICENSE file.
     6  
     7  package codegen
     8  
     9  import "math"
    10  
    11  var sink64 [8]float64
    12  
    13  func approx(x float64) {
    14  	// amd64/v2:-".*x86HasSSE41" amd64/v3:-".*x86HasSSE41"
    15  	// amd64:"ROUNDSD\t[$]2"
    16  	// s390x:"FIDBR\t[$]6"
    17  	// arm64:"FRINTPD"
    18  	// ppc64x:"FRIP"
    19  	// wasm:"F64Ceil"
    20  	sink64[0] = math.Ceil(x)
    21  
    22  	// amd64/v2:-".*x86HasSSE41" amd64/v3:-".*x86HasSSE41"
    23  	// amd64:"ROUNDSD\t[$]1"
    24  	// s390x:"FIDBR\t[$]7"
    25  	// arm64:"FRINTMD"
    26  	// ppc64x:"FRIM"
    27  	// wasm:"F64Floor"
    28  	sink64[1] = math.Floor(x)
    29  
    30  	// s390x:"FIDBR\t[$]1"
    31  	// arm64:"FRINTAD"
    32  	// ppc64x:"FRIN"
    33  	sink64[2] = math.Round(x)
    34  
    35  	// amd64/v2:-".*x86HasSSE41" amd64/v3:-".*x86HasSSE41"
    36  	// amd64:"ROUNDSD\t[$]3"
    37  	// s390x:"FIDBR\t[$]5"
    38  	// arm64:"FRINTZD"
    39  	// ppc64x:"FRIZ"
    40  	// wasm:"F64Trunc"
    41  	sink64[3] = math.Trunc(x)
    42  
    43  	// amd64/v2:-".*x86HasSSE41" amd64/v3:-".*x86HasSSE41"
    44  	// amd64:"ROUNDSD\t[$]0"
    45  	// s390x:"FIDBR\t[$]4"
    46  	// arm64:"FRINTND"
    47  	// wasm:"F64Nearest"
    48  	sink64[4] = math.RoundToEven(x)
    49  }
    50  
    51  func sqrt(x float64) float64 {
    52  	// amd64:"SQRTSD"
    53  	// 386/sse2:"SQRTSD" 386/softfloat:-"SQRTD"
    54  	// arm64:"FSQRTD"
    55  	// arm/7:"SQRTD"
    56  	// mips/hardfloat:"SQRTD" mips/softfloat:-"SQRTD"
    57  	// mips64/hardfloat:"SQRTD" mips64/softfloat:-"SQRTD"
    58  	// wasm:"F64Sqrt"
    59  	// ppc64x:"FSQRT"
    60  	// riscv64: "FSQRTD"
    61  	return math.Sqrt(x)
    62  }
    63  
    64  func sqrt32(x float32) float32 {
    65  	// amd64:"SQRTSS"
    66  	// 386/sse2:"SQRTSS" 386/softfloat:-"SQRTS"
    67  	// arm64:"FSQRTS"
    68  	// arm/7:"SQRTF"
    69  	// mips/hardfloat:"SQRTF" mips/softfloat:-"SQRTF"
    70  	// mips64/hardfloat:"SQRTF" mips64/softfloat:-"SQRTF"
    71  	// wasm:"F32Sqrt"
    72  	// ppc64x:"FSQRTS"
    73  	// riscv64: "FSQRTS"
    74  	return float32(math.Sqrt(float64(x)))
    75  }
    76  
    77  // Check that it's using integer registers
    78  func abs(x, y float64) {
    79  	// amd64:"BTRQ\t[$]63"
    80  	// arm64:"FABSD\t"
    81  	// s390x:"LPDFR\t",-"MOVD\t"     (no integer load/store)
    82  	// ppc64x:"FABS\t"
    83  	// riscv64:"FABSD\t"
    84  	// wasm:"F64Abs"
    85  	// arm/6:"ABSD\t"
    86  	// mips64/hardfloat:"ABSD\t"
    87  	// mips/hardfloat:"ABSD\t"
    88  	sink64[0] = math.Abs(x)
    89  
    90  	// amd64:"BTRQ\t[$]63","PXOR"    (TODO: this should be BTSQ)
    91  	// s390x:"LNDFR\t",-"MOVD\t"     (no integer load/store)
    92  	// ppc64x:"FNABS\t"
    93  	sink64[1] = -math.Abs(y)
    94  }
    95  
    96  // Check that it's using integer registers
    97  func abs32(x float32) float32 {
    98  	// s390x:"LPDFR",-"LDEBR",-"LEDBR"     (no float64 conversion)
    99  	return float32(math.Abs(float64(x)))
   100  }
   101  
   102  // Check that it's using integer registers
   103  func copysign(a, b, c float64) {
   104  	// amd64:"BTRQ\t[$]63","ANDQ","ORQ"
   105  	// s390x:"CPSDR",-"MOVD"         (no integer load/store)
   106  	// ppc64x:"FCPSGN"
   107  	// riscv64:"FSGNJD"
   108  	// wasm:"F64Copysign"
   109  	sink64[0] = math.Copysign(a, b)
   110  
   111  	// amd64:"BTSQ\t[$]63"
   112  	// s390x:"LNDFR\t",-"MOVD\t"     (no integer load/store)
   113  	// ppc64x:"FCPSGN"
   114  	// riscv64:"FSGNJD"
   115  	// arm64:"ORR", -"AND"
   116  	sink64[1] = math.Copysign(c, -1)
   117  
   118  	// Like math.Copysign(c, -1), but with integer operations. Useful
   119  	// for platforms that have a copysign opcode to see if it's detected.
   120  	// s390x:"LNDFR\t",-"MOVD\t"     (no integer load/store)
   121  	sink64[2] = math.Float64frombits(math.Float64bits(a) | 1<<63)
   122  
   123  	// amd64:"ANDQ","ORQ"
   124  	// s390x:"CPSDR\t",-"MOVD\t"     (no integer load/store)
   125  	// ppc64x:"FCPSGN"
   126  	// riscv64:"FSGNJD"
   127  	sink64[3] = math.Copysign(-1, c)
   128  }
   129  
   130  func fma(x, y, z float64) float64 {
   131  	// amd64/v3:-".*x86HasFMA"
   132  	// amd64:"VFMADD231SD"
   133  	// arm/6:"FMULAD"
   134  	// arm64:"FMADDD"
   135  	// s390x:"FMADD"
   136  	// ppc64x:"FMADD"
   137  	// riscv64:"FMADDD"
   138  	return math.FMA(x, y, z)
   139  }
   140  
   141  func fms(x, y, z float64) float64 {
   142  	// riscv64:"FMSUBD"
   143  	return math.FMA(x, y, -z)
   144  }
   145  
   146  func fnms(x, y, z float64) float64 {
   147  	// riscv64:"FNMSUBD",-"FNMADDD"
   148  	return math.FMA(-x, y, z)
   149  }
   150  
   151  func fnma(x, y, z float64) float64 {
   152  	// riscv64:"FNMADDD",-"FNMSUBD"
   153  	return math.FMA(x, -y, -z)
   154  }
   155  
   156  func fromFloat64(f64 float64) uint64 {
   157  	// amd64:"MOVQ\tX.*, [^X].*"
   158  	// arm64:"FMOVD\tF.*, R.*"
   159  	// ppc64x:"MFVSRD"
   160  	// mips64/hardfloat:"MOVV\tF.*, R.*"
   161  	return math.Float64bits(f64+1) + 1
   162  }
   163  
   164  func fromFloat32(f32 float32) uint32 {
   165  	// amd64:"MOVL\tX.*, [^X].*"
   166  	// arm64:"FMOVS\tF.*, R.*"
   167  	// mips64/hardfloat:"MOVW\tF.*, R.*"
   168  	return math.Float32bits(f32+1) + 1
   169  }
   170  
   171  func toFloat64(u64 uint64) float64 {
   172  	// amd64:"MOVQ\t[^X].*, X.*"
   173  	// arm64:"FMOVD\tR.*, F.*"
   174  	// ppc64x:"MTVSRD"
   175  	// mips64/hardfloat:"MOVV\tR.*, F.*"
   176  	return math.Float64frombits(u64+1) + 1
   177  }
   178  
   179  func toFloat32(u32 uint32) float32 {
   180  	// amd64:"MOVL\t[^X].*, X.*"
   181  	// arm64:"FMOVS\tR.*, F.*"
   182  	// mips64/hardfloat:"MOVW\tR.*, F.*"
   183  	return math.Float32frombits(u32+1) + 1
   184  }
   185  
   186  // Test that comparisons with constants converted to float
   187  // are evaluated at compile-time
   188  
   189  func constantCheck64() bool {
   190  	// amd64:"(MOVB\t[$]0)|(XORL\t[A-Z][A-Z0-9]+, [A-Z][A-Z0-9]+)",-"FCMP",-"MOVB\t[$]1"
   191  	// s390x:"MOV(B|BZ|D)\t[$]0,",-"FCMPU",-"MOV(B|BZ|D)\t[$]1,"
   192  	return 0.5 == float64(uint32(1)) || 1.5 > float64(uint64(1<<63))
   193  }
   194  
   195  func constantCheck32() bool {
   196  	// amd64:"MOV(B|L)\t[$]1",-"FCMP",-"MOV(B|L)\t[$]0"
   197  	// s390x:"MOV(B|BZ|D)\t[$]1,",-"FCMPU",-"MOV(B|BZ|D)\t[$]0,"
   198  	return float32(0.5) <= float32(int64(1)) && float32(1.5) >= float32(int32(-1<<31))
   199  }
   200  
   201  // Test that integer constants are converted to floating point constants
   202  // at compile-time
   203  
   204  func constantConvert32(x float32) float32 {
   205  	// amd64:"MOVSS\t[$]f32.3f800000\\(SB\\)"
   206  	// s390x:"FMOVS\t[$]f32.3f800000\\(SB\\)"
   207  	// ppc64x:"FMOVS\t[$]f32.3f800000\\(SB\\)"
   208  	// arm64:"FMOVS\t[$]\\(1.0\\)"
   209  	if x > math.Float32frombits(0x3f800000) {
   210  		return -x
   211  	}
   212  	return x
   213  }
   214  
   215  func constantConvertInt32(x uint32) uint32 {
   216  	// amd64:-"MOVSS"
   217  	// s390x:-"FMOVS"
   218  	// ppc64x:-"FMOVS"
   219  	// arm64:-"FMOVS"
   220  	if x > math.Float32bits(1) {
   221  		return -x
   222  	}
   223  	return x
   224  }
   225  
   226  func nanGenerate64() float64 {
   227  	// Test to make sure we don't generate a NaN while constant propagating.
   228  	// See issue 36400.
   229  	zero := 0.0
   230  	// amd64:-"DIVSD"
   231  	inf := 1 / zero // +inf. We can constant propagate this one.
   232  	negone := -1.0
   233  
   234  	// amd64:"DIVSD"
   235  	z0 := zero / zero
   236  	// amd64:"MULSD"
   237  	z1 := zero * inf
   238  	// amd64:"SQRTSD"
   239  	z2 := math.Sqrt(negone)
   240  	return z0 + z1 + z2
   241  }
   242  
   243  func nanGenerate32() float32 {
   244  	zero := float32(0.0)
   245  	// amd64:-"DIVSS"
   246  	inf := 1 / zero // +inf. We can constant propagate this one.
   247  
   248  	// amd64:"DIVSS"
   249  	z0 := zero / zero
   250  	// amd64:"MULSS"
   251  	z1 := zero * inf
   252  	return z0 + z1
   253  }
   254
View as plain text