aboutsummaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
Diffstat (limited to 'src')
-rw-r--r--src/cmd/5g/gsubr.go3
-rw-r--r--src/cmd/5g/peep.go1
-rw-r--r--src/cmd/5g/prog.go21
-rw-r--r--src/cmd/6g/gsubr.go3
-rw-r--r--src/cmd/6g/prog.go1
-rw-r--r--src/cmd/internal/gc/cgen.go9
-rw-r--r--src/cmd/internal/gc/gen.go3
-rw-r--r--src/cmd/internal/gc/syntax.go3
-rw-r--r--src/cmd/internal/gc/walk.go10
-rw-r--r--src/math/all_test.go45
-rw-r--r--src/math/sqrt.go5
11 files changed, 91 insertions, 13 deletions
diff --git a/src/cmd/5g/gsubr.go b/src/cmd/5g/gsubr.go
index 0d22f74c9a..fe4ed8d1f2 100644
--- a/src/cmd/5g/gsubr.go
+++ b/src/cmd/5g/gsubr.go
@@ -1055,6 +1055,9 @@ func optoas(op int, t *gc.Type) int {
case gc.ODIV<<16 | gc.TFLOAT64:
a = arm.ADIVD
+
+ case gc.OSQRT<<16 | gc.TFLOAT64:
+ a = arm.ASQRTD
}
return a
diff --git a/src/cmd/5g/peep.go b/src/cmd/5g/peep.go
index 5305e4b7f6..9ec3be2eec 100644
--- a/src/cmd/5g/peep.go
+++ b/src/cmd/5g/peep.go
@@ -1101,6 +1101,7 @@ func copyu(p *obj.Prog, v *obj.Addr, s *obj.Addr) int {
return 0
case obj.ANOP, /* read,, write */
+ arm.ASQRTD,
arm.AMOVW,
arm.AMOVF,
arm.AMOVD,
diff --git a/src/cmd/5g/prog.go b/src/cmd/5g/prog.go
index bfb703e56f..c472cdf042 100644
--- a/src/cmd/5g/prog.go
+++ b/src/cmd/5g/prog.go
@@ -70,16 +70,17 @@ var progtable = [arm.ALAST]obj.ProgInfo{
arm.ATST: {gc.SizeL | gc.LeftRead | gc.RightRead, 0, 0, 0},
// Floating point.
- arm.AADDD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
- arm.AADDF: {gc.SizeF | gc.LeftRead | RightRdwr, 0, 0, 0},
- arm.ACMPD: {gc.SizeD | gc.LeftRead | gc.RightRead, 0, 0, 0},
- arm.ACMPF: {gc.SizeF | gc.LeftRead | gc.RightRead, 0, 0, 0},
- arm.ADIVD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
- arm.ADIVF: {gc.SizeF | gc.LeftRead | RightRdwr, 0, 0, 0},
- arm.AMULD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
- arm.AMULF: {gc.SizeF | gc.LeftRead | RightRdwr, 0, 0, 0},
- arm.ASUBD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
- arm.ASUBF: {gc.SizeF | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.AADDD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.AADDF: {gc.SizeF | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.ACMPD: {gc.SizeD | gc.LeftRead | gc.RightRead, 0, 0, 0},
+ arm.ACMPF: {gc.SizeF | gc.LeftRead | gc.RightRead, 0, 0, 0},
+ arm.ADIVD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.ADIVF: {gc.SizeF | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.AMULD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.AMULF: {gc.SizeF | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.ASUBD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.ASUBF: {gc.SizeF | gc.LeftRead | RightRdwr, 0, 0, 0},
+ arm.ASQRTD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
// Conversions.
arm.AMOVWD: {gc.SizeD | gc.LeftRead | gc.RightWrite | gc.Conv, 0, 0, 0},
diff --git a/src/cmd/6g/gsubr.go b/src/cmd/6g/gsubr.go
index b2290af733..323ea69a98 100644
--- a/src/cmd/6g/gsubr.go
+++ b/src/cmd/6g/gsubr.go
@@ -1131,6 +1131,9 @@ func optoas(op int, t *gc.Type) int {
case gc.ODIV<<16 | gc.TFLOAT64:
a = x86.ADIVSD
+
+ case gc.OSQRT<<16 | gc.TFLOAT64:
+ a = x86.ASQRTSD
}
return a
diff --git a/src/cmd/6g/prog.go b/src/cmd/6g/prog.go
index 0644800257..fe9f013851 100644
--- a/src/cmd/6g/prog.go
+++ b/src/cmd/6g/prog.go
@@ -204,6 +204,7 @@ var progtable = [x86.ALAST]obj.ProgInfo{
x86.ASHRL: {gc.SizeL | gc.LeftRead | RightRdwr | gc.ShiftCX | gc.SetCarry, 0, 0, 0},
x86.ASHRQ: {gc.SizeQ | gc.LeftRead | RightRdwr | gc.ShiftCX | gc.SetCarry, 0, 0, 0},
x86.ASHRW: {gc.SizeW | gc.LeftRead | RightRdwr | gc.ShiftCX | gc.SetCarry, 0, 0, 0},
+ x86.ASQRTSD: {gc.SizeD | gc.LeftRead | RightRdwr, 0, 0, 0},
x86.ASTOSB: {gc.OK, AX | DI, DI, 0},
x86.ASTOSL: {gc.OK, AX | DI, DI, 0},
x86.ASTOSQ: {gc.OK, AX | DI, DI, 0},
diff --git a/src/cmd/internal/gc/cgen.go b/src/cmd/internal/gc/cgen.go
index b3524c26c4..d3921f7ece 100644
--- a/src/cmd/internal/gc/cgen.go
+++ b/src/cmd/internal/gc/cgen.go
@@ -409,6 +409,15 @@ func Cgen(n *Node, res *Node) {
cgen_norm(n, &n1, res)
return
+ case OSQRT:
+ var n1 Node
+ Regalloc(&n1, nl.Type, res)
+ Cgen(n.Left, &n1)
+ Thearch.Gins(Thearch.Optoas(OSQRT, nl.Type), &n1, &n1)
+ Thearch.Gmove(&n1, res)
+ Regfree(&n1)
+ return
+
// symmetric binary
case OAND,
OOR,
diff --git a/src/cmd/internal/gc/gen.go b/src/cmd/internal/gc/gen.go
index caae2f1ce1..e0659fc8a4 100644
--- a/src/cmd/internal/gc/gen.go
+++ b/src/cmd/internal/gc/gen.go
@@ -1002,6 +1002,9 @@ func gen(n *Node) {
case ORETURN, ORETJMP:
cgen_ret(n)
+ case OSQRT:
+ cgen_discard(n.Left)
+
case OCHECKNIL:
Cgen_checknil(n.Left)
diff --git a/src/cmd/internal/gc/syntax.go b/src/cmd/internal/gc/syntax.go
index 8f5b85db1f..671a624c1d 100644
--- a/src/cmd/internal/gc/syntax.go
+++ b/src/cmd/internal/gc/syntax.go
@@ -293,7 +293,7 @@ const (
OREGISTER // a register, such as AX.
OINDREG // offset plus indirect of a register, such as 8(SP).
- // 386/amd64-specific opcodes
+ // arch-specific opcodes
OCMP // compare: ACMP.
ODEC // decrement: ADEC.
OINC // increment: AINC.
@@ -303,6 +303,7 @@ const (
ORROTC // right rotate-carry: ARCR.
ORETJMP // return to other function
OPS // compare parity set (for x86 NaN check)
+ OSQRT // sqrt(float64), on systems that have hw support
OEND
)
diff --git a/src/cmd/internal/gc/walk.go b/src/cmd/internal/gc/walk.go
index c10201aa2e..a0a29d35ac 100644
--- a/src/cmd/internal/gc/walk.go
+++ b/src/cmd/internal/gc/walk.go
@@ -622,6 +622,16 @@ func walkexpr(np **Node, init **NodeList) {
walkexpr(&n.Left, init)
walkexprlist(n.List, init)
+ if n.Left.Op == ONAME && n.Left.Sym.Name == "Sqrt" && n.Left.Sym.Pkg.Path == "math" {
+ switch Thearch.Thechar {
+ case '5', '6':
+ n.Op = OSQRT
+ n.Left = n.List.N
+ n.List = nil
+ goto ret
+ }
+ }
+
ll := ascompatte(int(n.Op), n, n.Isddd, getinarg(t), n.List, 0, init)
n.List = reorder1(ll)
goto ret
diff --git a/src/math/all_test.go b/src/math/all_test.go
index c07ac740e3..84061be264 100644
--- a/src/math/all_test.go
+++ b/src/math/all_test.go
@@ -2977,15 +2977,56 @@ func BenchmarkSinh(b *testing.B) {
}
}
+var Global float64
+
func BenchmarkSqrt(b *testing.B) {
+ x, y := 0.0, 10.0
+ for i := 0; i < b.N; i++ {
+ x += Sqrt(y)
+ }
+ Global = x
+}
+
+func BenchmarkSqrtIndirect(b *testing.B) {
+ x, y := 0.0, 10.0
+ f := Sqrt
for i := 0; i < b.N; i++ {
- Sqrt(10)
+ x += f(y)
}
+ Global = x
}
func BenchmarkSqrtGo(b *testing.B) {
+ x, y := 0.0, 10.0
for i := 0; i < b.N; i++ {
- SqrtGo(10)
+ x += SqrtGo(y)
+ }
+ Global = x
+}
+
+func isPrime(i int) bool {
+ // Yes, this is a dumb way to write this code,
+ // but calling Sqrt repeatedly in this way demonstrates
+ // the benefit of using a direct SQRT instruction on systems
+ // that have one, whereas the obvious loop seems not to
+ // demonstrate such a benefit.
+ for j := 2; float64(j) <= Sqrt(float64(i)); j++ {
+ if i%j == 0 {
+ return false
+ }
+ }
+ return true
+}
+
+func BenchmarkSqrtPrime(b *testing.B) {
+ any := false
+ for i := 0; i < b.N; i++ {
+ if isPrime(100003) {
+ any = true
+ }
+ }
+ if any {
+ Global = 1
}
}
diff --git a/src/math/sqrt.go b/src/math/sqrt.go
index fdc869992e..23cf2996c2 100644
--- a/src/math/sqrt.go
+++ b/src/math/sqrt.go
@@ -91,6 +91,11 @@ package math
// Sqrt(NaN) = NaN
func Sqrt(x float64) float64
+// Note: Sqrt is implemented in assembly on some systems.
+// Others have assembly stubs that jump to func sqrt below.
+// On systems where Sqrt is a single instruction, the compiler
+// may turn a direct call into a direct use of that instruction instead.
+
func sqrt(x float64) float64 {
// special cases
switch {
generated by cgit v1.3-6-g1900 (git 2.53.0) at 2026-04-17 20:39:22 +0000