int: optimize performance by avoiding allocation

This change defines a new representation for Int on 64-bit machines
running a POSIX operating system, by reserving a 4GB portion of the
address space.

Pointers to addresses in this region represent int32 values,
and are disjoint from all *big.Int pointers,
allowing us to represent the union in a single pointer.
This means the conversion from Int to Value does not allocate.

The gauss benchmark (added in this CL) shows -40% ns, -48% bytes, -63% allocs:
Benchmark/bench_gauss-12  	      84	  13648744 ns/op	 3175816 B/op	  105862 allocs/op (before)
Benchmark/bench_gauss-12  	      55	  24283703 ns/op	 6119844 B/op	  289862 allocs/op (after)

On 32-but machines, or those running a non-POSIX system,
we continue to use the old representation.

Change-Id: Ib3b116760c5c9c0c8096314c90e68ddcb4b89bc0

.

Change-Id: I512065d0f09c10cf55b87d28673282c078ae2e3c

go.mod

@@ -6,5 +6,5 @@ require (
 	github.com/chzyer/logex v1.1.10 // indirect
 	github.com/chzyer/readline v0.0.0-20180603132655-2972be24d48e
 	github.com/chzyer/test v0.0.0-20180213035817-a1ea475d72b1 // indirect
-	golang.org/x/sys v0.0.0-20191002063906-3421d5a6bb1c // indirect
+	golang.org/x/sys v0.0.0-20191002063906-3421d5a6bb1c
 )

starlark/int.go

@@ -14,30 +14,9 @@ import (
 )
 
 // Int is the type of a Starlark int.
-type Int struct {
-	// We use only the signed 32 bit range of small to ensure
-	// that small+small and small*small do not overflow.
-	small_ int64    // minint32 <= small <= maxint32
-	big_   *big.Int // big != nil <=> value is not representable as int32
-}
-
-// --- low-level accessors ---
-
-// get returns the small and big components of the Int.
-// small is defined only if big is nil.
-func (i Int) get() (small int64, big *big.Int) {
-	return i.small_, i.big_
-}
-
-// Precondition: math.MinInt32 <= x && x <= math.MaxInt32
-func makeSmallInt(x int64) Int {
-	return Int{small_: x}
-}
-
-// Precondition: x cannot be represented as int32.
-func makeBigInt(x *big.Int) Int {
-	return Int{big_: x}
-}
+//
+// The zero value is not a legal value; use MakeInt(0).
+type Int struct{ impl intImpl }
 
 // --- high-level accessors ---
 

starlark/int_generic.go

@@ -0,0 +1,33 @@
+//+build !linux,!darwin
+//+build !amd64,!arm64,!mips64x,!ppc64x
+
+package starlark
+
+// generic Int implementation
+
+import "math/big"
+
+type intImpl struct {
+	// We use only the signed 32-bit range of small to ensure
+	// that small+small and small*small do not overflow.
+	small_ int64    // minint32 <= small <= maxint32
+	big_   *big.Int // big != nil <=> value is not representable as int32
+}
+
+// --- low-level accessors ---
+
+// get returns the small and big components of the Int.
+// small is defined only if big is nil.
+func (i Int) get() (small int64, big *big.Int) {
+	return i.small_, i.big_
+}
+
+// Precondition: math.MinInt32 <= x && x <= math.MaxInt32
+func makeSmallInt(x int64) Int {
+	return Int{intImpl{small_: x}}
+}
+
+// Precondition: x cannot be represented as int32.
+func makeBigInt(x *big.Int) Int {
+	return Int{intImpl{big_: x}}
+}

starlark/int_posix64.go

@@ -0,0 +1,55 @@
+//+build linux darwin
+//+build amd64 arm64 mips64x ppc64x
+
+package starlark
+
+// Optimized Int implementation for 64-bit machines running POSIX (for mmap).
+// It reserves a portion of the address space to represent int32 values.
+
+import (
+	"log"
+	"math"
+	"math/big"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// intImpl represents a union of (int32, *big.Int) in a single pointer,
+// so that Int-to-Value conversions need not allocate.
+//
+// The pointer is either a *big.Int, if the value is big, or a pointer into a
+// reserved portion of the address space (smallints), if the value is small.
+//
+// See int_generic.go for the basic representation concepts.
+type intImpl unsafe.Pointer
+
+// get returns the (small, big) arms of the union.
+func (i Int) get() (int64, *big.Int) {
+	ptr := uintptr(i.impl)
+	if ptr >= smallints && ptr < smallints+1<<32 {
+		return math.MinInt32 + int64(ptr-smallints), nil
+	}
+	return 0, (*big.Int)(i.impl)
+}
+
+// Precondition: math.MinInt32 <= x && x <= math.MaxInt32
+func makeSmallInt(x int64) Int {
+	return Int{intImpl(uintptr(x-math.MinInt32) + smallints)}
+}
+
+// Precondition: x cannot be represented as int32.
+func makeBigInt(x *big.Int) Int { return Int{intImpl(x)} }
+
+// smallints is the base address of a 2^32 byte memory region.
+// Pointers to addresses in this region represent int32 values.
+// We assume smallints is not at the very top of the address space.
+var smallints = reserveAddresses(1 << 32)
+
+func reserveAddresses(len int) uintptr {
+	b, err := unix.Mmap(-1, 0, len, unix.PROT_READ, unix.MAP_PRIVATE|unix.MAP_ANONYMOUS)
+	if err != nil {
+		log.Fatalf("mmap: %v", err)
+	}
+	return uintptr(unsafe.Pointer(&b[0]))
+}

starlark/testdata/benchmark.star

@@ -32,6 +32,12 @@ def bench_int():
         a += 1
 
 def bench_bigint():
-    a = 1 << 31 # maxint32 + 1
+    a = 1 << 31  # maxint32 + 1
     for _ in range1000:
         a += 1
+
+def bench_gauss():
+    # Sum of arithmetic series. All results fit in int32.
+    acc = 0
+    for x in range(92000):
+        acc += x