Create inlined fast versions of integer ops

1 files changed, 116 insertions, 13 deletions

diff --git a/builtins/integers.h b/builtins/integers.h
index 218d8d23..f9051ff5 100644
--- a/builtins/integers.h
+++ b/builtins/integers.h
@@ -82,22 +82,125 @@ Int_t Int$from_num(double n);
 double Int$as_num(Int_t i);
 #define I(i) ((int64_t)(i) == (int32_t)(i) ? ((Int_t){.small=((uint64_t)(i)<<2)|1}) : Int$from_i64(i))
 
-Int_t Int$plus(Int_t x, Int_t y);
-Int_t Int$minus(Int_t x, Int_t y);
-Int_t Int$times(Int_t x, Int_t y);
-Int_t Int$divided_by(Int_t x, Int_t y);
-Int_t Int$modulo(Int_t x, Int_t y);
-Int_t Int$modulo1(Int_t x, Int_t y);
-Int_t Int$left_shifted(Int_t x, Int_t y);
-Int_t Int$right_shifted(Int_t x, Int_t y);
-Int_t Int$bit_and(Int_t x, Int_t y);
-Int_t Int$bit_or(Int_t x, Int_t y);
-Int_t Int$bit_xor(Int_t x, Int_t y);
-Int_t Int$negative(Int_t x);
-Int_t Int$negated(Int_t x);
+Int_t Int$slow_plus(Int_t x, Int_t y);
+Int_t Int$slow_minus(Int_t x, Int_t y);
+Int_t Int$slow_times(Int_t x, Int_t y);
+Int_t Int$slow_divided_by(Int_t x, Int_t y);
+Int_t Int$slow_modulo(Int_t x, Int_t y);
+Int_t Int$slow_modulo1(Int_t x, Int_t y);
+Int_t Int$slow_left_shifted(Int_t x, Int_t y);
+Int_t Int$slow_right_shifted(Int_t x, Int_t y);
+Int_t Int$slow_bit_and(Int_t x, Int_t y);
+Int_t Int$slow_bit_or(Int_t x, Int_t y);
+Int_t Int$slow_bit_xor(Int_t x, Int_t y);
+Int_t Int$slow_negative(Int_t x);
+Int_t Int$slow_negated(Int_t x);
 Int_t Int$abs(Int_t x);
 
 extern const TypeInfo $Int;
 
+// Fast-path inline versions for the common case where integer arithmetic is
+// between two small ints.
+
+static inline Int_t Int$plus(Int_t x, Int_t y) {
+    const int64_t z = (int64_t)((uint64_t)x.small + (uint64_t)y.small);
+    if (__builtin_expect(((z|2) == (int32_t)z), 1))
+        return (Int_t){.small=(z-1)};
+    return Int$slow_plus(x, y);
+}
+
+static inline Int_t Int$minus(Int_t x, Int_t y) {
+    const int64_t z = (int64_t)(((uint64_t)x.small ^ 3) - (uint64_t)y.small);
+    if (__builtin_expect(((z & ~2) == (int32_t)z), 1))
+        return (Int_t){.small=z};
+    return Int$slow_minus(x, y);
+}
+
+static inline Int_t Int$times(Int_t x, Int_t y) {
+    if (__builtin_expect(((x.small & y.small) & 1) != 0, 1)) {
+        const int64_t z = (x.small>>1) * (y.small>>1);
+        if (__builtin_expect(z == (int32_t)z, 1))
+            return (Int_t){.small=z+1};
+    }
+    return Int$slow_times(x, y);
+}
+
+static inline Int_t Int$divided_by(Int_t x, Int_t y) {
+    if (__builtin_expect(((x.small & y.small) & 1) != 0, 1)) {
+        const int64_t z = 4*(x.small>>1) / (y.small>>1);
+        if (__builtin_expect(z == (int32_t)z, 1))
+            return (Int_t){.small=z+1};
+    }
+    return Int$slow_divided_by(x, y);
+}
+
+static inline Int_t Int$modulo(Int_t x, Int_t y) {
+    if (__builtin_expect(((x.small & y.small) & 1) != 0, 1)) {
+        int64_t mod = (x.small>>2) % (y.small>>2);
+        if (mod < 0) mod += (y.small>>2);
+        return (Int_t){.small=(mod<<2)+1};
+    }
+    return Int$slow_modulo(x, y);
+}
+
+static inline Int_t Int$modulo1(Int_t x, Int_t y) {
+    if (__builtin_expect(((x.small & y.small) & 1) != 0, 1)) {
+        int64_t mod = ((x.small>>2)-1) % (y.small>>2);
+        if (mod < 0) mod += (y.small>>2);
+        return (Int_t){.small=((mod+1)<<2)+1};
+    }
+    return Int$slow_modulo1(x, y);
+}
+
+static inline Int_t Int$left_shifted(Int_t x, Int_t y) {
+    if (__builtin_expect(((x.small & y.small) & 1) != 0, 1)) {
+        const int64_t z = 4*((x.small>>2) << (y.small>>2));
+        if (__builtin_expect(z == (int32_t)z, 1))
+            return (Int_t){.small=z+1};
+    }
+    return Int$slow_left_shifted(x, y);
+}
+
+static inline Int_t Int$right_shifted(Int_t x, Int_t y) {
+    if (__builtin_expect(((x.small & y.small) & 1) != 0, 1)) {
+        const int64_t z = 4*((x.small>>2) >> (y.small>>2));
+        if (__builtin_expect(z == (int32_t)z, 1))
+            return (Int_t){.small=z+1};
+    }
+    return Int$slow_right_shifted(x, y);
+}
+
+static inline Int_t Int$bit_and(Int_t x, Int_t y) {
+    const int64_t z = x.small & y.small;
+    if (__builtin_expect((z & 1) == 1, 1))
+        return (Int_t){.small=z};
+    return Int$slow_bit_and(x, y);
+}
+
+static inline Int_t Int$bit_or(Int_t x, Int_t y) {
+    if (__builtin_expect(((x.small & y.small) & 1) == 1, 1))
+        return (Int_t){.small=(x.small | y.small)};
+    return Int$slow_bit_or(x, y);
+}
+
+static inline Int_t Int$bit_xor(Int_t x, Int_t y) {
+    if (__builtin_expect(((x.small & y.small) & 1) == 1, 1))
+        return (Int_t){.small=(x.small ^ y.small) | 1};
+    return Int$slow_bit_xor(x, y);
+}
+
+static inline Int_t Int$negated(Int_t x)
+{
+    if (__builtin_expect((x.small & 1), 1))
+        return (Int_t){.small=(~x.small) ^ 3};
+    return Int$slow_negated(x);
+}
+
+static inline Int_t Int$negative(Int_t x)
+{
+    if (__builtin_expect((x.small & 1), 1))
+        return (Int_t){.small=4*-((x.small)>>2) + 1};
+    return Int$slow_negative(x);
+}
 
 // vim: ts=4 sw=0 et cino=L2,l1,(0,W4,m1,\:0