Fast numeric operations for Num

1 files changed, 49 insertions, 0 deletions

diff --git a/cbits/arith.c b/cbits/arith.c
new file mode 100644
index 0000000..02c8ce1
--- /dev/null
+++ b/cbits/arith.c
@@ -0,0 +1,49 @@
+#include <stdint.h>
+#include <stdlib.h>
+#include <math.h>
+
+typedef int32_t i32;
+typedef int64_t i64;
+
+#define COMM_OP(name, op, typ) \
+  void oxarop_ ## name ## _ ## typ ## _sv(i64 n, typ *out, typ x, typ *y) { \
+    for (i64 i = 0; i < n; i++) out[i] = x op y[i]; \
+  } \
+  void oxarop_ ## name ## _ ## typ ## _vv(i64 n, typ *out, const typ *x, const typ *y) { \
+    for (i64 i = 0; i < n; i++) out[i] = x[i] op y[i]; \
+  }
+
+#define NONCOMM_OP(name, op, typ) \
+  COMM_OP(name, op, typ) \
+  void oxarop_ ## name ## _ ## typ ## _vs(i64 n, typ *out, typ *x, typ y) { \
+    for (i64 i = 0; i < n; i++) out[i] = x[i] op y; \
+  }
+
+#define UNARY_OP(name, op, typ) \
+  void oxarop_ ## name ## _ ## typ(i64 n, typ *out, typ *x) { \
+    for (i64 i = 0; i < n; i++) out[i] = op(x[i]); \
+  }
+
+#define GEN_ABS(x) \
+  _Generic((x), \
+      int: abs, \
+      long: labs, \
+      long long: llabs, \
+      float: fabsf, \
+      double: fabs)(x)
+
+// This does not result in multiple loads with GCC 13.
+#define GEN_SIGNUM(x) ((x) < 0 ? -1 : (x) > 0 ? 1 : 0)
+
+#define NUM_TYPES_LOOP_XLIST \
+  X(i32) X(i64) X(double) X(float)
+
+#define X(typ) \
+  COMM_OP(add, +, typ) \
+  NONCOMM_OP(sub, -, typ) \
+  COMM_OP(mul, *, typ) \
+  UNARY_OP(neg, -, typ) \
+  UNARY_OP(abs, GEN_ABS, typ) \
+  UNARY_OP(signum, GEN_SIGNUM, typ)
+NUM_TYPES_LOOP_XLIST
+#undef X