Fold LIBC_BITS into LIBC_INTRIN

libc/intrin/midpoint.h

@@ -0,0 +1,32 @@
+#ifndef COSMOPOLITAN_LIBC_BITS_MIDPOINT_H_
+#define COSMOPOLITAN_LIBC_BITS_MIDPOINT_H_
+#include "libc/assert.h"
+#if !(__ASSEMBLER__ + __LINKER__ + 0)
+
+#if defined(__GNUC__) && !defined(__STRICT_ANSI__) && defined(__x86__)
+/**
+ * Computes `(a + b) / 2` assuming unsigned.
+ *
+ * This implementation is the fastest on AMD Zen architecture.
+ */
+#define _midpoint(a, b)         \
+  ({                            \
+    typeof((a) + (b)) a_ = (a); \
+    typeof(a_) b_ = (b);        \
+    assert(a_ >= 0);            \
+    assert(b_ >= 0);            \
+    asm("add\t%1,%0\n\t"        \
+        "rcr\t%0"               \
+        : "+r"(a_)              \
+        : "r"(b_));             \
+    a_;                         \
+  })
+#else
+/**
+ * Computes `(a + b) / 2` assuming unsigned.
+ */
+#define _midpoint(a, b) (((a) & (b)) + ((a) ^ (b)) / 2)
+#endif /* __GNUC__ && !__STRICT_ANSI__ && x86 */
+
+#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */
+#endif /* COSMOPOLITAN_LIBC_BITS_MIDPOINT_H_ */