Merge pull request #127 from startupfoundry/perf-guest-memcmp

port/android/guest/runtime/guest_string.c

@@ -0,0 +1,37 @@
+/*
+GUEST_STRING.C
+
+String functions faster than musl's generic C ones, which the guest takes
+in their place (tools/android_build.py MUSL_EXCLUDE). The renderer's state
+caches compare their keys with memcmp on every draw (the pixel shaders'
+keys, the vertex constants, the uniforms), and musl's compares a byte at a
+time.
+*/
+
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+/* eight bytes at a time while they match (AArch64 loads them unaligned),
+then the bytes that differ, or the tail. no_builtin: clang would turn the
+loops back into a call to memcmp */
+__attribute__((no_builtin)) int memcmp(const void *left, const void *right, size_t size)
+{
+	const unsigned char *l = left, *r = right;
+
+	for (; size >= 8; size -= 8, l += 8, r += 8)
+	{
+		uint64_t a, b;
+
+		__builtin_memcpy(&a, l, 8);
+		__builtin_memcpy(&b, r, 8);
+		if (a != b)
+			break;
+	}
+	for (; size; size--, l++, r++)
+	{
+		if (*l != *r)
+			return *l - *r;
+	}
+	return 0;
+}

tools/android_build.py

@@ -133,6 +133,8 @@ MUSL_EXCLUDE = {
     "env/__reset_tls.c", "malloc/oldmalloc", "thread/pthread_create.c",
     # unused, and its compiler barrier is an inline assembly statement
     "string/explicit_bzero.c",
+    # eight bytes at a time (guest_string.c)
+    "string/memcmp.c",
 }
 # game files that call variadic functions without a prototype in scope, which
 # only works under x86's calling convention (tools/android_abi_check.py)