hand-optimize memcpy/asm

2 files changed, 27 insertions, 42 deletions

diff --git a/emcc b/emcc
index 88f38bda..73f12cda 100755
--- a/emcc
+++ b/emcc
@@ -1089,36 +1089,6 @@ try:
       shared.Building.llvm_opt(in_temp(target_basename + '.bc'), link_opts)
       if DEBUG: save_intermediate('linktime', 'bc')
 
-    # Optimization and lto can add new intrinsics like memcpy that were not present before. We
-    # are now *after* linking in libc, so we missed our chance to get memcpy - check and add it now
-    # if necessary
-    final_symbols = shared.Building.llvm_nm(final)
-    need_memcpy = False
-    for symbol in final_symbols.undefs:
-      if symbol in MEMCPY_ALIASES:
-        need_memcpy = True
-        break
-    has_memcpy = False
-    for symbol in final_symbols.defs:
-      if symbol in MEMCPY_ALIASES:
-        has_memcpy = True
-        break
-    if need_memcpy and not has_memcpy:
-      if DEBUG: print >> sys.stderr, 'memcpy intrinsic added in optimizations, linking in optimized memcpy'
-      memcpy = in_temp('memcpy.bc')
-      force_cxx = os.environ.get('EMMAKEN_CXX')
-      if force_cxx is not None: del os.environ['EMMAKEN_CXX'] # memcpy must be compiled as C
-      execute([shared.PYTHON, shared.EMCC, shared.path_from_root('system', 'lib', 'libc', 'musl', 'memcpy.c'), '-o', memcpy], stdout=stdout, stderr=stderr)
-      if force_cxx is not None: os.environ['EMMAKEN_CXX'] = force_cxx
-      shared.Building.llvm_opt(memcpy, llvm_opts) # optimize it just like normal code; no point in lto though
-      next = final + '.postrinsics.bc'
-      shared.Building.link([final, memcpy], next)
-      final = next
-      if shared.Settings.ASM_JS: # export it so other library functions etc. can use it
-        if '_memcpy' not in shared.Settings.EXPORTED_FUNCTIONS:
-          shared.Settings.EXPORTED_FUNCTIONS.append('_memcpy')
-      if DEBUG: save_intermediate('postrinsics', 'bc')
-
   # Prepare .ll for Emscripten
   if not LEAVE_INPUTS_RAW:
     final = shared.Building.llvm_dis(final, final + '.ll')
diff --git a/src/library.js b/src/library.js
index 53fc8617..bc8502bc 100644
--- a/src/library.js
+++ b/src/library.js
@@ -4217,16 +4217,31 @@ LibraryManager.library = {
     return ret;
   },
 
+  memcpy__asm: 'true',
+  memcpy__sig: 'iiii',
   memcpy: function (dest, src, num) {
-    // simple version, in general it should not be used - we should pull it in from libc
-    if (!_memcpy.shown) {
-      _memcpy.shown = true;
-      Module.printErr('warning: library.js memcpy should not be running, it is only for testing!');
+    dest = dest|0; src = src|0; num = num|0;
+    if ((dest&3) == (src&3)) {
+      while (dest & 3 & num) {
+        {{{ makeSetValueAsm('dest', 0, makeGetValueAsm('src', 0, 'i8'), 'i8') }}};
+        dest = (dest+1)|0;
+        src = (src+1)|0;
+        num = (num-1)|0;
+      }
+      while (num|0 >= 4) {
+        {{{ makeSetValueAsm('dest', 0, makeGetValueAsm('src', 0, 'i32'), 'i32') }}};
+        dest = (dest+4)|0;
+        src = (src+4)|0;
+        num = (num-4)|0;
+      }
     }
-#endif
-    while (num--) {
-      HEAP8[dest++] = HEAP8[src++];
+    while (num|0 > 0) {
+      {{{ makeSetValueAsm('dest', 0, makeGetValueAsm('src', 0, 'i8'), 'i8') }}};
+      dest = (dest+1)|0;
+      src = (src+1)|0;
+      num = (num-1)|0;
     }
+    return dest|0;
   },
 
   wmemcpy: function() { throw 'wmemcpy not implemented' },
@@ -5601,11 +5616,11 @@ LibraryManager.library = {
   // ==========================================================================
 
   __utsname_struct_layout: Runtime.generateStructInfo([
-	  'sysname',
-	  'nodename',
-	  'release',
-	  'version',
-	  'machine'], '%struct.utsname'),
+    'sysname',
+    'nodename',
+    'release',
+    'version',
+    'machine'], '%struct.utsname'),
   uname__deps: ['__utsname_struct_layout'],
   uname: function(name) {
     // int uname(struct utsname *name);