Optimize llvm_cttz_i32() to remove the clumsy cttz_i8 lookup table to simplify code - (probably faster too, though did not benchmark; the arithmetic only form is greatly preferred anyway)

Author: juj

emscripten.py

@@ -1259,11 +1259,6 @@ def create_basic_vars(exported_implemented_functions, forwarded_json, metadata,
   basic_vars = ['DYNAMICTOP_PTR', 'tempDoublePtr', 'ABORT']
   if not (settings['WASM'] and settings['SIDE_MODULE']):
     basic_vars += ['STACKTOP', 'STACK_MAX']
-  if metadata.get('preciseI64MathUsed'):
-    basic_vars += ['cttz_i8']
-  else:
-    if forwarded_json['Functions']['libraryFunctions'].get('_llvm_cttz_i32'):
-      basic_vars += ['cttz_i8']
   if settings['RELOCATABLE']:
     if not (settings['WASM'] and settings['SIDE_MODULE']):
       basic_vars += ['gb', 'fb']

src/library.js

@@ -1067,37 +1067,13 @@ LibraryManager.library = {
     return ret | 0;
   },
 
-  llvm_cttz_i32__deps: [function() {
-    function cttz(x) {
-      for (var i = 0; i < 8; i++) {
-        if (x & (1 << i)) {
-          return i;
-        }
-      }
-      return 8;
-    }
-    if (SIDE_MODULE) return ''; // uses it from the parent
-
-#if USE_PTHREADS
-    return 'var cttz_i8; if (ENVIRONMENT_IS_PTHREAD) cttz_i8 = PthreadWorkerInit.cttz_i8; else PthreadWorkerInit.cttz_i8 = cttz_i8 = allocate([' + range(256).map(function(x) { return cttz(x) }).join(',') + '], "i8", ALLOC_STATIC);';
-#else
-    return 'var cttz_i8 = allocate([' + range(256).map(function(x) { return cttz(x) }).join(',') + '], "i8", ALLOC_STATIC);';
-#endif
-  }],
 #if WASM == 0 // binaryen will convert these calls to wasm anyhow
   llvm_cttz_i32__asm: true,
 #endif
   llvm_cttz_i32__sig: 'ii',
   llvm_cttz_i32: function(x) {
-    x = x|0;
-    var ret = 0;
-    ret = {{{ makeGetValueAsm('cttz_i8', 'x & 0xff', 'i8') }}};
-    if ((ret|0) < 8) return ret|0;
-    ret = {{{ makeGetValueAsm('cttz_i8', '(x >> 8)&0xff', 'i8') }}};
-    if ((ret|0) < 8) return (ret + 8)|0;
-    ret = {{{ makeGetValueAsm('cttz_i8', '(x >> 16)&0xff', 'i8') }}};
-    if ((ret|0) < 8) return (ret + 16)|0;
-    return ({{{ makeGetValueAsm('cttz_i8', 'x >>> 24', 'i8') }}} + 24)|0;
+    x = x | 0;
+    return x ? (31 - (Math_clz32((x ^ (x - 1))) | 0) | 0) : 32;
   },
 
   llvm_cttz_i64__deps: ['llvm_cttz_i32'],

tests/core/test_llvm_intrinsics.cpp

@@ -65,6 +65,16 @@ int main(void) {
   printf("%d,%d\n", (int)llvm_ctpop_i64((0x3101ULL << 32) | 1),
          llvm_ctpop_i32(0x3101));
 
+  printf("llvm_cttz_i32:\n");
+  printf("(0, 0)=%d\n", llvm_cttz_i32(0, 0));
+  printf("(1, 0)=%d\n", llvm_cttz_i32(1, 0));
+  printf("(2, 0)=%d\n", llvm_cttz_i32(2, 0));
+  printf("(0x0000FFFF, 0)=%d\n", llvm_cttz_i32(0x0000FFFF, 0));
+  printf("(0x7FFF0000, 0)=%d\n", llvm_cttz_i32(0x7FFF0000, 0));
+  printf("(0xFFFF0000, 0)=%d\n", llvm_cttz_i32(0xFFFF0000, 0));
+  printf("(0x7FFFFFFF, 0)=%d\n", llvm_cttz_i32(0x7FFFFFFF, 0));
+  printf("(0xFFFFFFFE, 0)=%d\n", llvm_cttz_i32(0xFFFFFFFE, 0));
+  printf("(0xFFFFFFFF, 0)=%d\n", llvm_cttz_i32(0xFFFFFFFF, 0));
   printf("small ctlz: %d,%d\n", (int)llvm_ctlz_i8(2, 0), llvm_ctlz_i16(2, 0));
 
   printf("llvm_ctpop_i32:\n");

tests/core/test_llvm_intrinsics.out

@@ -5,6 +5,16 @@ c5,de,15,8a
 23,21
 40,10
 5,4
+llvm_cttz_i32:
+(0, 0)=32
+(1, 0)=0
+(2, 0)=1
+(0x0000FFFF, 0)=0
+(0x7FFF0000, 0)=16
+(0xFFFF0000, 0)=16
+(0x7FFFFFFF, 0)=0
+(0xFFFFFFFE, 0)=1
+(0xFFFFFFFF, 0)=0
 small ctlz: 6,14
 llvm_ctpop_i32:
 22

tests/test_other.py

@@ -7976,7 +7976,7 @@ def test(filename, expectations):
                  0, [],                         ['tempDoublePtr', 'waka'],     8,   0,    0), # totally empty!
       # but we don't metadce with linkable code! other modules may want it
       (['-O3', '-s', 'MAIN_MODULE=1'],
-              1542, ['invoke_i'],               ['waka'],                 496958, 168, 2558),
+              1541, ['invoke_i'],               ['waka'],                 496958, 168, 2558),
     ])
 
     print('test on a minimal pure computational thing')