Merge pull request #1088 from juj/branchless_i64_ops

Remove 'if()' branches from i64 add and subtract code.

3 files changed, 12 insertions, 12 deletions

diff --git a/src/library.js b/src/library.js
index ae0c845c..5bbf4204 100644
--- a/src/library.js
+++ b/src/library.js
@@ -7511,10 +7511,7 @@ LibraryManager.library = {
     a = a|0; b = b|0; c = c|0; d = d|0;
     var l = 0, h = 0;
     l = (a + c)>>>0;
-    h = (b + d)>>>0;
-    if ((l>>>0) < (a>>>0)) { // iff we overflowed
-      h = (h+1)>>>0;
-    }
+    h = (((b + d)>>>0) + (((l>>>0) < (a>>>0))>>>0))>>>0; // Add carry from low word to high word on overflow.
     {{{ makeStructuralReturn(['l|0', 'h'], true) }}};
   },
   llvm_uadd_with_overflow_i64__asm: true,
@@ -7524,11 +7521,9 @@ LibraryManager.library = {
     var l = 0, h = 0, overflow = 0;
     l = (a + c)>>>0;
     h = (b + d)>>>0;
-    if ((h>>>0) < (b>>>0)) overflow = 1;
-    if ((l>>>0) < (a>>>0)) {
-      h = (h+1)>>>0;
-      if ((h>>>0) == 0) overflow = 1; // two possibilities to overflow here
-    }
+    overflow = ((h>>>0) < (b>>>0))>>>0; // Return whether addition overflowed even the high word.
+    h = h + (((l>>>0) < (a>>>0))>>>0)>>>0; // Add carry from low word to high word on overflow.
+    overflow = overflow | (h == 0); // Check again for overflow.
     {{{ makeStructuralReturn(['l|0', 'h', 'overflow'], true) }}};
   },
 
@@ -7539,9 +7534,7 @@ LibraryManager.library = {
     var l = 0, h = 0;
     l = (a - c)>>>0;
     h = (b - d)>>>0;
-    if ((l>>>0) > (a>>>0)) { // iff we overflowed
-      h = (h-1)>>>0;
-    }
+    h = (((b - d)>>>0) - (((c>>>0) > (a>>>0))>>>0))>>>0; // Borrow one from high word to low word on underflow.
     {{{ makeStructuralReturn(['l|0', 'h'], true) }}};
   },
 
diff --git a/tests/cases/uadd_overflow_64_ta2.ll b/tests/cases/uadd_overflow_64_ta2.ll
index a4f3e40b..6dfbc1e5 100644
--- a/tests/cases/uadd_overflow_64_ta2.ll
+++ b/tests/cases/uadd_overflow_64_ta2.ll
@@ -19,6 +19,12 @@ entry:
   %b2 = zext i1 %b1 to i32
   call i32 (i8*, ...)* @printf(i8* getelementptr inbounds ([11 x i8]* @.str2, i32 0, i32 0), i64 %b0, i32 %b2) ; [#uses=0]
 
+  %uadd3 = tail call { i64, i1 } @llvm.uadd.with.overflow.i64(i64 4294967297, i64 18446744073709551615)
+  %c0 = extractvalue { i64, i1 } %uadd3, 0
+  %c1 = extractvalue { i64, i1 } %uadd3, 1
+  %c2 = zext i1 %c1 to i32
+  call i32 (i8*, ...)* @printf(i8* getelementptr inbounds ([11 x i8]* @.str2, i32 0, i32 0), i64 %c0, i32 %c2) ; [#uses=0]
+
   ret i32 1
 }
 
diff --git a/tests/cases/uadd_overflow_64_ta2.txt b/tests/cases/uadd_overflow_64_ta2.txt
index f5711309..2abb084c 100644
--- a/tests/cases/uadd_overflow_64_ta2.txt
+++ b/tests/cases/uadd_overflow_64_ta2.txt
@@ -1,2 +1,3 @@
 *2705,1*
 *10c6f7a0dcfc,0*
+*100000000,1*
\ No newline at end of file