diff --git a/CMakeLists.txt b/CMakeLists.txt
index f8fc04ef7d..91272fc9cc 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -494,8 +494,7 @@ set(ZIG_STAGE2_SOURCES
     "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/modti3.zig"
     "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/mulXf3.zig"
     "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/muldi3.zig"
-    "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/mulodi4.zig"
-    "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/muloti4.zig"
+    "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/mulo.zig"
     "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/multi3.zig"
     "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/negXf2.zig"
     "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/negXi2.zig"
diff --git a/lib/std/special/compiler_rt.zig b/lib/std/special/compiler_rt.zig
index e9f422be59..310bffbbe2 100644
--- a/lib/std/special/compiler_rt.zig
+++ b/lib/std/special/compiler_rt.zig
@@ -77,9 +77,11 @@ comptime {
         @export(__extendhfsf2, .{ .name = "__gnu_h2f_ieee", .linkage = linkage });
 
         // Integral arithmetic which returns if overflow
-        const __mulodi4 = @import("compiler_rt/mulodi4.zig").__mulodi4;
+        const __mulosi4 = @import("compiler_rt/mulo.zig").__mulosi4;
+        @export(__mulosi4, .{ .name = "__mulosi4", .linkage = linkage });
+        const __mulodi4 = @import("compiler_rt/mulo.zig").__mulodi4;
         @export(__mulodi4, .{ .name = "__mulodi4", .linkage = linkage });
-        const __muloti4 = @import("compiler_rt/muloti4.zig").__muloti4;
+        const __muloti4 = @import("compiler_rt/mulo.zig").__muloti4;
         @export(__muloti4, .{ .name = "__muloti4", .linkage = linkage });
     }
 
diff --git a/lib/std/special/compiler_rt/bswap.zig b/lib/std/special/compiler_rt/bswap.zig
index d903039626..0646f68c1a 100644
--- a/lib/std/special/compiler_rt/bswap.zig
+++ b/lib/std/special/compiler_rt/bswap.zig
@@ -6,7 +6,6 @@ const builtin = @import("builtin");
 // ie for u32
 // DE AD BE EF   <- little|big endian
 // FE BE AD DE   <- big|little endian
-// ie for u32
 // ff 00 00 00 >> 3*8 (leftmost  byte)
 // 00 ff 00 00 >> 1*8 (2nd left  byte)
 // 00 00 ff 00 << 1*8 (2n right  byte)
diff --git a/lib/std/special/compiler_rt/mulo.zig b/lib/std/special/compiler_rt/mulo.zig
new file mode 100644
index 0000000000..1e812c7e99
--- /dev/null
+++ b/lib/std/special/compiler_rt/mulo.zig
@@ -0,0 +1,69 @@
+const builtin = @import("builtin");
+
+// mulo - multiplication overflow
+// - muloXi4_generic for unoptimized version
+
+// return a*b.
+// return if a*b overflows => 1 else => 0
+// see https://stackoverflow.com/a/26320664 for possible implementations
+
+fn muloXi4_generic(comptime ST: type) fn (a: ST, b: ST, overflow: *c_int) callconv(.C) ST {
+    return struct {
+        fn f(a: ST, b: ST, overflow: *c_int) callconv(.C) ST {
+            @setRuntimeSafety(builtin.is_test);
+            const BSIZE = @bitSizeOf(ST);
+            comptime var UT = switch (ST) {
+                i32 => u32,
+                i64 => u64,
+                i128 => u128,
+                else => unreachable,
+            };
+            const min = @bitCast(ST, @as(UT, 1 << (BSIZE - 1)));
+            const max = ~min;
+            overflow.* = 0;
+            const result = a *% b;
+
+            // edge cases
+            if (a == min) {
+                if (b != 0 and b != 1) overflow.* = 1;
+                return result;
+            }
+            if (b == min) {
+                if (a != 0 and a != 1) overflow.* = 1;
+                return result;
+            }
+
+            // take sign of x sx
+            const sa = a >> (BSIZE - 1);
+            const sb = b >> (BSIZE - 1);
+            // take absolute value of a and b via
+            // abs(x) = (x^sx)) - sx
+            const abs_a = (a ^ sa) -% sa;
+            const abs_b = (b ^ sb) -% sb;
+
+            // unitary magnitude, cannot have overflow
+            if (abs_a < 2 or abs_b < 2) return result;
+
+            // compare the signs of operands
+            if ((a ^ b) >> (BSIZE - 1) != 0) {
+                if (abs_a > @divTrunc(max, abs_b)) overflow.* = 1;
+            } else {
+                if (abs_a > @divTrunc(min, -abs_b)) overflow.* = 1;
+            }
+
+            return result;
+        }
+    }.f;
+}
+
+pub const __mulosi4 = muloXi4_generic(i32);
+
+pub const __mulodi4 = muloXi4_generic(i64);
+
+pub const __muloti4 = muloXi4_generic(i128);
+
+test {
+    _ = @import("mulosi4_test.zig");
+    _ = @import("mulodi4_test.zig");
+    _ = @import("muloti4_test.zig");
+}
diff --git a/lib/std/special/compiler_rt/mulodi4.zig b/lib/std/special/compiler_rt/mulodi4.zig
deleted file mode 100644
index d58cfbfe11..0000000000
--- a/lib/std/special/compiler_rt/mulodi4.zig
+++ /dev/null
@@ -1,42 +0,0 @@
-const builtin = @import("builtin");
-const compiler_rt = @import("../compiler_rt.zig");
-
-pub fn __mulodi4(a: i64, b: i64, overflow: *c_int) callconv(.C) i64 {
-    @setRuntimeSafety(builtin.is_test);
-
-    const min = @bitCast(i64, @as(u64, 1 << (64 - 1)));
-    const max = ~min;
-
-    overflow.* = 0;
-    const result = a *% b;
-
-    // Edge cases
-    if (a == min) {
-        if (b != 0 and b != 1) overflow.* = 1;
-        return result;
-    }
-    if (b == min) {
-        if (a != 0 and a != 1) overflow.* = 1;
-        return result;
-    }
-
-    // Take absolute value of a and b via abs(x) = (x^(x >> 63)) - (x >> 63).
-    const abs_a = (a ^ (a >> 63)) -% (a >> 63);
-    const abs_b = (b ^ (b >> 63)) -% (b >> 63);
-
-    // Unitary magnitude, cannot have overflow
-    if (abs_a < 2 or abs_b < 2) return result;
-
-    // Compare the signs of the operands
-    if ((a ^ b) >> 63 != 0) {
-        if (abs_a > @divTrunc(max, abs_b)) overflow.* = 1;
-    } else {
-        if (abs_a > @divTrunc(min, -abs_b)) overflow.* = 1;
-    }
-
-    return result;
-}
-
-test {
-    _ = @import("mulodi4_test.zig");
-}
diff --git a/lib/std/special/compiler_rt/mulodi4_test.zig b/lib/std/special/compiler_rt/mulodi4_test.zig
index b9fea2553f..3944f62ede 100644
--- a/lib/std/special/compiler_rt/mulodi4_test.zig
+++ b/lib/std/special/compiler_rt/mulodi4_test.zig
@@ -1,9 +1,11 @@
-const __mulodi4 = @import("mulodi4.zig").__mulodi4;
+const mulo = @import("mulo.zig");
 const testing = @import("std").testing;
 
+// ported from https://github.com/llvm-mirror/compiler-rt/tree/release_80/test/builtins/Unit
+
 fn test__mulodi4(a: i64, b: i64, expected: i64, expected_overflow: c_int) !void {
     var overflow: c_int = undefined;
-    const x = __mulodi4(a, b, &overflow);
+    const x = mulo.__mulodi4(a, b, &overflow);
     try testing.expect(overflow == expected_overflow and (expected_overflow != 0 or x == expected));
 }
 
diff --git a/lib/std/special/compiler_rt/mulosi4_test.zig b/lib/std/special/compiler_rt/mulosi4_test.zig
new file mode 100644
index 0000000000..523faa490f
--- /dev/null
+++ b/lib/std/special/compiler_rt/mulosi4_test.zig
@@ -0,0 +1,72 @@
+const mulo = @import("mulo.zig");
+const testing = @import("std").testing;
+
+// ported from https://github.com/llvm-mirror/compiler-rt/tree/release_80/test/builtins/Unit
+
+fn test__mulosi4(a: i32, b: i32, expected: i32, expected_overflow: c_int) !void {
+    var overflow: c_int = undefined;
+    const x = mulo.__mulosi4(a, b, &overflow);
+    try testing.expect(overflow == expected_overflow and (expected_overflow != 0 or x == expected));
+}
+
+test "mulosi4" {
+    try test__mulosi4(0, 0, 0, 0);
+    try test__mulosi4(0, 1, 0, 0);
+    try test__mulosi4(1, 0, 0, 0);
+    try test__mulosi4(0, 10, 0, 0);
+    try test__mulosi4(10, 0, 0, 0);
+    try test__mulosi4(0, 0x1234567, 0, 0);
+    try test__mulosi4(0x1234567, 0, 0, 0);
+
+    try test__mulosi4(0, -1, 0, 0);
+    try test__mulosi4(-1, 0, 0, 0);
+    try test__mulosi4(0, -10, 0, 0);
+    try test__mulosi4(-10, 0, 0, 0);
+    try test__mulosi4(0, -0x1234567, 0, 0);
+    try test__mulosi4(-0x1234567, 0, 0, 0);
+
+    try test__mulosi4(1, 1, 1, 0);
+    try test__mulosi4(1, 10, 10, 0);
+    try test__mulosi4(10, 1, 10, 0);
+    try test__mulosi4(1, 0x1234567, 0x1234567, 0);
+    try test__mulosi4(0x1234567, 1, 0x1234567, 0);
+
+    try test__mulosi4(1, -1, -1, 0);
+    try test__mulosi4(1, -10, -10, 0);
+    try test__mulosi4(-10, 1, -10, 0);
+    try test__mulosi4(1, -0x1234567, -0x1234567, 0);
+    try test__mulosi4(-0x1234567, 1, -0x1234567, 0);
+
+    try test__mulosi4(0x7FFFFFFF, -2, @bitCast(i32, @as(u32, 0x80000001)), 1);
+    try test__mulosi4(-2, 0x7FFFFFFF, @bitCast(i32, @as(u32, 0x80000001)), 1);
+    try test__mulosi4(0x7FFFFFFF, -1, @bitCast(i32, @as(u32, 0x80000001)), 0);
+    try test__mulosi4(-1, 0x7FFFFFFF, @bitCast(i32, @as(u32, 0x80000001)), 0);
+    try test__mulosi4(0x7FFFFFFF, 0, 0, 0);
+    try test__mulosi4(0, 0x7FFFFFFF, 0, 0);
+    try test__mulosi4(0x7FFFFFFF, 1, 0x7FFFFFFF, 0);
+    try test__mulosi4(1, 0x7FFFFFFF, 0x7FFFFFFF, 0);
+    try test__mulosi4(0x7FFFFFFF, 2, @bitCast(i32, @as(u32, 0x80000001)), 1);
+    try test__mulosi4(2, 0x7FFFFFFF, @bitCast(i32, @as(u32, 0x80000001)), 1);
+
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000000)), -2, @bitCast(i32, @as(u32, 0x80000000)), 1);
+    try test__mulosi4(-2, @bitCast(i32, @as(u32, 0x80000000)), @bitCast(i32, @as(u32, 0x80000000)), 1);
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000000)), -1, @bitCast(i32, @as(u32, 0x80000000)), 1);
+    try test__mulosi4(-1, @bitCast(i32, @as(u32, 0x80000000)), @bitCast(i32, @as(u32, 0x80000000)), 1);
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000000)), 0, 0, 0);
+    try test__mulosi4(0, @bitCast(i32, @as(u32, 0x80000000)), 0, 0);
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000000)), 1, @bitCast(i32, @as(u32, 0x80000000)), 0);
+    try test__mulosi4(1, @bitCast(i32, @as(u32, 0x80000000)), @bitCast(i32, @as(u32, 0x80000000)), 0);
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000000)), 2, @bitCast(i32, @as(u32, 0x80000000)), 1);
+    try test__mulosi4(2, @bitCast(i32, @as(u32, 0x80000000)), @bitCast(i32, @as(u32, 0x80000000)), 1);
+
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000001)), -2, @bitCast(i32, @as(u32, 0x80000001)), 1);
+    try test__mulosi4(-2, @bitCast(i32, @as(u32, 0x80000001)), @bitCast(i32, @as(u32, 0x80000001)), 1);
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000001)), -1, 0x7FFFFFFF, 0);
+    try test__mulosi4(-1, @bitCast(i32, @as(u32, 0x80000001)), 0x7FFFFFFF, 0);
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000001)), 0, 0, 0);
+    try test__mulosi4(0, @bitCast(i32, @as(u32, 0x80000001)), 0, 0);
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000001)), 1, @bitCast(i32, @as(u32, 0x80000001)), 0);
+    try test__mulosi4(1, @bitCast(i32, @as(u32, 0x80000001)), @bitCast(i32, @as(u32, 0x80000001)), 0);
+    try test__mulosi4(@bitCast(i32, @as(u32, 0x80000001)), 2, @bitCast(i32, @as(u32, 0x80000000)), 1);
+    try test__mulosi4(2, @bitCast(i32, @as(u32, 0x80000001)), @bitCast(i32, @as(u32, 0x80000000)), 1);
+}
diff --git a/lib/std/special/compiler_rt/muloti4.zig b/lib/std/special/compiler_rt/muloti4.zig
deleted file mode 100644
index 7e3497ed7b..0000000000
--- a/lib/std/special/compiler_rt/muloti4.zig
+++ /dev/null
@@ -1,49 +0,0 @@
-const builtin = @import("builtin");
-const compiler_rt = @import("../compiler_rt.zig");
-
-pub fn __muloti4(a: i128, b: i128, overflow: *c_int) callconv(.C) i128 {
-    @setRuntimeSafety(builtin.is_test);
-
-    const min = @bitCast(i128, @as(u128, 1 << (128 - 1)));
-    const max = ~min;
-    overflow.* = 0;
-
-    const r = a *% b;
-    if (a == min) {
-        if (b != 0 and b != 1) {
-            overflow.* = 1;
-        }
-        return r;
-    }
-    if (b == min) {
-        if (a != 0 and a != 1) {
-            overflow.* = 1;
-        }
-        return r;
-    }
-
-    const sa = a >> (128 - 1);
-    const abs_a = (a ^ sa) -% sa;
-    const sb = b >> (128 - 1);
-    const abs_b = (b ^ sb) -% sb;
-
-    if (abs_a < 2 or abs_b < 2) {
-        return r;
-    }
-
-    if (sa == sb) {
-        if (abs_a > @divTrunc(max, abs_b)) {
-            overflow.* = 1;
-        }
-    } else {
-        if (abs_a > @divTrunc(min, -abs_b)) {
-            overflow.* = 1;
-        }
-    }
-
-    return r;
-}
-
-test {
-    _ = @import("muloti4_test.zig");
-}
diff --git a/lib/std/special/compiler_rt/muloti4_test.zig b/lib/std/special/compiler_rt/muloti4_test.zig
index 08450797fa..6d204ff785 100644
--- a/lib/std/special/compiler_rt/muloti4_test.zig
+++ b/lib/std/special/compiler_rt/muloti4_test.zig
@@ -1,9 +1,11 @@
-const __muloti4 = @import("muloti4.zig").__muloti4;
+const mulo = @import("mulo.zig");
 const testing = @import("std").testing;
 
+// ported from https://github.com/llvm-mirror/compiler-rt/tree/release_80/test/builtins/Unit
+
 fn test__muloti4(a: i128, b: i128, expected: i128, expected_overflow: c_int) !void {
     var overflow: c_int = undefined;
-    const x = __muloti4(a, b, &overflow);
+    const x = mulo.__muloti4(a, b, &overflow);
     try testing.expect(overflow == expected_overflow and (expected_overflow != 0 or x == expected));
 }
 
@@ -13,7 +15,6 @@ test "muloti4" {
     try test__muloti4(1, 0, 0, 0);
     try test__muloti4(0, 10, 0, 0);
     try test__muloti4(10, 0, 0, 0);
-
     try test__muloti4(0, 81985529216486895, 0, 0);
     try test__muloti4(81985529216486895, 0, 0, 0);
 
@@ -24,6 +25,18 @@ test "muloti4" {
     try test__muloti4(0, -81985529216486895, 0, 0);
     try test__muloti4(-81985529216486895, 0, 0, 0);
 
+    try test__muloti4(1, 1, 1, 0);
+    try test__muloti4(1, 10, 10, 0);
+    try test__muloti4(10, 1, 10, 0);
+    try test__muloti4(1, 81985529216486895, 81985529216486895, 0);
+    try test__muloti4(81985529216486895, 1, 81985529216486895, 0);
+
+    try test__muloti4(1, -1, -1, 0);
+    try test__muloti4(1, -10, -10, 0);
+    try test__muloti4(-10, 1, -10, 0);
+    try test__muloti4(1, -81985529216486895, -81985529216486895, 0);
+    try test__muloti4(-81985529216486895, 1, -81985529216486895, 0);
+
     try test__muloti4(3037000499, 3037000499, 9223372030926249001, 0);
     try test__muloti4(-3037000499, 3037000499, -9223372030926249001, 0);
     try test__muloti4(3037000499, -3037000499, -9223372030926249001, 0);