Commit 2fcb2f5975

Jacob Young <jacobly0@users.noreply.github.com>
2024-02-15 10:37:52
Sema: implement vector coercions
These used to be lowered elementwise in air, and now are a single air instruction that can be lowered elementwise in the backend if necessary.
1 parent 2fdc9e6
lib/std/unicode.zig
@@ -602,9 +602,9 @@ fn testUtf8IteratorOnAscii() !void {
     const s = Utf8View.initComptime("abc");
 
     var it1 = s.iterator();
-    try testing.expect(std.mem.eql(u8, "a", it1.nextCodepointSlice().?));
-    try testing.expect(std.mem.eql(u8, "b", it1.nextCodepointSlice().?));
-    try testing.expect(std.mem.eql(u8, "c", it1.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "a", it1.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "b", it1.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "c", it1.nextCodepointSlice().?));
     try testing.expect(it1.nextCodepointSlice() == null);
 
     var it2 = s.iterator();
@@ -632,9 +632,9 @@ fn testUtf8ViewOk() !void {
     const s = Utf8View.initComptime("東京市");
 
     var it1 = s.iterator();
-    try testing.expect(std.mem.eql(u8, "東", it1.nextCodepointSlice().?));
-    try testing.expect(std.mem.eql(u8, "京", it1.nextCodepointSlice().?));
-    try testing.expect(std.mem.eql(u8, "市", it1.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "東", it1.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "京", it1.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "市", it1.nextCodepointSlice().?));
     try testing.expect(it1.nextCodepointSlice() == null);
 
     var it2 = s.iterator();
@@ -772,20 +772,20 @@ fn testUtf8Peeking() !void {
     const s = Utf8View.initComptime("noël");
     var it = s.iterator();
 
-    try testing.expect(std.mem.eql(u8, "n", it.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "n", it.nextCodepointSlice().?));
 
-    try testing.expect(std.mem.eql(u8, "o", it.peek(1)));
-    try testing.expect(std.mem.eql(u8, "oë", it.peek(2)));
-    try testing.expect(std.mem.eql(u8, "oël", it.peek(3)));
-    try testing.expect(std.mem.eql(u8, "oël", it.peek(4)));
-    try testing.expect(std.mem.eql(u8, "oël", it.peek(10)));
+    try testing.expect(mem.eql(u8, "o", it.peek(1)));
+    try testing.expect(mem.eql(u8, "oë", it.peek(2)));
+    try testing.expect(mem.eql(u8, "oël", it.peek(3)));
+    try testing.expect(mem.eql(u8, "oël", it.peek(4)));
+    try testing.expect(mem.eql(u8, "oël", it.peek(10)));
 
-    try testing.expect(std.mem.eql(u8, "o", it.nextCodepointSlice().?));
-    try testing.expect(std.mem.eql(u8, "ë", it.nextCodepointSlice().?));
-    try testing.expect(std.mem.eql(u8, "l", it.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "o", it.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "ë", it.nextCodepointSlice().?));
+    try testing.expect(mem.eql(u8, "l", it.nextCodepointSlice().?));
     try testing.expect(it.nextCodepointSlice() == null);
 
-    try testing.expect(std.mem.eql(u8, &[_]u8{}, it.peek(1)));
+    try testing.expect(mem.eql(u8, &[_]u8{}, it.peek(1)));
 }
 
 fn testError(bytes: []const u8, expected_err: anyerror) !void {
@@ -927,20 +927,16 @@ test "fmtUtf8" {
 }
 
 fn utf16LeToUtf8ArrayListImpl(
-    array_list: *std.ArrayList(u8),
+    result: *std.ArrayList(u8),
     utf16le: []const u16,
     comptime surrogates: Surrogates,
 ) (switch (surrogates) {
     .cannot_encode_surrogate_half => Utf16LeToUtf8AllocError,
     .can_encode_surrogate_half => mem.Allocator.Error,
 })!void {
-    // optimistically guess that it will all be ascii.
-    try array_list.ensureTotalCapacityPrecise(utf16le.len);
+    assert(result.capacity >= utf16le.len);
 
     var remaining = utf16le;
-    if (builtin.zig_backend != .stage2_x86_64 or
-        comptime (std.Target.x86.featureSetHas(builtin.cpu.features, .ssse3) and
-        !std.Target.x86.featureSetHasAny(builtin.cpu.features, .{ .prefer_256_bit, .avx })))
     vectorized: {
         const chunk_len = std.simd.suggestVectorLength(u16) orelse break :vectorized;
         const Chunk = @Vector(chunk_len, u16);
@@ -948,41 +944,33 @@ fn utf16LeToUtf8ArrayListImpl(
         // Fast path. Check for and encode ASCII characters at the start of the input.
         while (remaining.len >= chunk_len) {
             const chunk: Chunk = remaining[0..chunk_len].*;
-            const mask: Chunk = @splat(std.mem.nativeToLittle(u16, 0x7F));
+            const mask: Chunk = @splat(mem.nativeToLittle(u16, 0x7F));
             if (@reduce(.Or, chunk | mask != mask)) {
                 // found a non ASCII code unit
                 break;
             }
-            const chunk_byte_len = chunk_len * 2;
-            const chunk_bytes: @Vector(chunk_byte_len, u8) = (std.mem.sliceAsBytes(remaining)[0..chunk_byte_len]).*;
-            const deinterlaced_bytes = std.simd.deinterlace(2, chunk_bytes);
-            const ascii_bytes: [chunk_len]u8 = deinterlaced_bytes[0];
+            const ascii_chunk: @Vector(chunk_len, u8) = @truncate(mem.nativeToLittle(Chunk, chunk));
             // We allocated enough space to encode every UTF-16 code unit
             // as ASCII, so if the entire string is ASCII then we are
             // guaranteed to have enough space allocated
-            array_list.appendSliceAssumeCapacity(&ascii_bytes);
+            result.addManyAsArrayAssumeCapacity(chunk_len).* = ascii_chunk;
             remaining = remaining[chunk_len..];
         }
     }
 
-    var out_index: usize = array_list.items.len;
     switch (surrogates) {
         .cannot_encode_surrogate_half => {
             var it = Utf16LeIterator.init(remaining);
             while (try it.nextCodepoint()) |codepoint| {
                 const utf8_len = utf8CodepointSequenceLength(codepoint) catch unreachable;
-                try array_list.resize(array_list.items.len + utf8_len);
-                assert((utf8Encode(codepoint, array_list.items[out_index..]) catch unreachable) == utf8_len);
-                out_index += utf8_len;
+                assert((utf8Encode(codepoint, try result.addManyAsSlice(utf8_len)) catch unreachable) == utf8_len);
             }
         },
         .can_encode_surrogate_half => {
             var it = Wtf16LeIterator.init(remaining);
             while (it.nextCodepoint()) |codepoint| {
                 const utf8_len = utf8CodepointSequenceLength(codepoint) catch unreachable;
-                try array_list.resize(array_list.items.len + utf8_len);
-                assert((wtf8Encode(codepoint, array_list.items[out_index..]) catch unreachable) == utf8_len);
-                out_index += utf8_len;
+                assert((wtf8Encode(codepoint, try result.addManyAsSlice(utf8_len)) catch unreachable) == utf8_len);
             }
         },
     }
@@ -990,8 +978,9 @@ fn utf16LeToUtf8ArrayListImpl(
 
 pub const Utf16LeToUtf8AllocError = mem.Allocator.Error || Utf16LeToUtf8Error;
 
-pub fn utf16LeToUtf8ArrayList(array_list: *std.ArrayList(u8), utf16le: []const u16) Utf16LeToUtf8AllocError!void {
-    return utf16LeToUtf8ArrayListImpl(array_list, utf16le, .cannot_encode_surrogate_half);
+pub fn utf16LeToUtf8ArrayList(result: *std.ArrayList(u8), utf16le: []const u16) Utf16LeToUtf8AllocError!void {
+    try result.ensureTotalCapacityPrecise(utf16le.len);
+    return utf16LeToUtf8ArrayListImpl(result, utf16le, .cannot_encode_surrogate_half);
 }
 
 /// Deprecated; renamed to utf16LeToUtf8Alloc
@@ -1003,8 +992,7 @@ pub fn utf16LeToUtf8Alloc(allocator: mem.Allocator, utf16le: []const u16) Utf16L
     var result = try std.ArrayList(u8).initCapacity(allocator, utf16le.len);
     errdefer result.deinit();
 
-    try utf16LeToUtf8ArrayList(&result, utf16le);
-
+    try utf16LeToUtf8ArrayListImpl(&result, utf16le, .cannot_encode_surrogate_half);
     return result.toOwnedSlice();
 }
 
@@ -1017,8 +1005,7 @@ pub fn utf16LeToUtf8AllocZ(allocator: mem.Allocator, utf16le: []const u16) Utf16
     var result = try std.ArrayList(u8).initCapacity(allocator, utf16le.len + 1);
     errdefer result.deinit();
 
-    try utf16LeToUtf8ArrayList(&result, utf16le);
-
+    try utf16LeToUtf8ArrayListImpl(&result, utf16le, .cannot_encode_surrogate_half);
     return result.toOwnedSliceSentinel(0);
 }
 
@@ -1030,12 +1017,9 @@ fn utf16LeToUtf8Impl(utf8: []u8, utf16le: []const u16, comptime surrogates: Surr
     .cannot_encode_surrogate_half => Utf16LeToUtf8Error,
     .can_encode_surrogate_half => error{},
 })!usize {
-    var end_index: usize = 0;
+    var dest_index: usize = 0;
 
     var remaining = utf16le;
-    if (builtin.zig_backend != .stage2_x86_64 or
-        comptime (std.Target.x86.featureSetHas(builtin.cpu.features, .ssse3) and
-        !std.Target.x86.featureSetHasAny(builtin.cpu.features, .{ .prefer_256_bit, .avx })))
     vectorized: {
         const chunk_len = std.simd.suggestVectorLength(u16) orelse break :vectorized;
         const Chunk = @Vector(chunk_len, u16);
@@ -1043,17 +1027,14 @@ fn utf16LeToUtf8Impl(utf8: []u8, utf16le: []const u16, comptime surrogates: Surr
         // Fast path. Check for and encode ASCII characters at the start of the input.
         while (remaining.len >= chunk_len) {
             const chunk: Chunk = remaining[0..chunk_len].*;
-            const mask: Chunk = @splat(std.mem.nativeToLittle(u16, 0x7F));
+            const mask: Chunk = @splat(mem.nativeToLittle(u16, 0x7F));
             if (@reduce(.Or, chunk | mask != mask)) {
                 // found a non ASCII code unit
                 break;
             }
-            const chunk_byte_len = chunk_len * 2;
-            const chunk_bytes: @Vector(chunk_byte_len, u8) = (std.mem.sliceAsBytes(remaining)[0..chunk_byte_len]).*;
-            const deinterlaced_bytes = std.simd.deinterlace(2, chunk_bytes);
-            const ascii_bytes: [chunk_len]u8 = deinterlaced_bytes[0];
-            @memcpy(utf8[end_index .. end_index + chunk_len], &ascii_bytes);
-            end_index += chunk_len;
+            const ascii_chunk: @Vector(chunk_len, u8) = @truncate(mem.nativeToLittle(Chunk, chunk));
+            utf8[dest_index..][0..chunk_len].* = ascii_chunk;
+            dest_index += chunk_len;
             remaining = remaining[chunk_len..];
         }
     }
@@ -1062,7 +1043,7 @@ fn utf16LeToUtf8Impl(utf8: []u8, utf16le: []const u16, comptime surrogates: Surr
         .cannot_encode_surrogate_half => {
             var it = Utf16LeIterator.init(remaining);
             while (try it.nextCodepoint()) |codepoint| {
-                end_index += utf8Encode(codepoint, utf8[end_index..]) catch |err| switch (err) {
+                dest_index += utf8Encode(codepoint, utf8[dest_index..]) catch |err| switch (err) {
                     // The maximum possible codepoint encoded by UTF-16 is U+10FFFF,
                     // which is within the valid codepoint range.
                     error.CodepointTooLarge => unreachable,
@@ -1075,7 +1056,7 @@ fn utf16LeToUtf8Impl(utf8: []u8, utf16le: []const u16, comptime surrogates: Surr
         .can_encode_surrogate_half => {
             var it = Wtf16LeIterator.init(remaining);
             while (it.nextCodepoint()) |codepoint| {
-                end_index += wtf8Encode(codepoint, utf8[end_index..]) catch |err| switch (err) {
+                dest_index += wtf8Encode(codepoint, utf8[dest_index..]) catch |err| switch (err) {
                     // The maximum possible codepoint encoded by UTF-16 is U+10FFFF,
                     // which is within the valid codepoint range.
                     error.CodepointTooLarge => unreachable,
@@ -1083,7 +1064,7 @@ fn utf16LeToUtf8Impl(utf8: []u8, utf16le: []const u16, comptime surrogates: Surr
             }
         },
     }
-    return end_index;
+    return dest_index;
 }
 
 /// Deprecated; renamed to utf16LeToUtf8
@@ -1156,18 +1137,12 @@ test utf16LeToUtf8 {
     }
 }
 
-fn utf8ToUtf16LeArrayListImpl(array_list: *std.ArrayList(u16), utf8: []const u8, comptime surrogates: Surrogates) !void {
-    // optimistically guess that it will not require surrogate pairs
-    try array_list.ensureTotalCapacityPrecise(utf8.len);
+fn utf8ToUtf16LeArrayListImpl(result: *std.ArrayList(u16), utf8: []const u8, comptime surrogates: Surrogates) !void {
+    assert(result.capacity >= utf8.len);
 
     var remaining = utf8;
-    // Need support for std.simd.interlace
-    if ((builtin.zig_backend != .stage2_x86_64 or
-        comptime (std.Target.x86.featureSetHas(builtin.cpu.features, .ssse3) and
-        !std.Target.x86.featureSetHasAny(builtin.cpu.features, .{ .prefer_256_bit, .avx }))) and
-        comptime !builtin.cpu.arch.isMIPS())
     vectorized: {
-        const chunk_len = @divExact(std.simd.suggestVectorLength(u8) orelse break :vectorized, 2);
+        const chunk_len = std.simd.suggestVectorLength(u16) orelse break :vectorized;
         const Chunk = @Vector(chunk_len, u8);
 
         // Fast path. Check for and encode ASCII characters at the start of the input.
@@ -1178,9 +1153,8 @@ fn utf8ToUtf16LeArrayListImpl(array_list: *std.ArrayList(u16), utf8: []const u8,
                 // found a non ASCII code unit
                 break;
             }
-            const zeroes: Chunk = @splat(0);
-            const utf16_chunk: [chunk_len * 2]u8 align(@alignOf(u16)) = std.simd.interlace(.{ chunk, zeroes });
-            array_list.appendSliceAssumeCapacity(std.mem.bytesAsSlice(u16, &utf16_chunk));
+            const utf16_chunk = mem.nativeToLittle(@Vector(chunk_len, u16), chunk);
+            result.addManyAsArrayAssumeCapacity(chunk_len).* = utf16_chunk;
             remaining = remaining[chunk_len..];
         }
     }
@@ -1192,21 +1166,18 @@ fn utf8ToUtf16LeArrayListImpl(array_list: *std.ArrayList(u16), utf8: []const u8,
     var it = view.iterator();
     while (it.nextCodepoint()) |codepoint| {
         if (codepoint < 0x10000) {
-            const short = @as(u16, @intCast(codepoint));
-            try array_list.append(mem.nativeToLittle(u16, short));
+            try result.append(mem.nativeToLittle(u16, @intCast(codepoint)));
         } else {
             const high = @as(u16, @intCast((codepoint - 0x10000) >> 10)) + 0xD800;
             const low = @as(u16, @intCast(codepoint & 0x3FF)) + 0xDC00;
-            var out: [2]u16 = undefined;
-            out[0] = mem.nativeToLittle(u16, high);
-            out[1] = mem.nativeToLittle(u16, low);
-            try array_list.appendSlice(out[0..]);
+            try result.appendSlice(&.{ mem.nativeToLittle(u16, high), mem.nativeToLittle(u16, low) });
         }
     }
 }
 
-pub fn utf8ToUtf16LeArrayList(array_list: *std.ArrayList(u16), utf8: []const u8) error{ InvalidUtf8, OutOfMemory }!void {
-    return utf8ToUtf16LeArrayListImpl(array_list, utf8, .cannot_encode_surrogate_half);
+pub fn utf8ToUtf16LeArrayList(result: *std.ArrayList(u16), utf8: []const u8) error{ InvalidUtf8, OutOfMemory }!void {
+    try result.ensureTotalCapacityPrecise(utf8.len);
+    return utf8ToUtf16LeArrayListImpl(result, utf8, .cannot_encode_surrogate_half);
 }
 
 pub fn utf8ToUtf16LeAlloc(allocator: mem.Allocator, utf8: []const u8) error{ InvalidUtf8, OutOfMemory }![]u16 {
@@ -1215,7 +1186,6 @@ pub fn utf8ToUtf16LeAlloc(allocator: mem.Allocator, utf8: []const u8) error{ Inv
     errdefer result.deinit();
 
     try utf8ToUtf16LeArrayListImpl(&result, utf8, .cannot_encode_surrogate_half);
-
     return result.toOwnedSlice();
 }
 
@@ -1228,7 +1198,6 @@ pub fn utf8ToUtf16LeAllocZ(allocator: mem.Allocator, utf8: []const u8) error{ In
     errdefer result.deinit();
 
     try utf8ToUtf16LeArrayListImpl(&result, utf8, .cannot_encode_surrogate_half);
-
     return result.toOwnedSliceSentinel(0);
 }
 
@@ -1239,16 +1208,11 @@ pub fn utf8ToUtf16Le(utf16le: []u16, utf8: []const u8) error{InvalidUtf8}!usize
 }
 
 pub fn utf8ToUtf16LeImpl(utf16le: []u16, utf8: []const u8, comptime surrogates: Surrogates) !usize {
-    var dest_i: usize = 0;
+    var dest_index: usize = 0;
 
     var remaining = utf8;
-    // Need support for std.simd.interlace
-    if ((builtin.zig_backend != .stage2_x86_64 or
-        comptime (std.Target.x86.featureSetHas(builtin.cpu.features, .ssse3) and
-        !std.Target.x86.featureSetHasAny(builtin.cpu.features, .{ .prefer_256_bit, .avx }))) and
-        comptime !builtin.cpu.arch.isMIPS())
     vectorized: {
-        const chunk_len = @divExact(std.simd.suggestVectorLength(u8) orelse break :vectorized, 2);
+        const chunk_len = std.simd.suggestVectorLength(u16) orelse break :vectorized;
         const Chunk = @Vector(chunk_len, u8);
 
         // Fast path. Check for and encode ASCII characters at the start of the input.
@@ -1259,57 +1223,60 @@ pub fn utf8ToUtf16LeImpl(utf16le: []u16, utf8: []const u8, comptime surrogates:
                 // found a non ASCII code unit
                 break;
             }
-            const zeroes: Chunk = @splat(0);
-            const utf16_bytes: [chunk_len * 2]u8 align(@alignOf(u16)) = std.simd.interlace(.{ chunk, zeroes });
-            @memcpy(utf16le[dest_i..][0..chunk_len], std.mem.bytesAsSlice(u16, &utf16_bytes));
-            dest_i += chunk_len;
+            const utf16_chunk = mem.nativeToLittle(@Vector(chunk_len, u16), chunk);
+            utf16le[dest_index..][0..chunk_len].* = utf16_chunk;
+            dest_index += chunk_len;
             remaining = remaining[chunk_len..];
         }
     }
 
-    var src_i: usize = 0;
-    while (src_i < remaining.len) {
-        const n = utf8ByteSequenceLength(remaining[src_i]) catch return switch (surrogates) {
-            .cannot_encode_surrogate_half => error.InvalidUtf8,
-            .can_encode_surrogate_half => error.InvalidWtf8,
-        };
-        const next_src_i = src_i + n;
-        const codepoint = switch (surrogates) {
-            .cannot_encode_surrogate_half => utf8Decode(remaining[src_i..next_src_i]) catch return error.InvalidUtf8,
-            .can_encode_surrogate_half => wtf8Decode(remaining[src_i..next_src_i]) catch return error.InvalidWtf8,
-        };
+    const view = switch (surrogates) {
+        .cannot_encode_surrogate_half => try Utf8View.init(remaining),
+        .can_encode_surrogate_half => try Wtf8View.init(remaining),
+    };
+    var it = view.iterator();
+    while (it.nextCodepoint()) |codepoint| {
         if (codepoint < 0x10000) {
-            const short = @as(u16, @intCast(codepoint));
-            utf16le[dest_i] = mem.nativeToLittle(u16, short);
-            dest_i += 1;
+            utf16le[dest_index] = mem.nativeToLittle(u16, @intCast(codepoint));
+            dest_index += 1;
         } else {
             const high = @as(u16, @intCast((codepoint - 0x10000) >> 10)) + 0xD800;
             const low = @as(u16, @intCast(codepoint & 0x3FF)) + 0xDC00;
-            utf16le[dest_i] = mem.nativeToLittle(u16, high);
-            utf16le[dest_i + 1] = mem.nativeToLittle(u16, low);
-            dest_i += 2;
+            utf16le[dest_index..][0..2].* = .{ mem.nativeToLittle(u16, high), mem.nativeToLittle(u16, low) };
+            dest_index += 2;
         }
-        src_i = next_src_i;
     }
-    return dest_i;
+    return dest_index;
 }
 
 test "utf8ToUtf16Le" {
-    var utf16le: [2]u16 = [_]u16{0} ** 2;
+    var utf16le: [128]u16 = undefined;
     {
         const length = try utf8ToUtf16Le(utf16le[0..], "𐐷");
-        try testing.expectEqual(@as(usize, 2), length);
-        try testing.expectEqualSlices(u8, "\x01\xd8\x37\xdc", mem.sliceAsBytes(utf16le[0..]));
+        try testing.expectEqualSlices(u8, "\x01\xd8\x37\xdc", mem.sliceAsBytes(utf16le[0..length]));
     }
     {
         const length = try utf8ToUtf16Le(utf16le[0..], "\u{10FFFF}");
-        try testing.expectEqual(@as(usize, 2), length);
-        try testing.expectEqualSlices(u8, "\xff\xdb\xff\xdf", mem.sliceAsBytes(utf16le[0..]));
+        try testing.expectEqualSlices(u8, "\xff\xdb\xff\xdf", mem.sliceAsBytes(utf16le[0..length]));
     }
     {
         const result = utf8ToUtf16Le(utf16le[0..], "\xf4\x90\x80\x80");
         try testing.expectError(error.InvalidUtf8, result);
     }
+    {
+        const length = try utf8ToUtf16Le(utf16le[0..], "This string has been designed to test the vectorized implementat" ++
+            "ion by beginning with one hundred twenty-seven ASCII characters¡");
+        try testing.expectEqualSlices(u8, &.{
+            'T', 0, 'h', 0, 'i', 0, 's', 0, ' ', 0, 's', 0, 't', 0, 'r', 0, 'i', 0, 'n', 0, 'g', 0, ' ', 0, 'h', 0, 'a', 0, 's', 0, ' ',  0,
+            'b', 0, 'e', 0, 'e', 0, 'n', 0, ' ', 0, 'd', 0, 'e', 0, 's', 0, 'i', 0, 'g', 0, 'n', 0, 'e', 0, 'd', 0, ' ', 0, 't', 0, 'o',  0,
+            ' ', 0, 't', 0, 'e', 0, 's', 0, 't', 0, ' ', 0, 't', 0, 'h', 0, 'e', 0, ' ', 0, 'v', 0, 'e', 0, 'c', 0, 't', 0, 'o', 0, 'r',  0,
+            'i', 0, 'z', 0, 'e', 0, 'd', 0, ' ', 0, 'i', 0, 'm', 0, 'p', 0, 'l', 0, 'e', 0, 'm', 0, 'e', 0, 'n', 0, 't', 0, 'a', 0, 't',  0,
+            'i', 0, 'o', 0, 'n', 0, ' ', 0, 'b', 0, 'y', 0, ' ', 0, 'b', 0, 'e', 0, 'g', 0, 'i', 0, 'n', 0, 'n', 0, 'i', 0, 'n', 0, 'g',  0,
+            ' ', 0, 'w', 0, 'i', 0, 't', 0, 'h', 0, ' ', 0, 'o', 0, 'n', 0, 'e', 0, ' ', 0, 'h', 0, 'u', 0, 'n', 0, 'd', 0, 'r', 0, 'e',  0,
+            'd', 0, ' ', 0, 't', 0, 'w', 0, 'e', 0, 'n', 0, 't', 0, 'y', 0, '-', 0, 's', 0, 'e', 0, 'v', 0, 'e', 0, 'n', 0, ' ', 0, 'A',  0,
+            'S', 0, 'C', 0, 'I', 0, 'I', 0, ' ', 0, 'c', 0, 'h', 0, 'a', 0, 'r', 0, 'a', 0, 'c', 0, 't', 0, 'e', 0, 'r', 0, 's', 0, '¡', 0,
+        }, mem.sliceAsBytes(utf16le[0..length]));
+    }
 }
 
 test utf8ToUtf16LeArrayList {
@@ -1354,25 +1321,40 @@ test utf8ToUtf16LeAllocZ {
     {
         const utf16 = try utf8ToUtf16LeAllocZ(testing.allocator, "𐐷");
         defer testing.allocator.free(utf16);
-        try testing.expectEqualSlices(u8, "\x01\xd8\x37\xdc", mem.sliceAsBytes(utf16[0..]));
+        try testing.expectEqualSlices(u8, "\x01\xd8\x37\xdc", mem.sliceAsBytes(utf16));
         try testing.expect(utf16[2] == 0);
     }
     {
         const utf16 = try utf8ToUtf16LeAllocZ(testing.allocator, "\u{10FFFF}");
         defer testing.allocator.free(utf16);
-        try testing.expectEqualSlices(u8, "\xff\xdb\xff\xdf", mem.sliceAsBytes(utf16[0..]));
+        try testing.expectEqualSlices(u8, "\xff\xdb\xff\xdf", mem.sliceAsBytes(utf16));
         try testing.expect(utf16[2] == 0);
     }
     {
         const result = utf8ToUtf16LeAllocZ(testing.allocator, "\xf4\x90\x80\x80");
         try testing.expectError(error.InvalidUtf8, result);
     }
+    {
+        const utf16 = try utf8ToUtf16LeWithNull(testing.allocator, "This string has been designed to test the vectorized implementat" ++
+            "ion by beginning with one hundred twenty-seven ASCII characters¡");
+        defer testing.allocator.free(utf16);
+        try testing.expectEqualSlices(u8, &.{
+            'T', 0, 'h', 0, 'i', 0, 's', 0, ' ', 0, 's', 0, 't', 0, 'r', 0, 'i', 0, 'n', 0, 'g', 0, ' ', 0, 'h', 0, 'a', 0, 's', 0, ' ',  0,
+            'b', 0, 'e', 0, 'e', 0, 'n', 0, ' ', 0, 'd', 0, 'e', 0, 's', 0, 'i', 0, 'g', 0, 'n', 0, 'e', 0, 'd', 0, ' ', 0, 't', 0, 'o',  0,
+            ' ', 0, 't', 0, 'e', 0, 's', 0, 't', 0, ' ', 0, 't', 0, 'h', 0, 'e', 0, ' ', 0, 'v', 0, 'e', 0, 'c', 0, 't', 0, 'o', 0, 'r',  0,
+            'i', 0, 'z', 0, 'e', 0, 'd', 0, ' ', 0, 'i', 0, 'm', 0, 'p', 0, 'l', 0, 'e', 0, 'm', 0, 'e', 0, 'n', 0, 't', 0, 'a', 0, 't',  0,
+            'i', 0, 'o', 0, 'n', 0, ' ', 0, 'b', 0, 'y', 0, ' ', 0, 'b', 0, 'e', 0, 'g', 0, 'i', 0, 'n', 0, 'n', 0, 'i', 0, 'n', 0, 'g',  0,
+            ' ', 0, 'w', 0, 'i', 0, 't', 0, 'h', 0, ' ', 0, 'o', 0, 'n', 0, 'e', 0, ' ', 0, 'h', 0, 'u', 0, 'n', 0, 'd', 0, 'r', 0, 'e',  0,
+            'd', 0, ' ', 0, 't', 0, 'w', 0, 'e', 0, 'n', 0, 't', 0, 'y', 0, '-', 0, 's', 0, 'e', 0, 'v', 0, 'e', 0, 'n', 0, ' ', 0, 'A',  0,
+            'S', 0, 'C', 0, 'I', 0, 'I', 0, ' ', 0, 'c', 0, 'h', 0, 'a', 0, 'r', 0, 'a', 0, 'c', 0, 't', 0, 'e', 0, 'r', 0, 's', 0, '¡', 0,
+        }, mem.sliceAsBytes(utf16));
+    }
 }
 
 /// Converts a UTF-8 string literal into a UTF-16LE string literal.
-pub fn utf8ToUtf16LeStringLiteral(comptime utf8: []const u8) *const [calcUtf16LeLen(utf8) catch unreachable:0]u16 {
+pub fn utf8ToUtf16LeStringLiteral(comptime utf8: []const u8) *const [calcUtf16LeLen(utf8) catch |err| @compileError(err):0]u16 {
     return comptime blk: {
-        const len: usize = calcUtf16LeLen(utf8) catch |err| @compileError(err);
+        const len: usize = calcUtf16LeLen(utf8) catch unreachable;
         var utf16le: [len:0]u16 = [_:0]u16{0} ** len;
         const utf16le_len = utf8ToUtf16Le(&utf16le, utf8[0..]) catch |err| @compileError(err);
         assert(len == utf16le_len);
@@ -1453,12 +1435,12 @@ test "fmtUtf16Le" {
     try expectFmt("", "{}", .{fmtUtf16Le(utf8ToUtf16LeStringLiteral(""))});
     try expectFmt("foo", "{}", .{fmtUtf16Le(utf8ToUtf16LeStringLiteral("foo"))});
     try expectFmt("𐐷", "{}", .{fmtUtf16Le(utf8ToUtf16LeStringLiteral("𐐷"))});
-    try expectFmt("퟿", "{}", .{fmtUtf16Le(&[_]u16{std.mem.readInt(u16, "\xff\xd7", native_endian)})});
-    try expectFmt("�", "{}", .{fmtUtf16Le(&[_]u16{std.mem.readInt(u16, "\x00\xd8", native_endian)})});
-    try expectFmt("�", "{}", .{fmtUtf16Le(&[_]u16{std.mem.readInt(u16, "\xff\xdb", native_endian)})});
-    try expectFmt("�", "{}", .{fmtUtf16Le(&[_]u16{std.mem.readInt(u16, "\x00\xdc", native_endian)})});
-    try expectFmt("�", "{}", .{fmtUtf16Le(&[_]u16{std.mem.readInt(u16, "\xff\xdf", native_endian)})});
-    try expectFmt("", "{}", .{fmtUtf16Le(&[_]u16{std.mem.readInt(u16, "\x00\xe0", native_endian)})});
+    try expectFmt("퟿", "{}", .{fmtUtf16Le(&[_]u16{mem.readInt(u16, "\xff\xd7", native_endian)})});
+    try expectFmt("�", "{}", .{fmtUtf16Le(&[_]u16{mem.readInt(u16, "\x00\xd8", native_endian)})});
+    try expectFmt("�", "{}", .{fmtUtf16Le(&[_]u16{mem.readInt(u16, "\xff\xdb", native_endian)})});
+    try expectFmt("�", "{}", .{fmtUtf16Le(&[_]u16{mem.readInt(u16, "\x00\xdc", native_endian)})});
+    try expectFmt("�", "{}", .{fmtUtf16Le(&[_]u16{mem.readInt(u16, "\xff\xdf", native_endian)})});
+    try expectFmt("", "{}", .{fmtUtf16Le(&[_]u16{mem.readInt(u16, "\x00\xe0", native_endian)})});
 }
 
 test "utf8ToUtf16LeStringLiteral" {
@@ -1701,8 +1683,9 @@ pub const Wtf8Iterator = struct {
     }
 };
 
-pub fn wtf16LeToWtf8ArrayList(array_list: *std.ArrayList(u8), utf16le: []const u16) mem.Allocator.Error!void {
-    return utf16LeToUtf8ArrayListImpl(array_list, utf16le, .can_encode_surrogate_half);
+pub fn wtf16LeToWtf8ArrayList(result: *std.ArrayList(u8), utf16le: []const u16) mem.Allocator.Error!void {
+    try result.ensureTotalCapacityPrecise(utf16le.len);
+    return utf16LeToUtf8ArrayListImpl(result, utf16le, .can_encode_surrogate_half);
 }
 
 /// Caller must free returned memory.
@@ -1711,8 +1694,7 @@ pub fn wtf16LeToWtf8Alloc(allocator: mem.Allocator, wtf16le: []const u16) mem.Al
     var result = try std.ArrayList(u8).initCapacity(allocator, wtf16le.len);
     errdefer result.deinit();
 
-    try wtf16LeToWtf8ArrayList(&result, wtf16le);
-
+    try utf16LeToUtf8ArrayListImpl(&result, wtf16le, .can_encode_surrogate_half);
     return result.toOwnedSlice();
 }
 
@@ -1722,8 +1704,7 @@ pub fn wtf16LeToWtf8AllocZ(allocator: mem.Allocator, wtf16le: []const u16) mem.A
     var result = try std.ArrayList(u8).initCapacity(allocator, wtf16le.len + 1);
     errdefer result.deinit();
 
-    try wtf16LeToWtf8ArrayList(&result, wtf16le);
-
+    try utf16LeToUtf8ArrayListImpl(&result, wtf16le, .can_encode_surrogate_half);
     return result.toOwnedSliceSentinel(0);
 }
 
@@ -1731,8 +1712,9 @@ pub fn wtf16LeToWtf8(wtf8: []u8, wtf16le: []const u16) usize {
     return utf16LeToUtf8Impl(wtf8, wtf16le, .can_encode_surrogate_half) catch |err| switch (err) {};
 }
 
-pub fn wtf8ToWtf16LeArrayList(array_list: *std.ArrayList(u16), wtf8: []const u8) error{ InvalidWtf8, OutOfMemory }!void {
-    return utf8ToUtf16LeArrayListImpl(array_list, wtf8, .can_encode_surrogate_half);
+pub fn wtf8ToWtf16LeArrayList(result: *std.ArrayList(u16), wtf8: []const u8) error{ InvalidWtf8, OutOfMemory }!void {
+    try result.ensureTotalCapacityPrecise(wtf8.len);
+    return utf8ToUtf16LeArrayListImpl(result, wtf8, .can_encode_surrogate_half);
 }
 
 pub fn wtf8ToWtf16LeAlloc(allocator: mem.Allocator, wtf8: []const u8) error{ InvalidWtf8, OutOfMemory }![]u16 {
@@ -1741,7 +1723,6 @@ pub fn wtf8ToWtf16LeAlloc(allocator: mem.Allocator, wtf8: []const u8) error{ Inv
     errdefer result.deinit();
 
     try utf8ToUtf16LeArrayListImpl(&result, wtf8, .can_encode_surrogate_half);
-
     return result.toOwnedSlice();
 }
 
@@ -1751,7 +1732,6 @@ pub fn wtf8ToWtf16LeAllocZ(allocator: mem.Allocator, wtf8: []const u8) error{ In
     errdefer result.deinit();
 
     try utf8ToUtf16LeArrayListImpl(&result, wtf8, .can_encode_surrogate_half);
-
     return result.toOwnedSliceSentinel(0);
 }
 
@@ -1910,7 +1890,7 @@ pub const Wtf16LeIterator = struct {
 
     pub fn init(s: []const u16) Wtf16LeIterator {
         return Wtf16LeIterator{
-            .bytes = std.mem.sliceAsBytes(s),
+            .bytes = mem.sliceAsBytes(s),
             .i = 0,
         };
     }
@@ -1923,12 +1903,12 @@ pub const Wtf16LeIterator = struct {
         assert(it.i <= it.bytes.len);
         if (it.i == it.bytes.len) return null;
         var code_units: [2]u16 = undefined;
-        code_units[0] = std.mem.readInt(u16, it.bytes[it.i..][0..2], .little);
+        code_units[0] = mem.readInt(u16, it.bytes[it.i..][0..2], .little);
         it.i += 2;
         surrogate_pair: {
             if (utf16IsHighSurrogate(code_units[0])) {
                 if (it.i >= it.bytes.len) break :surrogate_pair;
-                code_units[1] = std.mem.readInt(u16, it.bytes[it.i..][0..2], .little);
+                code_units[1] = mem.readInt(u16, it.bytes[it.i..][0..2], .little);
                 const codepoint = utf16DecodeSurrogatePair(&code_units) catch break :surrogate_pair;
                 it.i += 2;
                 return codepoint;
@@ -2045,31 +2025,31 @@ fn testRoundtripWtf16(wtf16le: []const u16) !void {
 
 test "well-formed WTF-16 roundtrips" {
     try testRoundtripWtf16(&[_]u16{
-        std.mem.nativeToLittle(u16, 0xD83D), // high surrogate
-        std.mem.nativeToLittle(u16, 0xDCA9), // low surrogate
+        mem.nativeToLittle(u16, 0xD83D), // high surrogate
+        mem.nativeToLittle(u16, 0xDCA9), // low surrogate
     });
     try testRoundtripWtf16(&[_]u16{
-        std.mem.nativeToLittle(u16, 0xD83D), // high surrogate
-        std.mem.nativeToLittle(u16, ' '), // not surrogate
-        std.mem.nativeToLittle(u16, 0xDCA9), // low surrogate
+        mem.nativeToLittle(u16, 0xD83D), // high surrogate
+        mem.nativeToLittle(u16, ' '), // not surrogate
+        mem.nativeToLittle(u16, 0xDCA9), // low surrogate
     });
     try testRoundtripWtf16(&[_]u16{
-        std.mem.nativeToLittle(u16, 0xD800), // high surrogate
-        std.mem.nativeToLittle(u16, 0xDBFF), // high surrogate
+        mem.nativeToLittle(u16, 0xD800), // high surrogate
+        mem.nativeToLittle(u16, 0xDBFF), // high surrogate
     });
     try testRoundtripWtf16(&[_]u16{
-        std.mem.nativeToLittle(u16, 0xD800), // high surrogate
-        std.mem.nativeToLittle(u16, 0xE000), // not surrogate
+        mem.nativeToLittle(u16, 0xD800), // high surrogate
+        mem.nativeToLittle(u16, 0xE000), // not surrogate
     });
     try testRoundtripWtf16(&[_]u16{
-        std.mem.nativeToLittle(u16, 0xD7FF), // not surrogate
-        std.mem.nativeToLittle(u16, 0xDC00), // low surrogate
+        mem.nativeToLittle(u16, 0xD7FF), // not surrogate
+        mem.nativeToLittle(u16, 0xDC00), // low surrogate
     });
     try testRoundtripWtf16(&[_]u16{
-        std.mem.nativeToLittle(u16, 0x61), // not surrogate
-        std.mem.nativeToLittle(u16, 0xDC00), // low surrogate
+        mem.nativeToLittle(u16, 0x61), // not surrogate
+        mem.nativeToLittle(u16, 0xDC00), // low surrogate
     });
     try testRoundtripWtf16(&[_]u16{
-        std.mem.nativeToLittle(u16, 0xDC00), // low surrogate
+        mem.nativeToLittle(u16, 0xDC00), // low surrogate
     });
 }
src/arch/x86_64/CodeGen.zig
@@ -2853,11 +2853,14 @@ fn airFptrunc(self: *Self, inst: Air.Inst.Index) !void {
 }
 
 fn airFpext(self: *Self, inst: Air.Inst.Index) !void {
+    const mod = self.bin_file.comp.module.?;
     const ty_op = self.air.instructions.items(.data)[@intFromEnum(inst)].ty_op;
     const dst_ty = self.typeOfIndex(inst);
-    const dst_bits = dst_ty.floatBits(self.target.*);
+    const dst_scalar_ty = dst_ty.scalarType(mod);
+    const dst_bits = dst_scalar_ty.floatBits(self.target.*);
     const src_ty = self.typeOf(ty_op.operand);
-    const src_bits = src_ty.floatBits(self.target.*);
+    const src_scalar_ty = src_ty.scalarType(mod);
+    const src_bits = src_scalar_ty.floatBits(self.target.*);
 
     const result = result: {
         if (switch (src_bits) {
@@ -2881,94 +2884,290 @@ fn airFpext(self: *Self, inst: Air.Inst.Index) !void {
             },
             else => unreachable,
         }) {
+            if (dst_ty.isVector(mod)) break :result null;
             var callee_buf: ["__extend?f?f2".len]u8 = undefined;
             break :result try self.genCall(.{ .lib = .{
-                .return_type = self.floatCompilerRtAbiType(dst_ty, src_ty).toIntern(),
-                .param_types = &.{self.floatCompilerRtAbiType(src_ty, dst_ty).toIntern()},
+                .return_type = self.floatCompilerRtAbiType(dst_scalar_ty, src_scalar_ty).toIntern(),
+                .param_types = &.{self.floatCompilerRtAbiType(src_scalar_ty, dst_scalar_ty).toIntern()},
                 .callee = std.fmt.bufPrint(&callee_buf, "__extend{c}f{c}f2", .{
                     floatCompilerRtAbiName(src_bits),
                     floatCompilerRtAbiName(dst_bits),
                 }) catch unreachable,
-            } }, &.{src_ty}, &.{.{ .air_ref = ty_op.operand }});
+            } }, &.{src_scalar_ty}, &.{.{ .air_ref = ty_op.operand }});
         }
 
+        const src_abi_size: u32 = @intCast(src_ty.abiSize(mod));
         const src_mcv = try self.resolveInst(ty_op.operand);
         const dst_mcv = if (src_mcv.isRegister() and self.reuseOperand(inst, ty_op.operand, 0, src_mcv))
             src_mcv
         else
             try self.copyToRegisterWithInstTracking(inst, dst_ty, src_mcv);
-        const dst_reg = dst_mcv.getReg().?.to128();
+        const dst_reg = dst_mcv.getReg().?;
+        const dst_alias = registerAlias(dst_reg, @intCast(@max(dst_ty.abiSize(mod), 16)));
         const dst_lock = self.register_manager.lockReg(dst_reg);
         defer if (dst_lock) |lock| self.register_manager.unlockReg(lock);
 
+        const vec_len = if (dst_ty.isVector(mod)) dst_ty.vectorLen(mod) else 1;
         if (src_bits == 16) {
             assert(self.hasFeature(.f16c));
             const mat_src_reg = if (src_mcv.isRegister())
                 src_mcv.getReg().?
             else
                 try self.copyToTmpRegister(src_ty, src_mcv);
-            try self.asmRegisterRegister(.{ .v_ps, .cvtph2 }, dst_reg, mat_src_reg.to128());
+            try self.asmRegisterRegister(
+                .{ .v_ps, .cvtph2 },
+                dst_alias,
+                registerAlias(mat_src_reg, src_abi_size),
+            );
             switch (dst_bits) {
                 32 => {},
                 64 => try self.asmRegisterRegisterRegister(
                     .{ .v_sd, .cvtss2 },
-                    dst_reg,
-                    dst_reg,
-                    dst_reg,
+                    dst_alias,
+                    dst_alias,
+                    dst_alias,
                 ),
                 else => unreachable,
             }
         } else {
             assert(src_bits == 32 and dst_bits == 64);
-            if (self.hasFeature(.avx)) if (src_mcv.isMemory()) try self.asmRegisterRegisterMemory(
-                .{ .v_sd, .cvtss2 },
-                dst_reg,
-                dst_reg,
-                try src_mcv.mem(self, .dword),
-            ) else try self.asmRegisterRegisterRegister(
-                .{ .v_sd, .cvtss2 },
-                dst_reg,
-                dst_reg,
-                (if (src_mcv.isRegister())
-                    src_mcv.getReg().?
-                else
-                    try self.copyToTmpRegister(src_ty, src_mcv)).to128(),
-            ) else if (src_mcv.isMemory()) try self.asmRegisterMemory(
-                .{ ._sd, .cvtss2 },
-                dst_reg,
-                try src_mcv.mem(self, .dword),
+            if (self.hasFeature(.avx)) switch (vec_len) {
+                1 => if (src_mcv.isMemory()) try self.asmRegisterRegisterMemory(
+                    .{ .v_sd, .cvtss2 },
+                    dst_alias,
+                    dst_alias,
+                    try src_mcv.mem(self, self.memSize(src_ty)),
+                ) else try self.asmRegisterRegisterRegister(
+                    .{ .v_sd, .cvtss2 },
+                    dst_alias,
+                    dst_alias,
+                    registerAlias(if (src_mcv.isRegister())
+                        src_mcv.getReg().?
+                    else
+                        try self.copyToTmpRegister(src_ty, src_mcv), src_abi_size),
+                ),
+                2...4 => if (src_mcv.isMemory()) try self.asmRegisterMemory(
+                    .{ .v_pd, .cvtps2 },
+                    dst_alias,
+                    try src_mcv.mem(self, self.memSize(src_ty)),
+                ) else try self.asmRegisterRegister(
+                    .{ .v_pd, .cvtps2 },
+                    dst_alias,
+                    registerAlias(if (src_mcv.isRegister())
+                        src_mcv.getReg().?
+                    else
+                        try self.copyToTmpRegister(src_ty, src_mcv), src_abi_size),
+                ),
+                else => break :result null,
+            } else if (src_mcv.isMemory()) try self.asmRegisterMemory(
+                switch (vec_len) {
+                    1 => .{ ._sd, .cvtss2 },
+                    2 => .{ ._pd, .cvtps2 },
+                    else => break :result null,
+                },
+                dst_alias,
+                try src_mcv.mem(self, self.memSize(src_ty)),
             ) else try self.asmRegisterRegister(
-                .{ ._sd, .cvtss2 },
-                dst_reg,
-                (if (src_mcv.isRegister())
+                switch (vec_len) {
+                    1 => .{ ._sd, .cvtss2 },
+                    2 => .{ ._pd, .cvtps2 },
+                    else => break :result null,
+                },
+                dst_alias,
+                registerAlias(if (src_mcv.isRegister())
                     src_mcv.getReg().?
                 else
-                    try self.copyToTmpRegister(src_ty, src_mcv)).to128(),
+                    try self.copyToTmpRegister(src_ty, src_mcv), src_abi_size),
             );
         }
         break :result dst_mcv;
-    };
+    } orelse return self.fail("TODO implement airFpext from {} to {}", .{
+        src_ty.fmt(mod), dst_ty.fmt(mod),
+    });
     return self.finishAir(inst, result, .{ ty_op.operand, .none, .none });
 }
 
 fn airIntCast(self: *Self, inst: Air.Inst.Index) !void {
     const mod = self.bin_file.comp.module.?;
     const ty_op = self.air.instructions.items(.data)[@intFromEnum(inst)].ty_op;
-    const result: MCValue = result: {
-        const src_ty = self.typeOf(ty_op.operand);
-        const src_int_info = src_ty.intInfo(mod);
+    const src_ty = self.typeOf(ty_op.operand);
+    const dst_ty = self.typeOfIndex(inst);
 
-        const dst_ty = self.typeOfIndex(inst);
-        const dst_int_info = dst_ty.intInfo(mod);
-        const abi_size: u32 = @intCast(dst_ty.abiSize(mod));
+    const result = @as(?MCValue, result: {
+        const dst_abi_size: u32 = @intCast(dst_ty.abiSize(mod));
 
-        const min_ty = if (dst_int_info.bits < src_int_info.bits) dst_ty else src_ty;
+        const src_int_info = src_ty.intInfo(mod);
+        const dst_int_info = dst_ty.intInfo(mod);
         const extend = switch (src_int_info.signedness) {
             .signed => dst_int_info,
             .unsigned => src_int_info,
         }.signedness;
 
         const src_mcv = try self.resolveInst(ty_op.operand);
+        if (dst_ty.isVector(mod)) {
+            const src_abi_size: u32 = @intCast(src_ty.abiSize(mod));
+            const max_abi_size = @max(dst_abi_size, src_abi_size);
+            if (max_abi_size > @as(u32, if (self.hasFeature(.avx2)) 32 else 16)) break :result null;
+            const has_avx = self.hasFeature(.avx);
+
+            const dst_elem_abi_size = dst_ty.childType(mod).abiSize(mod);
+            const src_elem_abi_size = src_ty.childType(mod).abiSize(mod);
+            switch (math.order(dst_elem_abi_size, src_elem_abi_size)) {
+                .lt => {
+                    const mir_tag: Mir.Inst.FixedTag = switch (dst_elem_abi_size) {
+                        else => break :result null,
+                        1 => switch (src_elem_abi_size) {
+                            else => break :result null,
+                            2 => switch (dst_int_info.signedness) {
+                                .signed => if (has_avx) .{ .vp_b, .ackssw } else .{ .p_b, .ackssw },
+                                .unsigned => if (has_avx) .{ .vp_b, .ackusw } else .{ .p_b, .ackusw },
+                            },
+                        },
+                        2 => switch (src_elem_abi_size) {
+                            else => break :result null,
+                            4 => switch (dst_int_info.signedness) {
+                                .signed => if (has_avx) .{ .vp_w, .ackssd } else .{ .p_w, .ackssd },
+                                .unsigned => if (has_avx)
+                                    .{ .vp_w, .ackusd }
+                                else if (self.hasFeature(.sse4_1))
+                                    .{ .p_w, .ackusd }
+                                else
+                                    break :result null,
+                            },
+                        },
+                    };
+
+                    const dst_mcv: MCValue = if (src_mcv.isRegister() and
+                        self.reuseOperand(inst, ty_op.operand, 0, src_mcv))
+                        src_mcv
+                    else if (has_avx and src_mcv.isRegister())
+                        .{ .register = try self.register_manager.allocReg(inst, abi.RegisterClass.sse) }
+                    else
+                        try self.copyToRegisterWithInstTracking(inst, src_ty, src_mcv);
+                    const dst_reg = dst_mcv.getReg().?;
+                    const dst_alias = registerAlias(dst_reg, dst_abi_size);
+
+                    if (has_avx) try self.asmRegisterRegisterRegister(
+                        mir_tag,
+                        dst_alias,
+                        registerAlias(if (src_mcv.isRegister())
+                            src_mcv.getReg().?
+                        else
+                            dst_reg, src_abi_size),
+                        dst_alias,
+                    ) else try self.asmRegisterRegister(
+                        mir_tag,
+                        dst_alias,
+                        dst_alias,
+                    );
+                    break :result dst_mcv;
+                },
+                .eq => if (self.reuseOperand(inst, ty_op.operand, 0, src_mcv))
+                    break :result src_mcv
+                else {
+                    const dst_mcv = try self.allocRegOrMem(inst, true);
+                    try self.genCopy(dst_ty, dst_mcv, src_mcv, .{});
+                    break :result dst_mcv;
+                },
+                .gt => if (self.hasFeature(.sse4_1)) {
+                    const mir_tag: Mir.Inst.FixedTag = .{ switch (dst_elem_abi_size) {
+                        else => break :result null,
+                        2 => if (has_avx) .vp_w else .p_w,
+                        4 => if (has_avx) .vp_d else .p_d,
+                        8 => if (has_avx) .vp_q else .p_q,
+                    }, switch (src_elem_abi_size) {
+                        else => break :result null,
+                        1 => switch (extend) {
+                            .signed => .movsxb,
+                            .unsigned => .movzxb,
+                        },
+                        2 => switch (extend) {
+                            .signed => .movsxw,
+                            .unsigned => .movzxw,
+                        },
+                        4 => switch (extend) {
+                            .signed => .movsxd,
+                            .unsigned => .movzxd,
+                        },
+                    } };
+
+                    const dst_mcv: MCValue = if (src_mcv.isRegister() and
+                        self.reuseOperand(inst, ty_op.operand, 0, src_mcv))
+                        src_mcv
+                    else
+                        .{ .register = try self.register_manager.allocReg(inst, abi.RegisterClass.sse) };
+                    const dst_reg = dst_mcv.getReg().?;
+                    const dst_alias = registerAlias(dst_reg, dst_abi_size);
+
+                    if (src_mcv.isMemory()) try self.asmRegisterMemory(
+                        mir_tag,
+                        dst_alias,
+                        try src_mcv.mem(self, self.memSize(src_ty)),
+                    ) else try self.asmRegisterRegister(
+                        mir_tag,
+                        dst_alias,
+                        registerAlias(if (src_mcv.isRegister())
+                            src_mcv.getReg().?
+                        else
+                            try self.copyToTmpRegister(src_ty, src_mcv), src_abi_size),
+                    );
+                    break :result dst_mcv;
+                } else {
+                    const mir_tag: Mir.Inst.FixedTag = switch (dst_elem_abi_size) {
+                        else => break :result null,
+                        2 => switch (src_elem_abi_size) {
+                            else => break :result null,
+                            1 => .{ .p_, .unpcklbw },
+                        },
+                        4 => switch (src_elem_abi_size) {
+                            else => break :result null,
+                            2 => .{ .p_, .unpcklwd },
+                        },
+                        8 => switch (src_elem_abi_size) {
+                            else => break :result null,
+                            2 => .{ .p_, .unpckldq },
+                        },
+                    };
+
+                    const dst_mcv: MCValue = if (src_mcv.isRegister() and
+                        self.reuseOperand(inst, ty_op.operand, 0, src_mcv))
+                        src_mcv
+                    else
+                        try self.copyToRegisterWithInstTracking(inst, dst_ty, src_mcv);
+                    const dst_reg = dst_mcv.getReg().?;
+
+                    const ext_reg = try self.register_manager.allocReg(null, abi.RegisterClass.sse);
+                    const ext_alias = registerAlias(ext_reg, src_abi_size);
+                    const ext_lock = self.register_manager.lockRegAssumeUnused(ext_reg);
+                    defer self.register_manager.unlockReg(ext_lock);
+
+                    try self.asmRegisterRegister(.{ .p_, .xor }, ext_alias, ext_alias);
+                    switch (extend) {
+                        .signed => try self.asmRegisterRegister(
+                            .{ switch (src_elem_abi_size) {
+                                else => unreachable,
+                                1 => .p_b,
+                                2 => .p_w,
+                                4 => .p_d,
+                            }, .cmpgt },
+                            ext_alias,
+                            registerAlias(dst_reg, src_abi_size),
+                        ),
+                        .unsigned => {},
+                    }
+                    try self.asmRegisterRegister(
+                        mir_tag,
+                        registerAlias(dst_reg, dst_abi_size),
+                        registerAlias(ext_reg, dst_abi_size),
+                    );
+                    break :result dst_mcv;
+                },
+            }
+            @compileError("unreachable");
+        }
+
+        const min_ty = if (dst_int_info.bits < src_int_info.bits) dst_ty else src_ty;
+
         const src_storage_bits: u16 = switch (src_mcv) {
             .register, .register_offset => 64,
             .register_pair => 128,
@@ -2986,13 +3185,13 @@ fn airIntCast(self: *Self, inst: Air.Inst.Index) !void {
         };
 
         if (dst_int_info.bits <= src_int_info.bits) break :result if (dst_mcv.isRegister())
-            .{ .register = registerAlias(dst_mcv.getReg().?, abi_size) }
+            .{ .register = registerAlias(dst_mcv.getReg().?, dst_abi_size) }
         else
             dst_mcv;
 
         if (dst_mcv.isRegister()) {
             try self.truncateRegister(src_ty, dst_mcv.getReg().?);
-            break :result .{ .register = registerAlias(dst_mcv.getReg().?, abi_size) };
+            break :result .{ .register = registerAlias(dst_mcv.getReg().?, dst_abi_size) };
         }
 
         const src_limbs_len = math.divCeil(u16, src_int_info.bits, 64) catch unreachable;
@@ -3040,7 +3239,9 @@ fn airIntCast(self: *Self, inst: Air.Inst.Index) !void {
         );
 
         break :result dst_mcv;
-    };
+    }) orelse return self.fail("TODO implement airIntCast from {} to {}", .{
+        src_ty.fmt(mod), dst_ty.fmt(mod),
+    });
     return self.finishAir(inst, result, .{ ty_op.operand, .none, .none });
 }
 
@@ -3063,7 +3264,7 @@ fn airTrunc(self: *Self, inst: Air.Inst.Index) !void {
             src_mcv
         else if (dst_abi_size <= 8)
             try self.copyToRegisterWithInstTracking(inst, dst_ty, src_mcv)
-        else if (dst_abi_size <= 16) dst: {
+        else if (dst_abi_size <= 16 and !dst_ty.isVector(mod)) dst: {
             const dst_regs =
                 try self.register_manager.allocRegs(2, .{ inst, inst }, abi.RegisterClass.gp);
             const dst_mcv: MCValue = .{ .register_pair = dst_regs };
@@ -3080,19 +3281,22 @@ fn airTrunc(self: *Self, inst: Air.Inst.Index) !void {
 
         if (dst_ty.zigTypeTag(mod) == .Vector) {
             assert(src_ty.zigTypeTag(mod) == .Vector and dst_ty.vectorLen(mod) == src_ty.vectorLen(mod));
-            const dst_info = dst_ty.childType(mod).intInfo(mod);
-            const src_info = src_ty.childType(mod).intInfo(mod);
-            const mir_tag = @as(?Mir.Inst.FixedTag, switch (dst_info.bits) {
-                8 => switch (src_info.bits) {
-                    16 => switch (dst_ty.vectorLen(mod)) {
+            const dst_elem_ty = dst_ty.childType(mod);
+            const dst_elem_abi_size: u32 = @intCast(dst_elem_ty.abiSize(mod));
+            const src_elem_ty = src_ty.childType(mod);
+            const src_elem_abi_size: u32 = @intCast(src_elem_ty.abiSize(mod));
+
+            const mir_tag = @as(?Mir.Inst.FixedTag, switch (dst_elem_abi_size) {
+                1 => switch (src_elem_abi_size) {
+                    2 => switch (dst_ty.vectorLen(mod)) {
                         1...8 => if (self.hasFeature(.avx)) .{ .vp_b, .ackusw } else .{ .p_b, .ackusw },
                         9...16 => if (self.hasFeature(.avx2)) .{ .vp_b, .ackusw } else null,
                         else => null,
                     },
                     else => null,
                 },
-                16 => switch (src_info.bits) {
-                    32 => switch (dst_ty.vectorLen(mod)) {
+                2 => switch (src_elem_abi_size) {
+                    4 => switch (dst_ty.vectorLen(mod)) {
                         1...4 => if (self.hasFeature(.avx))
                             .{ .vp_w, .ackusd }
                         else if (self.hasFeature(.sse4_1))
@@ -3107,12 +3311,14 @@ fn airTrunc(self: *Self, inst: Air.Inst.Index) !void {
                 else => null,
             }) orelse return self.fail("TODO implement airTrunc for {}", .{dst_ty.fmt(mod)});
 
-            const elem_ty = src_ty.childType(mod);
-            const mask_val = try mod.intValue(elem_ty, @as(u64, math.maxInt(u64)) >> @intCast(64 - dst_info.bits));
+            const dst_info = dst_elem_ty.intInfo(mod);
+            const src_info = src_elem_ty.intInfo(mod);
+
+            const mask_val = try mod.intValue(src_elem_ty, @as(u64, math.maxInt(u64)) >> @intCast(64 - dst_info.bits));
 
             const splat_ty = try mod.vectorType(.{
                 .len = @intCast(@divExact(@as(u64, if (src_abi_size > 16) 256 else 128), src_info.bits)),
-                .child = elem_ty.ip_index,
+                .child = src_elem_ty.ip_index,
             });
             const splat_abi_size: u32 = @intCast(splat_ty.abiSize(mod));
 
@@ -4086,7 +4292,7 @@ fn airMulWithOverflow(self: *Self, inst: Air.Inst.Index) !void {
             if (dst_info.bits > 128 and dst_info.signedness == .unsigned) {
                 const slow_inc = self.hasFeature(.slow_incdec);
                 const abi_size: u32 = @intCast(dst_ty.abiSize(mod));
-                const limb_len = std.math.divCeil(u32, abi_size, 8) catch unreachable;
+                const limb_len = math.divCeil(u32, abi_size, 8) catch unreachable;
 
                 try self.spillRegisters(&.{ .rax, .rcx, .rdx });
                 const reg_locks = self.register_manager.lockRegsAssumeUnused(3, .{ .rax, .rcx, .rdx });
@@ -6935,7 +7141,7 @@ fn airAbs(self: *Self, inst: Air.Inst.Index) !void {
                 },
                 else => {
                     const abi_size: u31 = @intCast(ty.abiSize(mod));
-                    const limb_len = std.math.divCeil(u31, abi_size, 8) catch unreachable;
+                    const limb_len = math.divCeil(u31, abi_size, 8) catch unreachable;
 
                     const tmp_regs =
                         try self.register_manager.allocRegs(3, .{null} ** 3, abi.RegisterClass.gp);
@@ -8222,7 +8428,7 @@ fn genShiftBinOpMir(
                         try self.asmRegisterImmediate(
                             .{ ._, .@"and" },
                             .cl,
-                            Immediate.u(std.math.maxInt(u6)),
+                            Immediate.u(math.maxInt(u6)),
                         );
                         try self.asmRegisterImmediate(
                             .{ ._r, .sh },
@@ -8259,7 +8465,7 @@ fn genShiftBinOpMir(
                         try self.asmRegisterImmediate(
                             .{ ._, .@"and" },
                             .cl,
-                            Immediate.u(std.math.maxInt(u6)),
+                            Immediate.u(math.maxInt(u6)),
                         );
                         try self.asmRegisterImmediate(
                             .{ ._r, .sh },
@@ -8324,7 +8530,7 @@ fn genShiftBinOpMir(
                     }, .sh },
                     temp_regs[2].to64(),
                     temp_regs[3].to64(),
-                    Immediate.u(shift_imm & std.math.maxInt(u6)),
+                    Immediate.u(shift_imm & math.maxInt(u6)),
                 ),
                 else => try self.asmRegisterRegisterRegister(.{ switch (tag[0]) {
                     ._l => ._ld,
@@ -8379,7 +8585,7 @@ fn genShiftBinOpMir(
             .immediate => |shift_imm| try self.asmRegisterImmediate(
                 tag,
                 temp_regs[2].to64(),
-                Immediate.u(shift_imm & std.math.maxInt(u6)),
+                Immediate.u(shift_imm & math.maxInt(u6)),
             ),
             else => try self.asmRegisterRegister(tag, temp_regs[2].to64(), .cl),
         }
@@ -8974,7 +9180,7 @@ fn genMulDivBinOp(
         switch (tag) {
             .mul, .mul_wrap => {
                 const slow_inc = self.hasFeature(.slow_incdec);
-                const limb_len = std.math.divCeil(u32, src_abi_size, 8) catch unreachable;
+                const limb_len = math.divCeil(u32, src_abi_size, 8) catch unreachable;
 
                 try self.spillRegisters(&.{ .rax, .rcx, .rdx });
                 const reg_locks = self.register_manager.lockRegs(3, .{ .rax, .rcx, .rdx });
@@ -14535,7 +14741,7 @@ fn genSetReg(
                 ty,
                 dst_reg.class(),
                 self.getFrameAddrAlignment(frame_addr).compare(.gte, Alignment.fromLog2Units(
-                    std.math.log2_int_ceil(u10, @divExact(dst_reg.bitSize(), 8)),
+                    math.log2_int_ceil(u10, @divExact(dst_reg.bitSize(), 8)),
                 )),
             ),
             .lea_frame => .{ .move = .{ ._, .lea } },
@@ -16833,6 +17039,7 @@ fn airShuffle(self: *Self, inst: Air.Inst.Index) !void {
             @intCast(mask_elem_val.toSignedInt(mod));
     }
 
+    const has_avx = self.hasFeature(.avx);
     const result = @as(?MCValue, result: {
         for (mask_elems) |mask_elem| {
             if (mask_elem) |_| break;
@@ -16858,7 +17065,6 @@ fn airShuffle(self: *Self, inst: Air.Inst.Index) !void {
             break :result dst_mcv;
         }
 
-        const has_avx = self.hasFeature(.avx);
         shufpd: {
             if (elem_abi_size != 8) break :shufpd;
             if (max_abi_size > @as(u32, if (has_avx) 32 else 16)) break :shufpd;
src/arch/x86_64/Encoding.zig
@@ -335,6 +335,8 @@ pub const Mnemonic = enum {
     pextrb, pextrd, pextrq,
     pinsrb, pinsrd, pinsrq,
     pmaxsb, pmaxsd, pmaxud, pmaxuw, pminsb, pminsd, pminud, pminuw,
+    pmovsxbd, pmovsxbq, pmovsxbw, pmovsxdq, pmovsxwd, pmovsxwq,
+    pmovzxbd, pmovzxbq, pmovzxbw, pmovzxdq, pmovzxwd, pmovzxwq,
     pmulld,
     roundpd, roundps, roundsd, roundss,
     // SSE4.2
@@ -387,6 +389,8 @@ pub const Mnemonic = enum {
     vpmaxsb, vpmaxsd, vpmaxsw, vpmaxub, vpmaxud, vpmaxuw,
     vpminsb, vpminsd, vpminsw, vpminub, vpminud, vpminuw,
     vpmovmskb,
+    vpmovsxbd, vpmovsxbq, vpmovsxbw, vpmovsxdq, vpmovsxwd, vpmovsxwq,
+    vpmovzxbd, vpmovzxbq, vpmovzxbw, vpmovzxdq, vpmovzxwd, vpmovzxwq,
     vpmulhw, vpmulld, vpmullw,
     vpor,
     vpshufb, vpshufd, vpshufhw, vpshuflw,
src/arch/x86_64/encodings.zig
@@ -1235,6 +1235,20 @@ pub const table = [_]Entry{
 
     .{ .pminud, .rm, &.{ .xmm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x3b }, 0, .none, .sse4_1 },
 
+    .{ .pmovsxbw, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x20 }, 0, .none, .sse4_1 },
+    .{ .pmovsxbd, .rm, &.{ .xmm, .xmm_m32 }, &.{ 0x66, 0x0f, 0x38, 0x21 }, 0, .none, .sse4_1 },
+    .{ .pmovsxbq, .rm, &.{ .xmm, .xmm_m16 }, &.{ 0x66, 0x0f, 0x38, 0x22 }, 0, .none, .sse4_1 },
+    .{ .pmovsxwd, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x23 }, 0, .none, .sse4_1 },
+    .{ .pmovsxwq, .rm, &.{ .xmm, .xmm_m32 }, &.{ 0x66, 0x0f, 0x38, 0x24 }, 0, .none, .sse4_1 },
+    .{ .pmovsxdq, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x25 }, 0, .none, .sse4_1 },
+
+    .{ .pmovzxbw, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x30 }, 0, .none, .sse4_1 },
+    .{ .pmovzxbd, .rm, &.{ .xmm, .xmm_m32 }, &.{ 0x66, 0x0f, 0x38, 0x31 }, 0, .none, .sse4_1 },
+    .{ .pmovzxbq, .rm, &.{ .xmm, .xmm_m16 }, &.{ 0x66, 0x0f, 0x38, 0x32 }, 0, .none, .sse4_1 },
+    .{ .pmovzxwd, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x33 }, 0, .none, .sse4_1 },
+    .{ .pmovzxwq, .rm, &.{ .xmm, .xmm_m32 }, &.{ 0x66, 0x0f, 0x38, 0x34 }, 0, .none, .sse4_1 },
+    .{ .pmovzxdq, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x35 }, 0, .none, .sse4_1 },
+
     .{ .pmulld, .rm, &.{ .xmm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x40 }, 0, .none, .sse4_1 },
 
     .{ .roundpd, .rmi, &.{ .xmm, .xmm_m128, .imm8 }, &.{ 0x66, 0x0f, 0x3a, 0x09 }, 0, .none, .sse4_1 },
@@ -1587,6 +1601,20 @@ pub const table = [_]Entry{
     .{ .vpmovmskb, .rm, &.{ .r32, .xmm }, &.{ 0x66, 0x0f, 0xd7 }, 0, .vex_128_wig, .avx },
     .{ .vpmovmskb, .rm, &.{ .r64, .xmm }, &.{ 0x66, 0x0f, 0xd7 }, 0, .vex_128_wig, .avx },
 
+    .{ .vpmovsxbw, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x20 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovsxbd, .rm, &.{ .xmm, .xmm_m32 }, &.{ 0x66, 0x0f, 0x38, 0x21 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovsxbq, .rm, &.{ .xmm, .xmm_m16 }, &.{ 0x66, 0x0f, 0x38, 0x22 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovsxwd, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x23 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovsxwq, .rm, &.{ .xmm, .xmm_m32 }, &.{ 0x66, 0x0f, 0x38, 0x24 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovsxdq, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x25 }, 0, .vex_128_wig, .avx },
+
+    .{ .vpmovzxbw, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x30 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovzxbd, .rm, &.{ .xmm, .xmm_m32 }, &.{ 0x66, 0x0f, 0x38, 0x31 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovzxbq, .rm, &.{ .xmm, .xmm_m16 }, &.{ 0x66, 0x0f, 0x38, 0x32 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovzxwd, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x33 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovzxwq, .rm, &.{ .xmm, .xmm_m32 }, &.{ 0x66, 0x0f, 0x38, 0x34 }, 0, .vex_128_wig, .avx },
+    .{ .vpmovzxdq, .rm, &.{ .xmm, .xmm_m64 }, &.{ 0x66, 0x0f, 0x38, 0x35 }, 0, .vex_128_wig, .avx },
+
     .{ .vpmulhw, .rvm, &.{ .xmm, .xmm, .xmm_m128 }, &.{ 0x66, 0x0f, 0xe5 }, 0, .vex_128_wig, .avx },
 
     .{ .vpmulld, .rvm, &.{ .xmm, .xmm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x40 }, 0, .vex_128_wig, .avx },
@@ -1816,6 +1844,20 @@ pub const table = [_]Entry{
     .{ .vpmovmskb, .rm, &.{ .r32, .ymm }, &.{ 0x66, 0x0f, 0xd7 }, 0, .vex_256_wig, .avx2 },
     .{ .vpmovmskb, .rm, &.{ .r64, .ymm }, &.{ 0x66, 0x0f, 0xd7 }, 0, .vex_256_wig, .avx2 },
 
+    .{ .vpmovsxbw, .rm, &.{ .ymm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x20 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovsxbd, .rm, &.{ .ymm, .xmm_m64  }, &.{ 0x66, 0x0f, 0x38, 0x21 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovsxbq, .rm, &.{ .ymm, .xmm_m32  }, &.{ 0x66, 0x0f, 0x38, 0x22 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovsxwd, .rm, &.{ .ymm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x23 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovsxwq, .rm, &.{ .ymm, .xmm_m64  }, &.{ 0x66, 0x0f, 0x38, 0x24 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovsxdq, .rm, &.{ .ymm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x25 }, 0, .vex_256_wig, .avx2 },
+
+    .{ .vpmovzxbw, .rm, &.{ .ymm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x30 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovzxbd, .rm, &.{ .ymm, .xmm_m64  }, &.{ 0x66, 0x0f, 0x38, 0x31 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovzxbq, .rm, &.{ .ymm, .xmm_m32  }, &.{ 0x66, 0x0f, 0x38, 0x32 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovzxwd, .rm, &.{ .ymm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x33 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovzxwq, .rm, &.{ .ymm, .xmm_m64  }, &.{ 0x66, 0x0f, 0x38, 0x34 }, 0, .vex_256_wig, .avx2 },
+    .{ .vpmovzxdq, .rm, &.{ .ymm, .xmm_m128 }, &.{ 0x66, 0x0f, 0x38, 0x35 }, 0, .vex_256_wig, .avx2 },
+
     .{ .vpmulhw, .rvm, &.{ .ymm, .ymm, .ymm_m256 }, &.{ 0x66, 0x0f, 0xe5 }, 0, .vex_256_wig, .avx2 },
 
     .{ .vpmulld, .rvm, &.{ .ymm, .ymm, .ymm_m256 }, &.{ 0x66, 0x0f, 0x38, 0x40 }, 0, .vex_256_wig, .avx2 },
src/arch/x86_64/Mir.zig
@@ -658,6 +658,14 @@ pub const Inst = struct {
         /// Insert scalar single-precision floating-point value
         /// Insert packed floating-point values
         insert,
+        /// Packed move with sign extend
+        movsxb,
+        movsxd,
+        movsxw,
+        /// Packed move with zero extend
+        movzxb,
+        movzxd,
+        movzxw,
         /// Round packed single-precision floating-point values
         /// Round scalar single-precision floating-point value
         /// Round packed double-precision floating-point values
src/codegen/c.zig
@@ -6109,41 +6109,48 @@ fn airFloatCast(f: *Function, inst: Air.Inst.Index) !CValue {
     const ty_op = f.air.instructions.items(.data)[@intFromEnum(inst)].ty_op;
 
     const inst_ty = f.typeOfIndex(inst);
+    const inst_scalar_ty = inst_ty.scalarType(mod);
     const operand = try f.resolveInst(ty_op.operand);
     try reap(f, inst, &.{ty_op.operand});
     const operand_ty = f.typeOf(ty_op.operand);
+    const scalar_ty = operand_ty.scalarType(mod);
     const target = f.object.dg.module.getTarget();
-    const operation = if (inst_ty.isRuntimeFloat() and operand_ty.isRuntimeFloat())
-        if (inst_ty.floatBits(target) < operand_ty.floatBits(target)) "trunc" else "extend"
-    else if (inst_ty.isInt(mod) and operand_ty.isRuntimeFloat())
-        if (inst_ty.isSignedInt(mod)) "fix" else "fixuns"
-    else if (inst_ty.isRuntimeFloat() and operand_ty.isInt(mod))
-        if (operand_ty.isSignedInt(mod)) "float" else "floatun"
+    const operation = if (inst_scalar_ty.isRuntimeFloat() and scalar_ty.isRuntimeFloat())
+        if (inst_scalar_ty.floatBits(target) < scalar_ty.floatBits(target)) "trunc" else "extend"
+    else if (inst_scalar_ty.isInt(mod) and scalar_ty.isRuntimeFloat())
+        if (inst_scalar_ty.isSignedInt(mod)) "fix" else "fixuns"
+    else if (inst_scalar_ty.isRuntimeFloat() and scalar_ty.isInt(mod))
+        if (scalar_ty.isSignedInt(mod)) "float" else "floatun"
     else
         unreachable;
 
     const writer = f.object.writer();
     const local = try f.allocLocal(inst, inst_ty);
+    const v = try Vectorize.start(f, inst, writer, operand_ty);
+    const a = try Assignment.start(f, writer, scalar_ty);
     try f.writeCValue(writer, local, .Other);
-
-    try writer.writeAll(" = ");
-    if (inst_ty.isInt(mod) and operand_ty.isRuntimeFloat()) {
+    try v.elem(f, writer);
+    try a.assign(f, writer);
+    if (inst_scalar_ty.isInt(mod) and scalar_ty.isRuntimeFloat()) {
         try writer.writeAll("zig_wrap_");
-        try f.object.dg.renderTypeForBuiltinFnName(writer, inst_ty);
+        try f.object.dg.renderTypeForBuiltinFnName(writer, inst_scalar_ty);
         try writer.writeByte('(');
     }
     try writer.writeAll("zig_");
     try writer.writeAll(operation);
-    try writer.writeAll(compilerRtAbbrev(operand_ty, mod));
-    try writer.writeAll(compilerRtAbbrev(inst_ty, mod));
+    try writer.writeAll(compilerRtAbbrev(scalar_ty, mod));
+    try writer.writeAll(compilerRtAbbrev(inst_scalar_ty, mod));
     try writer.writeByte('(');
     try f.writeCValue(writer, operand, .FunctionArgument);
+    try v.elem(f, writer);
     try writer.writeByte(')');
-    if (inst_ty.isInt(mod) and operand_ty.isRuntimeFloat()) {
-        try f.object.dg.renderBuiltinInfo(writer, inst_ty, .bits);
+    if (inst_scalar_ty.isInt(mod) and scalar_ty.isRuntimeFloat()) {
+        try f.object.dg.renderBuiltinInfo(writer, inst_scalar_ty, .bits);
         try writer.writeByte(')');
     }
-    try writer.writeAll(";\n");
+    try a.end(f, writer);
+    try v.end(f, inst, writer);
+
     return local;
 }
 
src/codegen/llvm.zig
@@ -8648,8 +8648,6 @@ pub const FuncGen = struct {
         const operand_ty = self.typeOf(ty_op.operand);
         const dest_ty = self.typeOfIndex(inst);
         const target = mod.getTarget();
-        const dest_bits = dest_ty.floatBits(target);
-        const src_bits = operand_ty.floatBits(target);
 
         if (intrinsicsAllowed(dest_ty, target) and intrinsicsAllowed(operand_ty, target)) {
             return self.wip.cast(.fpext, operand, try o.lowerType(dest_ty), "");
@@ -8657,11 +8655,19 @@ pub const FuncGen = struct {
             const operand_llvm_ty = try o.lowerType(operand_ty);
             const dest_llvm_ty = try o.lowerType(dest_ty);
 
+            const dest_bits = dest_ty.scalarType(mod).floatBits(target);
+            const src_bits = operand_ty.scalarType(mod).floatBits(target);
             const fn_name = try o.builder.fmt("__extend{s}f{s}f2", .{
                 compilerRtFloatAbbrev(src_bits), compilerRtFloatAbbrev(dest_bits),
             });
 
             const libc_fn = try self.getLibcFunction(fn_name, &.{operand_llvm_ty}, dest_llvm_ty);
+            if (dest_ty.isVector(mod)) return self.buildElementwiseCall(
+                libc_fn,
+                &.{operand},
+                try o.builder.poisonValue(dest_llvm_ty),
+                dest_ty.vectorLen(mod),
+            );
             return self.wip.call(
                 .normal,
                 .ccc,
src/InternPool.zig
@@ -7404,10 +7404,14 @@ pub fn isIntegerType(ip: *const InternPool, ty: Index) bool {
         .c_ulong_type,
         .c_longlong_type,
         .c_ulonglong_type,
-        .c_longdouble_type,
         .comptime_int_type,
         => true,
-        else => ip.indexToKey(ty) == .int_type,
+        else => switch (ip.items.items(.tag)[@intFromEnum(ty)]) {
+            .type_int_signed,
+            .type_int_unsigned,
+            => true,
+            else => false,
+        },
     };
 }
 
src/Sema.zig
@@ -23328,7 +23328,8 @@ fn checkVectorElemType(
     const mod = sema.mod;
     switch (ty.zigTypeTag(mod)) {
         .Int, .Float, .Bool => return,
-        else => if (ty.isPtrAtRuntime(mod)) return,
+        .Optional, .Pointer => if (ty.isPtrAtRuntime(mod)) return,
+        else => {},
     }
     return sema.fail(block, ty_src, "expected integer, float, bool, or pointer for the vector element type; found '{}'", .{ty.fmt(mod)});
 }
@@ -28455,7 +28456,7 @@ const CoerceOpts = struct {
     report_err: bool = true,
     /// Ignored if `report_err == false`.
     is_ret: bool = false,
-    /// Should coercion to comptime_int ermit an error message.
+    /// Should coercion to comptime_int emit an error message.
     no_cast_to_comptime_int: bool = false,
 
     param_src: struct {
@@ -31858,6 +31859,34 @@ fn coerceArrayLike(
     }
 
     const dest_elem_ty = dest_ty.childType(mod);
+    if (dest_ty.isVector(mod) and inst_ty.isVector(mod) and (try sema.resolveValue(inst)) == null) {
+        const inst_elem_ty = inst_ty.childType(mod);
+        switch (dest_elem_ty.zigTypeTag(mod)) {
+            .Int => if (inst_elem_ty.isInt(mod)) {
+                // integer widening
+                const dst_info = dest_elem_ty.intInfo(mod);
+                const src_info = inst_elem_ty.intInfo(mod);
+                if ((src_info.signedness == dst_info.signedness and dst_info.bits >= src_info.bits) or
+                    // small enough unsigned ints can get casted to large enough signed ints
+                    (dst_info.signedness == .signed and dst_info.bits > src_info.bits))
+                {
+                    try sema.requireRuntimeBlock(block, inst_src, null);
+                    return block.addTyOp(.intcast, dest_ty, inst);
+                }
+            },
+            .Float => if (inst_elem_ty.isRuntimeFloat()) {
+                // float widening
+                const src_bits = inst_elem_ty.floatBits(target);
+                const dst_bits = dest_elem_ty.floatBits(target);
+                if (dst_bits >= src_bits) {
+                    try sema.requireRuntimeBlock(block, inst_src, null);
+                    return block.addTyOp(.fpext, dest_ty, inst);
+                }
+            },
+            else => {},
+        }
+    }
+
     const element_vals = try sema.arena.alloc(InternPool.Index, dest_len);
     const element_refs = try sema.arena.alloc(Air.Inst.Ref, dest_len);
     var runtime_src: ?LazySrcLoc = null;
src/type.zig
@@ -2134,7 +2134,8 @@ pub const Type = struct {
 
     /// Returns true if and only if the type is a fixed-width integer.
     pub fn isInt(self: Type, mod: *const Module) bool {
-        return self.isSignedInt(mod) or self.isUnsignedInt(mod);
+        return self.toIntern() != .comptime_int_type and
+            mod.intern_pool.isIntegerType(self.toIntern());
     }
 
     /// Returns true if and only if the type is a fixed-width, signed integer.
test/behavior/cast.zig
@@ -601,25 +601,25 @@ test "cast *[1][*]const u8 to [*]const ?[*]const u8" {
 
 test "@intCast on vector" {
     if (builtin.zig_backend == .stage2_wasm) return error.SkipZigTest; // TODO
-    if (builtin.zig_backend == .stage2_x86_64) return error.SkipZigTest; // TODO
     if (builtin.zig_backend == .stage2_aarch64) return error.SkipZigTest; // TODO
     if (builtin.zig_backend == .stage2_arm) return error.SkipZigTest; // TODO
     if (builtin.zig_backend == .stage2_sparc64) return error.SkipZigTest; // TODO
+    if (builtin.zig_backend == .stage2_x86_64) return error.SkipZigTest;
 
     const S = struct {
         fn doTheTest() !void {
             // Upcast (implicit, equivalent to @intCast)
             var up0: @Vector(2, u8) = [_]u8{ 0x55, 0xaa };
             _ = &up0;
-            const up1 = @as(@Vector(2, u16), up0);
-            const up2 = @as(@Vector(2, u32), up0);
-            const up3 = @as(@Vector(2, u64), up0);
+            const up1: @Vector(2, u16) = up0;
+            const up2: @Vector(2, u32) = up0;
+            const up3: @Vector(2, u64) = up0;
             // Downcast (safety-checked)
             var down0 = up3;
             _ = &down0;
-            const down1 = @as(@Vector(2, u32), @intCast(down0));
-            const down2 = @as(@Vector(2, u16), @intCast(down0));
-            const down3 = @as(@Vector(2, u8), @intCast(down0));
+            const down1: @Vector(2, u32) = @intCast(down0);
+            const down2: @Vector(2, u16) = @intCast(down0);
+            const down3: @Vector(2, u8) = @intCast(down0);
 
             try expect(mem.eql(u16, &@as([2]u16, up1), &[2]u16{ 0x55, 0xaa }));
             try expect(mem.eql(u32, &@as([2]u32, up2), &[2]u32{ 0x55, 0xaa }));
@@ -629,20 +629,10 @@ test "@intCast on vector" {
             try expect(mem.eql(u16, &@as([2]u16, down2), &[2]u16{ 0x55, 0xaa }));
             try expect(mem.eql(u8, &@as([2]u8, down3), &[2]u8{ 0x55, 0xaa }));
         }
-
-        fn doTheTestFloat() !void {
-            var vec: @Vector(2, f32) = @splat(1234.0);
-            _ = &vec;
-            const wider: @Vector(2, f64) = vec;
-            try expect(wider[0] == 1234.0);
-            try expect(wider[1] == 1234.0);
-        }
     };
 
     try S.doTheTest();
     try comptime S.doTheTest();
-    try S.doTheTestFloat();
-    try comptime S.doTheTestFloat();
 }
 
 test "@floatCast cast down" {
@@ -2340,10 +2330,31 @@ test "@floatCast on vector" {
 
     const S = struct {
         fn doTheTest() !void {
-            var a: @Vector(3, f64) = .{ 1.5, 2.5, 3.5 };
-            _ = &a;
-            const b: @Vector(3, f32) = @floatCast(a);
-            try expectEqual(@Vector(3, f32){ 1.5, 2.5, 3.5 }, b);
+            {
+                var a: @Vector(2, f64) = .{ 1.5, 2.5 };
+                _ = &a;
+                const b: @Vector(2, f32) = @floatCast(a);
+                try expectEqual(@Vector(2, f32){ 1.5, 2.5 }, b);
+            }
+            {
+                var a: @Vector(2, f32) = .{ 3.25, 4.25 };
+                _ = &a;
+                const b: @Vector(2, f64) = @floatCast(a);
+                try expectEqual(@Vector(2, f64){ 3.25, 4.25 }, b);
+            }
+            {
+                var a: @Vector(2, f32) = .{ 5.75, 6.75 };
+                _ = &a;
+                const b: @Vector(2, f64) = a;
+                try expectEqual(@Vector(2, f64){ 5.75, 6.75 }, b);
+            }
+            {
+                var vec: @Vector(2, f32) = @splat(1234.0);
+                _ = &vec;
+                const wider: @Vector(2, f64) = vec;
+                try expect(wider[0] == 1234.0);
+                try expect(wider[1] == 1234.0);
+            }
         }
     };