Commit e313b387a0

Jacob Young <jacobly0@users.noreply.github.com>
2025-09-11 04:25:55
x86_64: delete usages of avx2 `vpack?s??`
This instruction actually has fairly useless semantics, and even the cases that were semantically correct could save 1 cycle of latency by using a different sequnce involving the avx version instead. Closes #25174
1 parent 1a0a9d7
Changed files (2)
src
arch
test
behavior
src/arch/x86_64/CodeGen.zig
@@ -22283,8 +22283,8 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                         .{ ._, .vp_w, .movsxb, .dst0y, .src0x, ._, ._ },
                         .{ ._, .vp_w, .movsxb, .tmp0y, .src1x, ._, ._ },
                         .{ ._, .vp_w, .mull, .dst0y, .dst0y, .tmp0y, ._ },
-                        .{ ._, .vp_b, .ackssw, .dst0y, .dst0y, .dst0y, ._ },
-                        .{ ._, .v_q, .perm, .dst0y, .dst0y, .ui(0b10_00_10_00), ._ },
+                        .{ ._, .v_i128, .extract, .tmp0x, .dst0y, .ui(1), ._ },
+                        .{ ._, .vp_b, .ackssw, .dst0x, .dst0x, .tmp0x, ._ },
                     } },
                 }, .{
                     .required_features = .{ .avx, null, null, null },
@@ -22414,8 +22414,8 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                         .{ ._, .vp_w, .movzxb, .dst0y, .src0x, ._, ._ },
                         .{ ._, .vp_w, .movzxb, .tmp0y, .src1x, ._, ._ },
                         .{ ._, .vp_w, .mull, .dst0y, .dst0y, .tmp0y, ._ },
-                        .{ ._, .vp_b, .ackusw, .dst0y, .dst0y, .dst0y, ._ },
-                        .{ ._, .v_q, .perm, .dst0y, .dst0y, .ui(0b10_00_10_00), ._ },
+                        .{ ._, .v_i128, .extract, .tmp0x, .dst0y, .ui(1), ._ },
+                        .{ ._, .vp_b, .ackusw, .dst0x, .dst0x, .tmp0x, ._ },
                     } },
                 }, .{
                     .required_features = .{ .avx2, null, null, null },
@@ -22447,8 +22447,8 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                         .{ .@"0:", .vp_w, .movsxb, .tmp1y, .memia(.src0x, .tmp0, .add_unaligned_size), ._, ._ },
                         .{ ._, .vp_w, .movsxb, .tmp2y, .memia(.src1x, .tmp0, .add_unaligned_size), ._, ._ },
                         .{ ._, .vp_w, .mull, .tmp1y, .tmp1y, .tmp2y, ._ },
-                        .{ ._, .vp_b, .ackssw, .tmp1y, .tmp1y, .tmp1y, ._ },
-                        .{ ._, .v_q, .perm, .tmp1y, .tmp1y, .ui(0b10_00_10_00), ._ },
+                        .{ ._, .v_i128, .extract, .tmp2x, .tmp1y, .ui(1), ._ },
+                        .{ ._, .vp_b, .ackssw, .tmp1x, .tmp1x, .tmp2x, ._ },
                         .{ ._, .v_dqa, .mov, .memia(.dst0x, .tmp0, .add_unaligned_size), .tmp1x, ._, ._ },
                         .{ ._, ._, .add, .tmp0p, .si(16), ._, ._ },
                         .{ ._, ._nc, .j, .@"0b", ._, ._, ._ },
@@ -22659,8 +22659,8 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                         .{ .@"0:", .vp_w, .movzxb, .tmp1y, .memia(.src0x, .tmp0, .add_unaligned_size), ._, ._ },
                         .{ ._, .vp_w, .movzxb, .tmp2y, .memia(.src1x, .tmp0, .add_unaligned_size), ._, ._ },
                         .{ ._, .vp_w, .mull, .tmp1y, .tmp1y, .tmp2y, ._ },
-                        .{ ._, .vp_b, .ackusw, .tmp1y, .tmp1y, .tmp1y, ._ },
-                        .{ ._, .v_q, .perm, .tmp1y, .tmp1y, .ui(0b10_00_10_00), ._ },
+                        .{ ._, .v_i128, .extract, .tmp2x, .tmp1y, .ui(1), ._ },
+                        .{ ._, .vp_b, .ackusw, .tmp1x, .tmp1x, .tmp2x, ._ },
                         .{ ._, .v_dqa, .mov, .memia(.dst0x, .tmp0, .add_unaligned_size), .tmp1x, ._, ._ },
                         .{ ._, ._, .add, .tmp0p, .si(16), ._, ._ },
                         .{ ._, ._nc, .j, .@"0b", ._, ._, ._ },
@@ -82559,7 +82559,7 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                                 .{ .type = .u32, .kind = .{ .rc = .general_purpose } },
                                 .{ .type = .u16, .kind = .{ .rc = .general_purpose } },
                                 .{ .kind = .{ .rc = .sse } },
-                                .unused,
+                                .{ .kind = .{ .rc = .sse } },
                                 .unused,
                                 .unused,
                                 .unused,
@@ -82576,8 +82576,9 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                                     .{ ._, ._, .xor, .tmp1d, .tmp1d, ._, ._ },
                                     .{ .@"0:", .v_dqu, .mov, .tmp3y, .memia(.src0y, .tmp0, .add_size), ._, ._ },
                                     .{ ._, .vp_w, .cmpeq, .tmp3y, .tmp3y, .memia(.src1y, .tmp0, .add_size), ._ },
-                                    .{ ._, .vp_b, .ackssw, .tmp3y, .tmp3y, .tmp3y, ._ },
-                                    .{ ._, .vp_b, .movmsk, .tmp2d, .tmp3y, ._, ._ },
+                                    .{ ._, .v_i128, .extract, .tmp4x, .tmp3y, .ui(1), ._ },
+                                    .{ ._, .vp_b, .ackssw, .tmp3x, .tmp3x, .tmp4x, ._ },
+                                    .{ ._, .vp_b, .movmsk, .tmp2d, .tmp3x, ._, ._ },
                                     .{ ._, ._, .mov, .memi(.dst0w, .tmp1), .tmp2w, ._, ._ },
                                     .{ ._, ._, .lea, .tmp1d, .lead(.tmp1, 2), ._, ._ },
                                     .{ ._, ._, .add, .tmp0p, .si(32), ._, ._ },
@@ -82588,8 +82589,9 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                                     .{ ._, ._, .xor, .tmp1d, .tmp1d, ._, ._ },
                                     .{ .@"0:", .v_dqu, .mov, .tmp3y, .memia(.src0y, .tmp0, .add_size), ._, ._ },
                                     .{ ._, .vp_w, .cmpeq, .tmp3y, .tmp3y, .memia(.src1y, .tmp0, .add_size), ._ },
-                                    .{ ._, .vp_b, .ackssw, .tmp3y, .tmp3y, .tmp3y, ._ },
-                                    .{ ._, .vp_b, .movmsk, .tmp2d, .tmp3y, ._, ._ },
+                                    .{ ._, .v_i128, .extract, .tmp4x, .tmp3y, .ui(1), ._ },
+                                    .{ ._, .vp_b, .ackssw, .tmp3x, .tmp3x, .tmp4x, ._ },
+                                    .{ ._, .vp_b, .movmsk, .tmp2d, .tmp3x, ._, ._ },
                                     .{ ._, ._, .not, .tmp2d, ._, ._, ._ },
                                     .{ ._, ._, .mov, .memi(.dst0w, .tmp1), .tmp2w, ._, ._ },
                                     .{ ._, ._, .lea, .tmp1d, .lead(.tmp1, 2), ._, ._ },
@@ -90324,7 +90326,7 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                 }, .{
                     .required_features = .{ .avx, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .xword, .is = .word } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .qword, .is = .byte } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .qword, .is = .byte } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
@@ -90346,7 +90348,7 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                 }, .{
                     .required_features = .{ .sse2, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .xword, .is = .word } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .qword, .is = .byte } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .qword, .is = .byte } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_mut_sse, .none, .none } },
                     },
@@ -90361,20 +90363,46 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
-                    .dst_temps = .{ .{ .mut_rc = .{ .ref = .src0, .rc = .sse } }, .unused },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
+                    .each = .{ .once = &.{
+                        .{ ._, .v_i128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_b, .ackssw, .dst0x, .src0x, .dst0x, ._ },
+                    } },
+                }, .{
+                    .required_features = .{ .avx, null, null, null },
+                    .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .word } }, .any, .any },
+                    .dst_constraints = .{ .{ .scalar_signed_int = .{ .of = .xword, .is = .byte } }, .any },
+                    .patterns = &.{
+                        .{ .src = .{ .to_sse, .none, .none } },
+                    },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
                     .each = .{ .once = &.{
-                        .{ ._, .vp_b, .ackssw, .dst0y, .src0y, .dst0y, ._ },
+                        .{ ._, .v_f128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_b, .ackssw, .dst0x, .src0x, .dst0x, ._ },
                     } },
                 }, .{
                     .required_features = .{ .avx2, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .word } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .xword, .is = .byte } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .xword, .is = .byte } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
-                    .dst_temps = .{ .{ .mut_rc = .{ .ref = .src0, .rc = .sse } }, .unused },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
                     .each = .{ .once = &.{
-                        .{ ._, .vp_b, .ackusw, .dst0y, .src0y, .dst0y, ._ },
+                        .{ ._, .v_i128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_b, .ackusw, .dst0x, .src0x, .dst0x, ._ },
+                    } },
+                }, .{
+                    .required_features = .{ .avx, null, null, null },
+                    .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .word } }, .any, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .xword, .is = .byte } }, .any },
+                    .patterns = &.{
+                        .{ .src = .{ .to_sse, .none, .none } },
+                    },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
+                    .each = .{ .once = &.{
+                        .{ ._, .v_f128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_b, .ackusw, .dst0x, .src0x, .dst0x, ._ },
                     } },
                 }, .{
                     .required_features = .{ .slow_incdec, null, null, null },
@@ -90448,7 +90476,7 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                 }, .{
                     .required_features = .{ .avx, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .xword, .is = .dword } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .dword, .is = .byte } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .dword, .is = .byte } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
@@ -90472,7 +90500,7 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                 }, .{
                     .required_features = .{ .sse4_1, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .xword, .is = .dword } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .dword, .is = .byte } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .dword, .is = .byte } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_mut_sse, .none, .none } },
                     },
@@ -90488,22 +90516,50 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
-                    .dst_temps = .{ .{ .mut_rc = .{ .ref = .src0, .rc = .sse } }, .unused },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
                     .each = .{ .once = &.{
-                        .{ ._, .vp_w, .ackssd, .dst0y, .src0y, .dst0y, ._ },
-                        .{ ._, .vp_b, .ackssw, .dst0y, .dst0y, .dst0y, ._ },
+                        .{ ._, .v_i128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_w, .ackssd, .dst0x, .src0x, .dst0x, ._ },
+                        .{ ._, .vp_b, .ackssw, .dst0x, .dst0x, .dst0x, ._ },
+                    } },
+                }, .{
+                    .required_features = .{ .avx, null, null, null },
+                    .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .dword } }, .any, .any },
+                    .dst_constraints = .{ .{ .scalar_signed_int = .{ .of = .qword, .is = .byte } }, .any },
+                    .patterns = &.{
+                        .{ .src = .{ .to_sse, .none, .none } },
+                    },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
+                    .each = .{ .once = &.{
+                        .{ ._, .v_f128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_w, .ackssd, .dst0x, .src0x, .dst0x, ._ },
+                        .{ ._, .vp_b, .ackssw, .dst0x, .dst0x, .dst0x, ._ },
                     } },
                 }, .{
                     .required_features = .{ .avx2, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .dword } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .qword, .is = .byte } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .qword, .is = .byte } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
-                    .dst_temps = .{ .{ .mut_rc = .{ .ref = .src0, .rc = .sse } }, .unused },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
+                    .each = .{ .once = &.{
+                        .{ ._, .v_i128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_w, .ackusd, .dst0x, .src0x, .dst0x, ._ },
+                        .{ ._, .vp_b, .ackusw, .dst0x, .dst0x, .dst0x, ._ },
+                    } },
+                }, .{
+                    .required_features = .{ .avx, null, null, null },
+                    .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .dword } }, .any, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .qword, .is = .byte } }, .any },
+                    .patterns = &.{
+                        .{ .src = .{ .to_sse, .none, .none } },
+                    },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
                     .each = .{ .once = &.{
-                        .{ ._, .vp_w, .ackusd, .dst0y, .src0y, .dst0y, ._ },
-                        .{ ._, .vp_b, .ackusw, .dst0y, .dst0y, .dst0y, ._ },
+                        .{ ._, .v_f128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_w, .ackusd, .dst0x, .src0x, .dst0x, ._ },
+                        .{ ._, .vp_b, .ackusw, .dst0x, .dst0x, .dst0x, ._ },
                     } },
                 }, .{
                     .required_features = .{ .slow_incdec, null, null, null },
@@ -90722,7 +90778,7 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                 }, .{
                     .required_features = .{ .avx, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .xword, .is = .dword } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .qword, .is = .word } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .qword, .is = .word } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
@@ -90744,7 +90800,7 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                 }, .{
                     .required_features = .{ .sse4_1, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .xword, .is = .dword } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .qword, .is = .word } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .qword, .is = .word } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_mut_sse, .none, .none } },
                     },
@@ -90759,20 +90815,46 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
-                    .dst_temps = .{ .{ .mut_rc = .{ .ref = .src0, .rc = .sse } }, .unused },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
                     .each = .{ .once = &.{
-                        .{ ._, .vp_w, .ackssd, .dst0y, .src0y, .dst0y, ._ },
+                        .{ ._, .v_i128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_w, .ackssd, .dst0x, .src0x, .dst0x, ._ },
+                    } },
+                }, .{
+                    .required_features = .{ .avx, null, null, null },
+                    .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .dword } }, .any, .any },
+                    .dst_constraints = .{ .{ .scalar_signed_int = .{ .of = .xword, .is = .word } }, .any },
+                    .patterns = &.{
+                        .{ .src = .{ .to_sse, .none, .none } },
+                    },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
+                    .each = .{ .once = &.{
+                        .{ ._, .v_f128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_w, .ackssd, .dst0x, .src0x, .dst0x, ._ },
                     } },
                 }, .{
                     .required_features = .{ .avx2, null, null, null },
                     .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .dword } }, .any, .any },
-                    .dst_constraints = .{ .{ .scalar_int = .{ .of = .xword, .is = .word } }, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .xword, .is = .word } }, .any },
                     .patterns = &.{
                         .{ .src = .{ .to_sse, .none, .none } },
                     },
-                    .dst_temps = .{ .{ .mut_rc = .{ .ref = .src0, .rc = .sse } }, .unused },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
                     .each = .{ .once = &.{
-                        .{ ._, .vp_w, .ackusd, .dst0y, .src0y, .dst0y, ._ },
+                        .{ ._, .v_i128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_w, .ackusd, .dst0x, .src0x, .dst0x, ._ },
+                    } },
+                }, .{
+                    .required_features = .{ .avx, null, null, null },
+                    .src_constraints = .{ .{ .scalar_int = .{ .of = .yword, .is = .dword } }, .any, .any },
+                    .dst_constraints = .{ .{ .scalar_unsigned_int = .{ .of = .xword, .is = .word } }, .any },
+                    .patterns = &.{
+                        .{ .src = .{ .to_sse, .none, .none } },
+                    },
+                    .dst_temps = .{ .{ .rc = .sse }, .unused },
+                    .each = .{ .once = &.{
+                        .{ ._, .v_f128, .extract, .dst0x, .src0y, .ui(1), ._ },
+                        .{ ._, .vp_w, .ackusd, .dst0x, .src0x, .dst0x, ._ },
                     } },
                 }, .{
                     .src_constraints = .{ .{ .multiple_scalar_int = .{ .of = .dword, .is = .dword } }, .any, .any },
@@ -92413,7 +92495,7 @@ fn genBody(cg: *CodeGen, body: []const Air.Inst.Index) InnerError!void {
                     .clobbers = .{ .eflags = true },
                     .each = .{ .once = &.{
                         .{ ._, ._, .mov, .tmp0p, .sa(.src0, .sub_unaligned_size), ._, ._ },
-                        .{ .@"0:", .vp_d, .movzxb, .tmp1y, .memia(.src0x, .tmp0, .add_unaligned_size), ._, ._ },
+                        .{ .@"0:", .vp_d, .movzxb, .tmp1y, .memia(.src0q, .tmp0, .add_unaligned_size), ._, ._ },
                         .{ ._, .v_dqa, .mov, .memsia(.dst0y, .@"4", .tmp0, .add_unaligned_size), .tmp1y, ._, ._ },
                         .{ ._, ._, .add, .tmp0p, .si(8), ._, ._ },
                         .{ ._, ._nc, .j, .@"0b", ._, ._, ._ },
test/behavior/cast.zig
@@ -607,26 +607,212 @@ test "@intCast on vector" {
 
     const S = struct {
         fn doTheTest() !void {
-            // Upcast (implicit, equivalent to @intCast)
-            var up0: @Vector(2, u8) = [_]u8{ 0x55, 0xaa };
-            _ = &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: @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 }));
-            try expect(mem.eql(u64, &@as([2]u64, up3), &[2]u64{ 0x55, 0xaa }));
-
-            try expect(mem.eql(u32, &@as([2]u32, down1), &[2]u32{ 0x55, 0xaa }));
-            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 }));
+            {
+                // Upcast (implicit, equivalent to @intCast)
+                var up0: @Vector(2, u8) = .{ 0x55, 0xaa };
+                _ = &up0;
+                const up1: @Vector(2, u16) = up0;
+                const up2: @Vector(2, u32) = up0;
+                const up3: @Vector(2, u64) = up0;
+
+                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 }));
+                try expect(mem.eql(u64, &@as([2]u64, up3), &[2]u64{ 0x55, 0xaa }));
+
+                {
+                    // Downcast (safety-checked)
+                    const down2: @Vector(2, u32) = @intCast(up3);
+                    const down1: @Vector(2, u16) = @intCast(up3);
+                    const down0: @Vector(2, u8) = @intCast(up3);
+
+                    try expect(mem.eql(u32, &@as([2]u32, down2), &[2]u32{ 0x55, 0xaa }));
+                    try expect(mem.eql(u16, &@as([2]u16, down1), &[2]u16{ 0x55, 0xaa }));
+                    try expect(mem.eql(u8, &@as([2]u8, down0), &[2]u8{ 0x55, 0xaa }));
+                }
+
+                {
+                    // Downcast (safety-checked)
+                    const down1: @Vector(2, u16) = @intCast(up2);
+                    const down0: @Vector(2, u8) = @intCast(up2);
+
+                    try expect(mem.eql(u16, &@as([2]u16, down1), &[2]u16{ 0x55, 0xaa }));
+                    try expect(mem.eql(u8, &@as([2]u8, down0), &[2]u8{ 0x55, 0xaa }));
+                }
+
+                {
+                    // Downcast (safety-checked)
+                    const down0: @Vector(2, u8) = @intCast(up1);
+
+                    try expect(mem.eql(u8, &@as([2]u8, down0), &[2]u8{ 0x55, 0xaa }));
+                }
+            }
+            {
+                // Upcast (implicit, equivalent to @intCast)
+                var up0: @Vector(4, u8) = .{ 0x00, 0x55, 0xaa, 0xff };
+                _ = &up0;
+                const up1: @Vector(4, u16) = up0;
+                const up2: @Vector(4, u32) = up0;
+                const up3: @Vector(4, u64) = up0;
+
+                try expect(mem.eql(u16, &@as([4]u16, up1), &[4]u16{ 0x00, 0x55, 0xaa, 0xff }));
+                try expect(mem.eql(u32, &@as([4]u32, up2), &[4]u32{ 0x00, 0x55, 0xaa, 0xff }));
+                try expect(mem.eql(u64, &@as([4]u64, up3), &[4]u64{ 0x00, 0x55, 0xaa, 0xff }));
+
+                {
+                    // Downcast (safety-checked)
+                    const down2: @Vector(4, u32) = @intCast(up3);
+                    const down1: @Vector(4, u16) = @intCast(up3);
+                    const down0: @Vector(4, u8) = @intCast(up3);
+
+                    try expect(mem.eql(u32, &@as([4]u32, down2), &[4]u32{ 0x00, 0x55, 0xaa, 0xff }));
+                    try expect(mem.eql(u16, &@as([4]u16, down1), &[4]u16{ 0x00, 0x55, 0xaa, 0xff }));
+                    try expect(mem.eql(u8, &@as([4]u8, down0), &[4]u8{ 0x00, 0x55, 0xaa, 0xff }));
+                }
+
+                {
+                    // Downcast (safety-checked)
+                    const down1: @Vector(4, u16) = @intCast(up2);
+                    const down0: @Vector(4, u8) = @intCast(up2);
+
+                    try expect(mem.eql(u16, &@as([4]u16, down1), &[4]u16{ 0x00, 0x55, 0xaa, 0xff }));
+                    try expect(mem.eql(u8, &@as([4]u8, down0), &[4]u8{ 0x00, 0x55, 0xaa, 0xff }));
+                }
+
+                {
+                    // Downcast (safety-checked)
+                    const down0: @Vector(4, u8) = @intCast(up1);
+
+                    try expect(mem.eql(u8, &@as([4]u8, down0), &[4]u8{ 0x00, 0x55, 0xaa, 0xff }));
+                }
+            }
+            {
+                // Upcast (implicit, equivalent to @intCast)
+                var up0: @Vector(8, u8) = .{
+                    0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                };
+                _ = &up0;
+                const up1: @Vector(8, u16) = up0;
+                const up2: @Vector(8, u32) = up0;
+                const up3: @Vector(8, u64) = up0;
+
+                try expect(mem.eql(u16, &@as([8]u16, up1), &[8]u16{
+                    0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                }));
+                try expect(mem.eql(u32, &@as([8]u32, up2), &[8]u32{
+                    0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                }));
+                try expect(mem.eql(u64, &@as([8]u64, up3), &[8]u64{
+                    0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                }));
+
+                {
+                    // Downcast (safety-checked)
+                    const down2: @Vector(8, u32) = @intCast(up3);
+                    const down1: @Vector(8, u16) = @intCast(up3);
+                    const down0: @Vector(8, u8) = @intCast(up3);
+
+                    try expect(mem.eql(u32, &@as([8]u32, down2), &[8]u32{
+                        0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                    }));
+                    try expect(mem.eql(u16, &@as([8]u16, down1), &[8]u16{
+                        0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                    }));
+                    try expect(mem.eql(u8, &@as([8]u8, down0), &[8]u8{
+                        0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                    }));
+                }
+
+                {
+                    // Downcast (safety-checked)
+                    const down1: @Vector(8, u16) = @intCast(up2);
+                    const down0: @Vector(8, u8) = @intCast(up2);
+
+                    try expect(mem.eql(u16, &@as([8]u16, down1), &[8]u16{
+                        0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                    }));
+                    try expect(mem.eql(u8, &@as([8]u8, down0), &[8]u8{
+                        0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                    }));
+                }
+
+                {
+                    // Downcast (safety-checked)
+                    const down0: @Vector(8, u8) = @intCast(up1);
+
+                    try expect(mem.eql(u8, &@as([8]u8, down0), &[8]u8{
+                        0x01, 0x23, 0x45, 0x67, 0x89, 0xab, 0xcd, 0xef,
+                    }));
+                }
+            }
+            {
+                // Upcast (implicit, equivalent to @intCast)
+                var up0: @Vector(16, u8) = .{
+                    0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                    0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                };
+                _ = &up0;
+                const up1: @Vector(16, u16) = up0;
+                const up2: @Vector(16, u32) = up0;
+                const up3: @Vector(16, u64) = up0;
+
+                try expect(mem.eql(u16, &@as([16]u16, up1), &[16]u16{
+                    0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                    0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                }));
+                try expect(mem.eql(u32, &@as([16]u32, up2), &[16]u32{
+                    0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                    0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                }));
+                try expect(mem.eql(u64, &@as([16]u64, up3), &[16]u64{
+                    0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                    0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                }));
+
+                {
+                    // Downcast (safety-checked)
+                    const down2: @Vector(16, u32) = @intCast(up3);
+                    const down1: @Vector(16, u16) = @intCast(up3);
+                    const down0: @Vector(16, u8) = @intCast(up3);
+
+                    try expect(mem.eql(u32, &@as([16]u32, down2), &[16]u32{
+                        0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                        0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                    }));
+                    try expect(mem.eql(u16, &@as([16]u16, down1), &[16]u16{
+                        0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                        0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                    }));
+                    try expect(mem.eql(u8, &@as([16]u8, down0), &[16]u8{
+                        0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                        0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                    }));
+                }
+
+                {
+                    // Downcast (safety-checked)
+                    const down1: @Vector(16, u16) = @intCast(up2);
+                    const down0: @Vector(16, u8) = @intCast(up2);
+
+                    try expect(mem.eql(u16, &@as([16]u16, down1), &[16]u16{
+                        0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                        0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                    }));
+                    try expect(mem.eql(u8, &@as([16]u8, down0), &[16]u8{
+                        0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                        0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                    }));
+                }
+
+                {
+                    // Downcast (safety-checked)
+                    const down0: @Vector(16, u8) = @intCast(up1);
+
+                    try expect(mem.eql(u8, &@as([16]u8, down0), &[16]u8{
+                        0x00, 0x11, 0x22, 0x33, 0x44, 0x55, 0x66, 0x77,
+                        0x88, 0x99, 0xaa, 0xbb, 0xcc, 0xdd, 0xee, 0xff,
+                    }));
+                }
+            }
         }
     };