port packed vector elem ptr logic from stage1

Closes #12812
Closes #13925

Author: Vexu

src/Air.zig

@@ -737,6 +737,10 @@ pub const Inst = struct {
         /// Uses the `ty_pl` field.
         save_err_return_trace_index,
 
+        /// Store an element to a vector pointer at an index.
+        /// Uses the `vector_store_elem` field.
+        vector_store_elem,
+
         pub fn fromCmpOp(op: std.math.CompareOperator, optimized: bool) Tag {
             switch (op) {
                 .lt => return if (optimized) .cmp_lt_optimized else .cmp_lt,
@@ -814,6 +818,11 @@ pub const Inst = struct {
             operand: Ref,
             operation: std.builtin.ReduceOp,
         },
+        vector_store_elem: struct {
+            vector_ptr: Ref,
+            // Index into a different array.
+            payload: u32,
+        },
 
         // Make sure we don't accidentally add a field to make this union
         // bigger than expected. Note that in Debug builds, Zig is allowed
@@ -1177,6 +1186,7 @@ pub fn typeOfIndex(air: Air, inst: Air.Inst.Index) Type {
         .set_union_tag,
         .prefetch,
         .set_err_return_trace,
+        .vector_store_elem,
         => return Type.void,
 
         .ptrtoint,

src/Liveness.zig

@@ -212,6 +212,15 @@ pub fn categorizeOperand(
             return .write;
         },
 
+        .vector_store_elem => {
+            const o = air_datas[inst].vector_store_elem;
+            const extra = air.extraData(Air.Bin, o.payload).data;
+            if (o.vector_ptr == operand_ref) return matchOperandSmallIndex(l, inst, 0, .write);
+            if (extra.lhs == operand_ref) return matchOperandSmallIndex(l, inst, 1, .none);
+            if (extra.rhs == operand_ref) return matchOperandSmallIndex(l, inst, 2, .none);
+            return .write;
+        },
+
         .arg,
         .alloc,
         .ret_ptr,
@@ -824,6 +833,12 @@ fn analyzeInst(
             return trackOperands(a, new_set, inst, main_tomb, .{ o.lhs, o.rhs, .none });
         },
 
+        .vector_store_elem => {
+            const o = inst_datas[inst].vector_store_elem;
+            const extra = a.air.extraData(Air.Bin, o.payload).data;
+            return trackOperands(a, new_set, inst, main_tomb, .{ o.vector_ptr, extra.lhs, extra.rhs });
+        },
+
         .arg,
         .alloc,
         .ret_ptr,

src/Sema.zig

@@ -25952,6 +25952,30 @@ fn storePtr2(
 
     try sema.requireRuntimeBlock(block, src, runtime_src);
     try sema.queueFullTypeResolution(elem_ty);
+
+    if (ptr_ty.ptrInfo().data.vector_index == .runtime) {
+        const ptr_inst = Air.refToIndex(ptr).?;
+        const air_tags = sema.air_instructions.items(.tag);
+        if (air_tags[ptr_inst] == .ptr_elem_ptr) {
+            const ty_pl = sema.air_instructions.items(.data)[ptr_inst].ty_pl;
+            const bin_op = sema.getTmpAir().extraData(Air.Bin, ty_pl.payload).data;
+            _ = try block.addInst(.{
+                .tag = .vector_store_elem,
+                .data = .{ .vector_store_elem = .{
+                    .vector_ptr = bin_op.lhs,
+                    .payload = try block.sema.addExtra(Air.Bin{
+                        .lhs = bin_op.rhs,
+                        .rhs = operand,
+                    }),
+                } },
+            });
+            return;
+        }
+        return sema.fail(block, ptr_src, "unable to determine vector element index of type '{}'", .{
+            ptr_ty.fmt(sema.mod),
+        });
+    }
+
     if (is_ret) {
         _ = try block.addBinOp(.store, ptr, operand);
     } else {
@@ -27827,6 +27851,19 @@ fn analyzeLoad(
         }
     }
 
+    if (ptr_ty.ptrInfo().data.vector_index == .runtime) {
+        const ptr_inst = Air.refToIndex(ptr).?;
+        const air_tags = sema.air_instructions.items(.tag);
+        if (air_tags[ptr_inst] == .ptr_elem_ptr) {
+            const ty_pl = sema.air_instructions.items(.data)[ptr_inst].ty_pl;
+            const bin_op = sema.getTmpAir().extraData(Air.Bin, ty_pl.payload).data;
+            return block.addBinOp(.ptr_elem_val, bin_op.lhs, bin_op.rhs);
+        }
+        return sema.fail(block, ptr_src, "unable to determine vector element index of type '{}'", .{
+            ptr_ty.fmt(sema.mod),
+        });
+    }
+
     return block.addTyOp(.load, elem_ty, ptr);
 }
 
@@ -32697,23 +32734,24 @@ fn elemPtrType(sema: *Sema, ptr_ty: Type, offset: ?usize) !Type {
     const target = sema.mod.getTarget();
     const parent_ty = ptr_ty.childType();
 
+    const VI = Type.Payload.Pointer.Data.VectorIndex;
+
     const vector_info: struct {
-        host_size: u16,
-        bit_offset: u16,
-        alignment: u32,
+        host_size: u16 = 0,
+        alignment: u32 = 0,
+        vector_index: VI = .none,
     } = if (parent_ty.tag() == .vector) blk: {
         const elem_bits = elem_ty.bitSize(target);
-        const is_packed = elem_bits != 0 and (elem_bits & (elem_bits - 1)) != 0;
-        // TODO: runtime-known index
-        assert(!is_packed or offset != null);
-        const is_packed_with_offset = is_packed and offset != null and offset.? != 0;
-        const target_offset = if (is_packed_with_offset) (if (target.cpu.arch.endian() == .Big) (parent_ty.vectorLen() - 1 - offset.?) else offset.?) else 0;
+        if (elem_bits == 0) break :blk .{};
+        const is_packed = elem_bits < 8 or !std.math.isPowerOfTwo(elem_bits);
+        if (!is_packed) break :blk .{};
+
         break :blk .{
-            .host_size = if (is_packed_with_offset) @intCast(u16, parent_ty.abiSize(target)) else 0,
-            .bit_offset = if (is_packed_with_offset) @intCast(u16, elem_bits * target_offset) else 0,
-            .alignment = if (is_packed_with_offset) @intCast(u16, parent_ty.abiAlignment(target)) else 0,
+            .host_size = @intCast(u16, parent_ty.arrayLen()),
+            .alignment = @intCast(u16, parent_ty.abiAlignment(target)),
+            .vector_index = if (offset) |some| @intToEnum(VI, some) else .runtime,
         };
-    } else .{ .host_size = 0, .bit_offset = 0, .alignment = 0 };
+    } else .{};
 
     const alignment: u32 = a: {
         // Calculate the new pointer alignment.
@@ -32741,6 +32779,6 @@ fn elemPtrType(sema: *Sema, ptr_ty: Type, offset: ?usize) !Type {
         .@"volatile" = ptr_info.@"volatile",
         .@"align" = alignment,
         .host_size = vector_info.host_size,
-        .bit_offset = vector_info.bit_offset,
+        .vector_index = vector_info.vector_index,
     });
 }

src/arch/aarch64/CodeGen.zig

@@ -873,6 +873,7 @@ fn genBody(self: *Self, body: []const Air.Inst.Index) InnerError!void {
 
             .is_named_enum_value => return self.fail("TODO implement is_named_enum_value", .{}),
             .error_set_has_value => return self.fail("TODO implement error_set_has_value", .{}),
+            .vector_store_elem => return self.fail("TODO implement vector_store_elem", .{}),
 
             .wasm_memory_size => unreachable,
             .wasm_memory_grow => unreachable,

src/arch/arm/CodeGen.zig

@@ -783,6 +783,7 @@ fn genBody(self: *Self, body: []const Air.Inst.Index) InnerError!void {
 
             .is_named_enum_value => return self.fail("TODO implement is_named_enum_value", .{}),
             .error_set_has_value => return self.fail("TODO implement error_set_has_value", .{}),
+            .vector_store_elem => return self.fail("TODO implement vector_store_elem", .{}),
 
             .wasm_memory_size => unreachable,
             .wasm_memory_grow => unreachable,

src/arch/riscv64/CodeGen.zig

@@ -697,6 +697,7 @@ fn genBody(self: *Self, body: []const Air.Inst.Index) InnerError!void {
 
             .is_named_enum_value => return self.fail("TODO implement is_named_enum_value", .{}),
             .error_set_has_value => return self.fail("TODO implement error_set_has_value", .{}),
+            .vector_store_elem => return self.fail("TODO implement vector_store_elem", .{}),
 
             .wasm_memory_size => unreachable,
             .wasm_memory_grow => unreachable,

src/arch/sparc64/CodeGen.zig

@@ -714,6 +714,7 @@ fn genBody(self: *Self, body: []const Air.Inst.Index) InnerError!void {
 
             .is_named_enum_value => @panic("TODO implement is_named_enum_value"),
             .error_set_has_value => @panic("TODO implement error_set_has_value"),
+            .vector_store_elem => @panic("TODO implement vector_store_elem"),
 
             .wasm_memory_size => unreachable,
             .wasm_memory_grow => unreachable,

src/arch/wasm/CodeGen.zig

@@ -1971,6 +1971,7 @@ fn genInst(func: *CodeGen, inst: Air.Inst.Index) InnerError!void {
         .is_named_enum_value,
         .error_set_has_value,
         .addrspace_cast,
+        .vector_store_elem,
         => |tag| return func.fail("TODO: Implement wasm inst: {s}", .{@tagName(tag)}),
 
         .add_optimized,
@@ -2213,6 +2214,7 @@ fn airStore(func: *CodeGen, inst: Air.Inst.Index) InnerError!void {
     const ptr_ty = func.air.typeOf(bin_op.lhs);
     const ptr_info = ptr_ty.ptrInfo().data;
     const ty = ptr_ty.childType();
+
     if (ptr_info.host_size == 0) {
         try func.store(lhs, rhs, ty, 0);
     } else {

src/arch/x86_64/CodeGen.zig

@@ -785,6 +785,7 @@ fn genBody(self: *Self, body: []const Air.Inst.Index) InnerError!void {
 
             .is_named_enum_value => return self.fail("TODO implement is_named_enum_value", .{}),
             .error_set_has_value => return self.fail("TODO implement error_set_has_value", .{}),
+            .vector_store_elem => return self.fail("TODO implement vector_store_elem", .{}),
 
             .wasm_memory_size => unreachable,
             .wasm_memory_grow => unreachable,

src/codegen/c.zig

@@ -2908,6 +2908,7 @@ fn genBodyInner(f: *Function, body: []const Air.Inst.Index) error{ AnalysisFail,
 
             .is_named_enum_value => return f.fail("TODO: C backend: implement is_named_enum_value", .{}),
             .error_set_has_value => return f.fail("TODO: C backend: implement error_set_has_value", .{}),
+            .vector_store_elem => return f.fail("TODO: C backend: implement vector_store_elem", .{}),
             // zig fmt: on
         };
         if (result_value == .local) {