From de6df2bc12c1cec81bb3c562a9395098c92d8239 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sat, 15 May 2021 01:04:02 +0200
Subject: [PATCH 01/21] SPIR-V: Restructure codegen a bit

---
 src/codegen/spirv.zig | 64 ++++++++++++++++++++++++++-----------------
 src/link/SpirV.zig    | 36 +++++++++++++++++++-----
 2 files changed, 68 insertions(+), 32 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index 077e71d4e1..9ceaf107d8 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -6,6 +6,7 @@ const spec = @import("spirv/spec.zig");
 const Module = @import("../Module.zig");
 const Decl = Module.Decl;
 const Type = @import("../type.zig").Type;
+const LazySrcLoc = Module.LazySrcLoc;
 
 pub const TypeMap = std.HashMap(Type, u32, Type.hash, Type.eql, std.hash_map.default_max_load_percentage);
 
@@ -15,30 +16,24 @@ pub fn writeInstruction(code: *std.ArrayList(u32), instr: spec.Opcode, args: []c
     try code.appendSlice(args);
 }
 
+/// This structure represents a SPIR-V binary module being compiled, and keeps track of relevant information
+/// such as code for the different logical sections, and the next result-id.
 pub const SPIRVModule = struct {
-    next_result_id: u32 = 0,
-
-    target: std.Target,
-
-    types: TypeMap,
-
+    next_result_id: u32,
     types_and_globals: std.ArrayList(u32),
     fn_decls: std.ArrayList(u32),
 
-    pub fn init(target: std.Target, allocator: *Allocator) SPIRVModule {
+    pub fn init(allocator: *Allocator) SPIRVModule {
         return .{
-            .target = target,
-            .types = TypeMap.init(allocator),
+            .next_result_id = 0,
             .types_and_globals = std.ArrayList(u32).init(allocator),
             .fn_decls = std.ArrayList(u32).init(allocator),
         };
     }
 
     pub fn deinit(self: *SPIRVModule) void {
-        self.fn_decls.deinit();
         self.types_and_globals.deinit();
-        self.types.deinit();
-        self.* = undefined;
+        self.fn_decls.deinit();
     }
 
     pub fn allocResultId(self: *SPIRVModule) u32 {
@@ -49,21 +44,40 @@ pub const SPIRVModule = struct {
     pub fn resultIdBound(self: *SPIRVModule) u32 {
         return self.next_result_id;
     }
+};
 
-    pub fn getOrGenType(self: *SPIRVModule, t: Type) !u32 {
+/// This structure is used to compile a declaration, and contains all relevant meta-information to deal with that.
+pub const DeclGen = struct {
+    module: *Module,
+    spv: *SPIRVModule,
+
+    types: TypeMap,
+
+    decl: *Decl,
+    error_msg: ?*Module.ErrorMsg,
+
+    fn fail(self: *DeclGen, src: LazySrcLoc, comptime format: []const u8, args: anytype) error{ AnalysisFail, OutOfMemory } {
+        @setCold(true);
+        const src_loc = src.toSrcLocWithDecl(self.decl);
+        self.error_msg = try Module.ErrorMsg.create(self.module.gpa, src_loc, format, args);
+        return error.AnalysisFail;
+    }
+
+    pub fn getOrGenType(self: *DeclGen, t: Type) !u32 {
         // We can't use getOrPut here so we can recursively generate types.
         if (self.types.get(t)) |already_generated| {
             return already_generated;
         }
 
-        const result = self.allocResultId();
+        const result = self.spv.allocResultId();
 
         switch (t.zigTypeTag()) {
-            .Void => try writeInstruction(&self.types_and_globals, .OpTypeVoid, &[_]u32{ result }),
-            .Bool => try writeInstruction(&self.types_and_globals, .OpTypeBool, &[_]u32{ result }),
+            .Void => try writeInstruction(&self.spv.types_and_globals, .OpTypeVoid, &[_]u32{ result }),
+            .Bool => try writeInstruction(&self.spv.types_and_globals, .OpTypeBool, &[_]u32{ result }),
             .Int => {
-                const int_info = t.intInfo(self.target);
-                try writeInstruction(&self.types_and_globals, .OpTypeInt, &[_]u32{
+                const int_info = t.intInfo(self.module.getTarget());
+                // TODO: Capabilities.
+                try writeInstruction(&self.spv.types_and_globals, .OpTypeInt, &[_]u32{
                     result,
                     int_info.bits,
                     switch (int_info.signedness) {
@@ -72,8 +86,8 @@ pub const SPIRVModule = struct {
                     },
                 });
             },
-            // TODO: Verify that floatBits() will be correct.
-            .Float => try writeInstruction(&self.types_and_globals, .OpTypeFloat, &[_]u32{ result, t.floatBits(self.target) }),
+            // TODO: Capabilities.
+            .Float => try writeInstruction(&self.spv.types_and_globals, .OpTypeFloat, &[_]u32{ result, t.floatBits(self.module.getTarget()) }),
             .Null,
             .Undefined,
             .EnumLiteral,
@@ -84,23 +98,23 @@ pub const SPIRVModule = struct {
 
             .BoundFn => unreachable, // this type will be deleted from the language.
 
-            else => return error.TODO,
+            else => |tag| return self.fail(.{ .node_offset = 0 }, "TODO: SPIR-V backend: implement type with tag {}", .{ tag }),
         }
 
         try self.types.put(t, result);
         return result;
     }
 
-    pub fn gen(self: *SPIRVModule, decl: *Decl) !void {
-        const typed_value = decl.typed_value.most_recent.typed_value;
+    pub fn gen(self: *DeclGen) !void {
+        const typed_value = self.decl.typed_value.most_recent.typed_value;
 
         switch (typed_value.ty.zigTypeTag()) {
             .Fn => {
-                log.debug("Generating code for function '{s}'", .{ std.mem.spanZ(decl.name) });
+                log.debug("Generating code for function '{s}'", .{ std.mem.spanZ(self.decl.name) });
 
                 _ = try self.getOrGenType(typed_value.ty.fnReturnType());
             },
-            else => return error.TODO,
+            else => |tag| return self.fail(.{ .node_offset = 0 }, "TODO: SPIR-V backend: generate decl with tag {}", .{ tag }),
         }
     }
 };
diff --git a/src/link/SpirV.zig b/src/link/SpirV.zig
index 95c747c170..8d5feafa5a 100644
--- a/src/link/SpirV.zig
+++ b/src/link/SpirV.zig
@@ -118,8 +118,8 @@ pub fn flushModule(self: *SpirV, comp: *Compilation) !void {
     const module = self.base.options.module.?;
     const target = comp.getTarget();
 
-    var spirv_module = codegen.SPIRVModule.init(target, self.base.allocator);
-    defer spirv_module.deinit();
+    var spv = codegen.SPIRVModule.init(self.base.allocator);
+    defer spv.deinit();
 
     // Allocate an ID for every declaration before generating code,
     // so that we can access them before processing them.
@@ -132,19 +132,41 @@ pub fn flushModule(self: *SpirV, comp: *Compilation) !void {
             if (decl.typed_value != .most_recent)
                 continue;
 
-            decl.fn_link.spirv.id = spirv_module.allocResultId();
+            decl.fn_link.spirv.id = spv.allocResultId();
             log.debug("Allocating id {} to '{s}'", .{ decl.fn_link.spirv.id, std.mem.spanZ(decl.name) });
         }
     }
 
     // Now, actually generate the code for all declarations.
     {
+        // We are just going to re-use this same DeclGen for every Decl, and we are just going to
+        // change the decl. Otherwise, we would have to keep a separate `types`, and re-construct this
+        // structure every time.
+        var decl_gen = codegen.DeclGen{
+            .module = module,
+            .spv = &spv,
+            .types = codegen.TypeMap.init(self.base.allocator),
+            .decl = undefined,
+            .error_msg = undefined,
+        };
+
+        defer decl_gen.types.deinit();
+
         for (module.decl_table.items()) |entry| {
             const decl = entry.value;
             if (decl.typed_value != .most_recent)
                 continue;
 
-            try spirv_module.gen(decl);
+            decl_gen.decl = decl;
+            decl_gen.error_msg = null;
+
+            decl_gen.gen() catch |err| switch (err) {
+                error.AnalysisFail => {
+                    try module.failed_decls.put(module.gpa, decl, decl_gen.error_msg.?);
+                    return;
+                },
+                else => |e| return e,
+            };
         }
     }
 
@@ -155,7 +177,7 @@ pub fn flushModule(self: *SpirV, comp: *Compilation) !void {
         spec.magic_number,
         (spec.version.major << 16) | (spec.version.minor << 8),
         0, // TODO: Register Zig compiler magic number.
-        spirv_module.resultIdBound(), // ID bound.
+        spv.resultIdBound(), // ID bound.
         0, // Schema (currently reserved for future use in the SPIR-V spec).
     });
 
@@ -166,8 +188,8 @@ pub fn flushModule(self: *SpirV, comp: *Compilation) !void {
     // follows the SPIR-V logical module format!
     var all_buffers = [_]std.os.iovec_const{
         wordsToIovConst(binary.items),
-        wordsToIovConst(spirv_module.types_and_globals.items),
-        wordsToIovConst(spirv_module.fn_decls.items),
+        wordsToIovConst(spv.types_and_globals.items),
+        wordsToIovConst(spv.fn_decls.items),
     };
 
     const file = self.base.file.?;

From 458c338aeb0ce72e772c19efcfc633f908ee91b3 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sat, 15 May 2021 01:38:39 +0200
Subject: [PATCH 02/21] SPIR-V: Compute backing integer bits

---
 src/codegen/spirv.zig | 45 +++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 43 insertions(+), 2 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index 9ceaf107d8..fd153533c8 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -2,6 +2,8 @@ const std = @import("std");
 const Allocator = std.mem.Allocator;
 const log = std.log.scoped(.codegen);
 
+const Target = std.Target;
+
 const spec = @import("spirv/spec.zig");
 const Module = @import("../Module.zig");
 const Decl = Module.Decl;
@@ -63,6 +65,43 @@ pub const DeclGen = struct {
         return error.AnalysisFail;
     }
 
+    /// SPIR-V requires enabling specific integer sizes through capabilities, and so if they are not enabled, we need
+    /// to emulate them in other instructions/types. This function returns, given an integer bit width (signed or unsigned, sign
+    /// included), the width of the underlying type which represents it, given the enabled features for the current target.
+    /// If the result is `null`, the largest type the target platform supports natively is not able to perform computations using
+    /// that size. In this case, multiple elements of the largest type should be used.
+    /// The backing type will be chosen as the smallest supported integer larger or equal to it in number of bits.
+    /// The result is valid to be used with OpTypeInt.
+    /// TODO: The extension SPV_INTEL_arbitrary_precision_integers allows any integer size (at least up to 32 bits).
+    /// TODO: This probably needs an ABI-version as well (especially in combination with SPV_INTEL_arbitrary_precision_integers).
+    fn backingIntBits(self: *DeclGen, bits: u32) ?u32 {
+        // TODO: Figure out what to do with u0/i0.
+        std.debug.assert(bits != 0);
+
+        const target = self.module.getTarget();
+
+        // 8, 16 and 64-bit integers require the Int8, Int16 and Inr64 capabilities respectively.
+        const ints = [_]struct{ bits: u32, feature: ?Target.spirv.Feature } {
+            .{ .bits = 8, .feature = .Int8 },
+            .{ .bits = 16, .feature = .Int16 },
+            .{ .bits = 32, .feature = null },
+            .{ .bits = 64, .feature = .Int64 },
+        };
+
+        for (ints) |int| {
+            const has_feature = if (int.feature) |feature|
+                Target.spirv.featureSetHas(target.cpu.features, feature)
+            else
+                true;
+
+            if (bits <= int.bits and has_feature) {
+                return int.bits;
+            }
+        }
+
+        return null;
+    }
+
     pub fn getOrGenType(self: *DeclGen, t: Type) !u32 {
         // We can't use getOrPut here so we can recursively generate types.
         if (self.types.get(t)) |already_generated| {
@@ -76,10 +115,12 @@ pub const DeclGen = struct {
             .Bool => try writeInstruction(&self.spv.types_and_globals, .OpTypeBool, &[_]u32{ result }),
             .Int => {
                 const int_info = t.intInfo(self.module.getTarget());
-                // TODO: Capabilities.
+                const backing_bits = self.backingIntBits(int_info.bits) orelse
+                    return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement fallback for integer of {} bits", .{ int_info.bits });
+
                 try writeInstruction(&self.spv.types_and_globals, .OpTypeInt, &[_]u32{
                     result,
-                    int_info.bits,
+                    backing_bits,
                     switch (int_info.signedness) {
                         .unsigned => 0,
                         .signed => 1,

From 38cdfebad3889853e5393b9f7b63e3c85e9d793f Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sat, 15 May 2021 02:22:12 +0200
Subject: [PATCH 03/21] SPIR-V: Function prototype generation

---
 src/codegen/spirv.zig | 85 ++++++++++++++++++++++++++++++-------------
 1 file changed, 60 insertions(+), 25 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index fd153533c8..076cce4a89 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -12,9 +12,13 @@ const LazySrcLoc = Module.LazySrcLoc;
 
 pub const TypeMap = std.HashMap(Type, u32, Type.hash, Type.eql, std.hash_map.default_max_load_percentage);
 
-pub fn writeInstruction(code: *std.ArrayList(u32), instr: spec.Opcode, args: []const u32) !void {
-    const word_count = @intCast(u32, args.len + 1);
-    try code.append((word_count << 16) | @enumToInt(instr));
+pub fn writeOpcode(code: *std.ArrayList(u32), opcode: spec.Opcode, arg_count: u32) !void {
+    const word_count = arg_count + 1;
+    try code.append((word_count << 16) | @enumToInt(opcode));
+}
+
+pub fn writeInstruction(code: *std.ArrayList(u32), opcode: spec.Opcode, args: []const u32) !void {
+    try writeOpcode(code, opcode, @intCast(u32, args.len));
     try code.appendSlice(args);
 }
 
@@ -58,7 +62,12 @@ pub const DeclGen = struct {
     decl: *Decl,
     error_msg: ?*Module.ErrorMsg,
 
-    fn fail(self: *DeclGen, src: LazySrcLoc, comptime format: []const u8, args: anytype) error{ AnalysisFail, OutOfMemory } {
+    const Error = error{
+        AnalysisFail,
+        OutOfMemory
+    };
+
+    fn fail(self: *DeclGen, src: LazySrcLoc, comptime format: []const u8, args: anytype) Error {
         @setCold(true);
         const src_loc = src.toSrcLocWithDecl(self.decl);
         self.error_msg = try Module.ErrorMsg.create(self.module.gpa, src_loc, format, args);
@@ -102,24 +111,25 @@ pub const DeclGen = struct {
         return null;
     }
 
-    pub fn getOrGenType(self: *DeclGen, t: Type) !u32 {
+    fn getOrGenType(self: *DeclGen, ty: Type) Error!u32 {
         // We can't use getOrPut here so we can recursively generate types.
-        if (self.types.get(t)) |already_generated| {
+        if (self.types.get(ty)) |already_generated| {
             return already_generated;
         }
 
-        const result = self.spv.allocResultId();
+        const code = &self.spv.types_and_globals;
+        const result_id = self.spv.allocResultId();
 
-        switch (t.zigTypeTag()) {
-            .Void => try writeInstruction(&self.spv.types_and_globals, .OpTypeVoid, &[_]u32{ result }),
-            .Bool => try writeInstruction(&self.spv.types_and_globals, .OpTypeBool, &[_]u32{ result }),
+        switch (ty.zigTypeTag()) {
+            .Void => try writeInstruction(code, .OpTypeVoid, &[_]u32{ result_id }),
+            .Bool => try writeInstruction(code, .OpTypeBool, &[_]u32{ result_id }),
             .Int => {
-                const int_info = t.intInfo(self.module.getTarget());
+                const int_info = ty.intInfo(self.module.getTarget());
                 const backing_bits = self.backingIntBits(int_info.bits) orelse
                     return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement fallback for integer of {} bits", .{ int_info.bits });
 
-                try writeInstruction(&self.spv.types_and_globals, .OpTypeInt, &[_]u32{
-                    result,
+                try writeInstruction(code, .OpTypeInt, &[_]u32{
+                    result_id,
                     backing_bits,
                     switch (int_info.signedness) {
                         .unsigned => 0,
@@ -128,7 +138,34 @@ pub const DeclGen = struct {
                 });
             },
             // TODO: Capabilities.
-            .Float => try writeInstruction(&self.spv.types_and_globals, .OpTypeFloat, &[_]u32{ result, t.floatBits(self.module.getTarget()) }),
+            .Float => try writeInstruction(code, .OpTypeFloat, &[_]u32{ result_id, ty.floatBits(self.module.getTarget()) }),
+            .Fn => {
+                // We only support zig-calling-convention functions, no varargs.
+                if (ty.fnCallingConvention() != .Unspecified)
+                    return self.fail(.{.node_offset = 0}, "Invalid calling convention for SPIR-V", .{});
+                if (ty.fnIsVarArgs())
+                    return self.fail(.{.node_offset = 0}, "VarArgs are not supported for SPIR-V", .{});
+
+                // In order to avoid a temporary here, first generate all the required types and then simply look them up
+                // when generating the function type.
+                const params = ty.fnParamLen();
+                var i: usize = 0;
+                while (i < params) : (i += 1) {
+                    _ = try self.getOrGenType(ty.fnParamType(i));
+                }
+
+                const return_type_id = try self.getOrGenType(ty.fnReturnType());
+
+                // result id + result type id + parameter type ids.
+                try writeOpcode(code, .OpTypeFunction, 2 + @intCast(u32, ty.fnParamLen()) );
+                try code.appendSlice(&.{ result_id, return_type_id });
+
+                i = 0;
+                while (i < params) : (i += 1) {
+                    const param_type_id = self.types.get(ty.fnParamType(i)).?;
+                    try code.append(param_type_id);
+                }
+            },
             .Null,
             .Undefined,
             .EnumLiteral,
@@ -139,23 +176,21 @@ pub const DeclGen = struct {
 
             .BoundFn => unreachable, // this type will be deleted from the language.
 
-            else => |tag| return self.fail(.{ .node_offset = 0 }, "TODO: SPIR-V backend: implement type with tag {}", .{ tag }),
+            else => |tag| return self.fail(.{ .node_offset = 0 }, "TODO: SPIR-V backend: implement type {}", .{ tag }),
         }
 
-        try self.types.put(t, result);
-        return result;
+        try self.types.put(ty, result_id);
+        return result_id;
     }
 
     pub fn gen(self: *DeclGen) !void {
-        const typed_value = self.decl.typed_value.most_recent.typed_value;
+        const tv = self.decl.typed_value.most_recent.typed_value;
 
-        switch (typed_value.ty.zigTypeTag()) {
-            .Fn => {
-                log.debug("Generating code for function '{s}'", .{ std.mem.spanZ(self.decl.name) });
-
-                _ = try self.getOrGenType(typed_value.ty.fnReturnType());
-            },
-            else => |tag| return self.fail(.{ .node_offset = 0 }, "TODO: SPIR-V backend: generate decl with tag {}", .{ tag }),
+        if (tv.val.castTag(.function)) |func_payload| {
+            std.debug.assert(tv.ty.zigTypeTag() == .Fn);
+            _ = try self.getOrGenType(tv.ty);
+        } else {
+            return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: generate decl type {}", .{ tv.ty.zigTypeTag() });
         }
     }
 };

From 4403f3598a97d93466c62a91cf3a5aacc6d113a6 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sat, 15 May 2021 02:39:58 +0200
Subject: [PATCH 04/21] SPIR-V: Proper floating point type generation

---
 src/codegen/spirv.zig | 29 ++++++++++++++++++++++++-----
 1 file changed, 24 insertions(+), 5 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index 076cce4a89..b297b1181c 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -117,6 +117,7 @@ pub const DeclGen = struct {
             return already_generated;
         }
 
+        const target = self.module.getTarget();
         const code = &self.spv.types_and_globals;
         const result_id = self.spv.allocResultId();
 
@@ -126,7 +127,7 @@ pub const DeclGen = struct {
             .Int => {
                 const int_info = ty.intInfo(self.module.getTarget());
                 const backing_bits = self.backingIntBits(int_info.bits) orelse
-                    return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement fallback for integer of {} bits", .{ int_info.bits });
+                    return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement fallback for {}", .{ ty });
 
                 try writeInstruction(code, .OpTypeInt, &[_]u32{
                     result_id,
@@ -137,14 +138,32 @@ pub const DeclGen = struct {
                     },
                 });
             },
-            // TODO: Capabilities.
-            .Float => try writeInstruction(code, .OpTypeFloat, &[_]u32{ result_id, ty.floatBits(self.module.getTarget()) }),
+            .Float => {
+                // We can (and want) not really emulate floating points with other floating point types like with the integer types,
+                // so if the float is not supported, just return an error.
+                const bits = ty.floatBits(target);
+                const supported = switch (bits) {
+                    16 => Target.spirv.featureSetHas(target.cpu.features, .Float16),
+                    32 => true,
+                    64 => Target.spirv.featureSetHas(target.cpu.features, .Float64),
+                    else => false
+                };
+
+                if (!supported) {
+                    return self.fail(.{.node_offset = 0}, "Floating point width of {} bits is not supported for the current SPIR-V feature set", .{ bits });
+                }
+
+                try writeInstruction(code, .OpTypeFloat, &.{
+                    result_id,
+                    bits
+                });
+            },
             .Fn => {
                 // We only support zig-calling-convention functions, no varargs.
                 if (ty.fnCallingConvention() != .Unspecified)
-                    return self.fail(.{.node_offset = 0}, "Invalid calling convention for SPIR-V", .{});
+                    return self.fail(.{.node_offset = 0}, "Unsupported calling convention for SPIR-V", .{});
                 if (ty.fnIsVarArgs())
-                    return self.fail(.{.node_offset = 0}, "VarArgs are not supported for SPIR-V", .{});
+                    return self.fail(.{.node_offset = 0}, "VarArgs unsupported for SPIR-V", .{});
 
                 // In order to avoid a temporary here, first generate all the required types and then simply look them up
                 // when generating the function type.

From 074cb9f1daff0f9d8c3bc5736b8990aa53eb8226 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sat, 15 May 2021 03:14:12 +0200
Subject: [PATCH 05/21] SPIR-V: OpFunction/OpFunctionEnd generation

---
 src/codegen/spirv.zig | 25 +++++++++++++++++--------
 1 file changed, 17 insertions(+), 8 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index b297b1181c..db4ddbd657 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -31,7 +31,7 @@ pub const SPIRVModule = struct {
 
     pub fn init(allocator: *Allocator) SPIRVModule {
         return .{
-            .next_result_id = 0,
+            .next_result_id = 1, // 0 is an invalid SPIR-V result ID.
             .types_and_globals = std.ArrayList(u32).init(allocator),
             .fn_decls = std.ArrayList(u32).init(allocator),
         };
@@ -146,17 +146,14 @@ pub const DeclGen = struct {
                     16 => Target.spirv.featureSetHas(target.cpu.features, .Float16),
                     32 => true,
                     64 => Target.spirv.featureSetHas(target.cpu.features, .Float64),
-                    else => false
+                    else => false,
                 };
 
                 if (!supported) {
                     return self.fail(.{.node_offset = 0}, "Floating point width of {} bits is not supported for the current SPIR-V feature set", .{ bits });
                 }
 
-                try writeInstruction(code, .OpTypeFloat, &.{
-                    result_id,
-                    bits
-                });
+                try writeInstruction(code, .OpTypeFloat, &[_]u32{ result_id, bits });
             },
             .Fn => {
                 // We only support zig-calling-convention functions, no varargs.
@@ -195,7 +192,7 @@ pub const DeclGen = struct {
 
             .BoundFn => unreachable, // this type will be deleted from the language.
 
-            else => |tag| return self.fail(.{ .node_offset = 0 }, "TODO: SPIR-V backend: implement type {}", .{ tag }),
+            else => |tag| return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement type {}", .{ tag }),
         }
 
         try self.types.put(ty, result_id);
@@ -203,11 +200,23 @@ pub const DeclGen = struct {
     }
 
     pub fn gen(self: *DeclGen) !void {
+        const result_id = self.decl.fn_link.spirv.id;
         const tv = self.decl.typed_value.most_recent.typed_value;
 
         if (tv.val.castTag(.function)) |func_payload| {
             std.debug.assert(tv.ty.zigTypeTag() == .Fn);
-            _ = try self.getOrGenType(tv.ty);
+            const prototype_id = try self.getOrGenType(tv.ty);
+            try writeInstruction(&self.spv.fn_decls, .OpFunction, &[_]u32{
+                self.types.get(tv.ty.fnReturnType()).?, // This type should be generated along with the prototype.
+                result_id,
+                @bitCast(u32, spec.FunctionControl{}), // TODO: We can set inline here if the type requires it.
+                prototype_id,
+            });
+
+            // TODO: Parameters
+            // TODO: Body
+
+            try writeInstruction(&self.spv.fn_decls, .OpFunctionEnd, &[_]u32{});
         } else {
             return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: generate decl type {}", .{ tv.ty.zigTypeTag() });
         }

From da0cc732ea899d2284200faf54c3c12e8c798b7f Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sat, 15 May 2021 03:27:59 +0200
Subject: [PATCH 06/21] SPIR-V: Function parameter generation

---
 src/codegen/spirv.zig | 13 ++++++++++++-
 src/link/SpirV.zig    |  5 ++++-
 2 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index db4ddbd657..a6b8216774 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -57,6 +57,7 @@ pub const DeclGen = struct {
     module: *Module,
     spv: *SPIRVModule,
 
+    args: std.ArrayList(u32),
     types: TypeMap,
 
     decl: *Decl,
@@ -213,7 +214,17 @@ pub const DeclGen = struct {
                 prototype_id,
             });
 
-            // TODO: Parameters
+            const params = tv.ty.fnParamLen();
+            var i: usize = 0;
+
+            try self.args.ensureCapacity(params);
+            while (i < params) : (i += 1) {
+                const param_type_id = self.types.get(tv.ty.fnParamType(i)).?;
+                const arg_result_id = self.spv.allocResultId();
+                try writeInstruction(&self.spv.fn_decls, .OpFunctionParameter, &[_]u32{ param_type_id, arg_result_id });
+                self.args.appendAssumeCapacity(arg_result_id);
+            }
+
             // TODO: Body
 
             try writeInstruction(&self.spv.fn_decls, .OpFunctionEnd, &[_]u32{});
diff --git a/src/link/SpirV.zig b/src/link/SpirV.zig
index 8d5feafa5a..8bc0d9fe9f 100644
--- a/src/link/SpirV.zig
+++ b/src/link/SpirV.zig
@@ -140,23 +140,26 @@ pub fn flushModule(self: *SpirV, comp: *Compilation) !void {
     // Now, actually generate the code for all declarations.
     {
         // We are just going to re-use this same DeclGen for every Decl, and we are just going to
-        // change the decl. Otherwise, we would have to keep a separate `types`, and re-construct this
+        // change the decl. Otherwise, we would have to keep a separate `args` and `types`, and re-construct this
         // structure every time.
         var decl_gen = codegen.DeclGen{
             .module = module,
             .spv = &spv,
+            .args = std.ArrayList(u32).init(self.base.allocator),
             .types = codegen.TypeMap.init(self.base.allocator),
             .decl = undefined,
             .error_msg = undefined,
         };
 
         defer decl_gen.types.deinit();
+        defer decl_gen.args.deinit();
 
         for (module.decl_table.items()) |entry| {
             const decl = entry.value;
             if (decl.typed_value != .most_recent)
                 continue;
 
+            decl_gen.args.items.len = 0;
             decl_gen.decl = decl;
             decl_gen.error_msg = null;
 

From cbf5280f54509e7aa58d8fd14258274a12efeee1 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sat, 15 May 2021 09:43:57 +0200
Subject: [PATCH 07/21] SPIR-V: Some instructions + constant generation setup

---
 src/codegen/spirv.zig | 150 ++++++++++++++++++++++++++++++++++++------
 src/link/SpirV.zig    |   6 +-
 2 files changed, 135 insertions(+), 21 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index a6b8216774..8be42627ff 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -1,16 +1,19 @@
 const std = @import("std");
 const Allocator = std.mem.Allocator;
-const log = std.log.scoped(.codegen);
-
 const Target = std.Target;
+const log = std.log.scoped(.codegen);
 
 const spec = @import("spirv/spec.zig");
 const Module = @import("../Module.zig");
 const Decl = Module.Decl;
 const Type = @import("../type.zig").Type;
+const Value = @import("../value.zig").Value;
 const LazySrcLoc = Module.LazySrcLoc;
+const ir = @import("../ir.zig");
+const Inst = ir.Inst;
 
 pub const TypeMap = std.HashMap(Type, u32, Type.hash, Type.eql, std.hash_map.default_max_load_percentage);
+pub const ValueMap = std.AutoHashMap(*Inst, u32);
 
 pub fn writeOpcode(code: *std.ArrayList(u32), opcode: spec.Opcode, arg_count: u32) !void {
     const word_count = arg_count + 1;
@@ -26,19 +29,19 @@ pub fn writeInstruction(code: *std.ArrayList(u32), opcode: spec.Opcode, args: []
 /// such as code for the different logical sections, and the next result-id.
 pub const SPIRVModule = struct {
     next_result_id: u32,
-    types_and_globals: std.ArrayList(u32),
+    types_globals_constants: std.ArrayList(u32),
     fn_decls: std.ArrayList(u32),
 
     pub fn init(allocator: *Allocator) SPIRVModule {
         return .{
             .next_result_id = 1, // 0 is an invalid SPIR-V result ID.
-            .types_and_globals = std.ArrayList(u32).init(allocator),
+            .types_globals_constants = std.ArrayList(u32).init(allocator),
             .fn_decls = std.ArrayList(u32).init(allocator),
         };
     }
 
     pub fn deinit(self: *SPIRVModule) void {
-        self.types_and_globals.deinit();
+        self.types_globals_constants.deinit();
         self.fn_decls.deinit();
     }
 
@@ -58,7 +61,10 @@ pub const DeclGen = struct {
     spv: *SPIRVModule,
 
     args: std.ArrayList(u32),
+    next_arg_index: u32,
+
     types: TypeMap,
+    values: ValueMap,
 
     decl: *Decl,
     error_msg: ?*Module.ErrorMsg,
@@ -75,6 +81,14 @@ pub const DeclGen = struct {
         return error.AnalysisFail;
     }
 
+    fn resolve(self: *DeclGen, inst: *Inst) !u32 {
+        if (inst.value()) |val| {
+            return self.genConstant(inst.ty, val);
+        }
+
+        return self.values.get(inst).?; // Instruction does not dominate all uses!
+    }
+
     /// SPIR-V requires enabling specific integer sizes through capabilities, and so if they are not enabled, we need
     /// to emulate them in other instructions/types. This function returns, given an integer bit width (signed or unsigned, sign
     /// included), the width of the underlying type which represents it, given the enabled features for the current target.
@@ -82,13 +96,16 @@ pub const DeclGen = struct {
     /// that size. In this case, multiple elements of the largest type should be used.
     /// The backing type will be chosen as the smallest supported integer larger or equal to it in number of bits.
     /// The result is valid to be used with OpTypeInt.
+    /// asserts `ty` is an integer.
     /// TODO: The extension SPV_INTEL_arbitrary_precision_integers allows any integer size (at least up to 32 bits).
     /// TODO: This probably needs an ABI-version as well (especially in combination with SPV_INTEL_arbitrary_precision_integers).
-    fn backingIntBits(self: *DeclGen, bits: u32) ?u32 {
-        // TODO: Figure out what to do with u0/i0.
-        std.debug.assert(bits != 0);
-
+    /// TODO: Should the result of this function be cached?
+    fn backingIntBits(self: *DeclGen, ty: Type) ?u32 {
         const target = self.module.getTarget();
+        const int_info = ty.intInfo(target);
+
+        // TODO: Figure out what to do with u0/i0.
+        std.debug.assert(int_info.bits != 0);
 
         // 8, 16 and 64-bit integers require the Int8, Int16 and Inr64 capabilities respectively.
         const ints = [_]struct{ bits: u32, feature: ?Target.spirv.Feature } {
@@ -104,7 +121,7 @@ pub const DeclGen = struct {
             else
                 true;
 
-            if (bits <= int.bits and has_feature) {
+            if (int_info.bits <= int.bits and has_feature) {
                 return int.bits;
             }
         }
@@ -112,6 +129,43 @@ pub const DeclGen = struct {
         return null;
     }
 
+    /// Return the amount of bits in the largest supported integer type. This is either 32 (always supported), or 64 (if
+    /// the Int64 capability is enabled).
+    /// Note: The extension SPV_INTEL_arbitrary_precision_integers allows any integer size (at least up to 32 bits).
+    /// In theory that could also be used, but since the spec says that it only guarantees support up to 32-bit ints there
+    /// is no way of knowing whether those are actually supported.
+    /// TODO: Maybe this should be cached?
+    fn largestSupportedIntBits(self: *DeclGen) u32 {
+        const target = self.module.getTarget();
+        return if (Target.spirv.featureSetHas(target.cpu.features, .Int64))
+            64
+        else
+            32;
+    }
+
+    /// Generate a constant representing `val`.
+    /// TODO: Deduplication?
+    fn genConstant(self: *DeclGen, ty: Type, val: Value) Error!u32 {
+        const code = &self.spv.types_globals_constants;
+        const result_id = self.spv.allocResultId();
+        const result_type_id = try self.getOrGenType(ty);
+
+        if (val.isUndef()) {
+            try writeInstruction(code, .OpUndef, &[_]u32{ result_type_id, result_id });
+            return result_id;
+        }
+
+        switch (ty.zigTypeTag()) {
+            .Bool => {
+                const opcode: spec.Opcode = if (val.toBool()) .OpConstantTrue else .OpConstantFalse;
+                try writeInstruction(code, opcode, &[_]u32{ result_type_id, result_id });
+            },
+            else => return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: constant generation of type {s}\n", .{ ty.zigTypeTag() }),
+        }
+
+        return result_id;
+    }
+
     fn getOrGenType(self: *DeclGen, ty: Type) Error!u32 {
         // We can't use getOrPut here so we can recursively generate types.
         if (self.types.get(ty)) |already_generated| {
@@ -119,24 +173,21 @@ pub const DeclGen = struct {
         }
 
         const target = self.module.getTarget();
-        const code = &self.spv.types_and_globals;
+        const code = &self.spv.types_globals_constants;
         const result_id = self.spv.allocResultId();
 
         switch (ty.zigTypeTag()) {
             .Void => try writeInstruction(code, .OpTypeVoid, &[_]u32{ result_id }),
             .Bool => try writeInstruction(code, .OpTypeBool, &[_]u32{ result_id }),
             .Int => {
-                const int_info = ty.intInfo(self.module.getTarget());
-                const backing_bits = self.backingIntBits(int_info.bits) orelse
+                const backing_bits = self.backingIntBits(ty) orelse
                     return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement fallback for {}", .{ ty });
 
+                // TODO: If backing_bits != int_info.bits, a duplicate type might be generated here.
                 try writeInstruction(code, .OpTypeInt, &[_]u32{
                     result_id,
                     backing_bits,
-                    switch (int_info.signedness) {
-                        .unsigned => 0,
-                        .signed => 1,
-                    },
+                    @boolToInt(ty.isSignedInt()),
                 });
             },
             .Float => {
@@ -183,6 +234,15 @@ pub const DeclGen = struct {
                     try code.append(param_type_id);
                 }
             },
+            .Vector => {
+                // Although not 100% the same, Zig vectors map quite neatly to SPIR-V vectors (including many integer and float operations
+                // which work on them), so simply use those.
+                // Note: SPIR-V vectors only support bools, ints and floats, so pointer vectors need to be supported another way.
+                // "big integers" (larger than the largest supported native type) can probably be represented by an array of vectors.
+
+                // TODO: Vectors are not yet supported by the self-hosted compiler itself it seems.
+                return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement type Vector", .{});
+            },
             .Null,
             .Undefined,
             .EnumLiteral,
@@ -193,10 +253,10 @@ pub const DeclGen = struct {
 
             .BoundFn => unreachable, // this type will be deleted from the language.
 
-            else => |tag| return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement type {}", .{ tag }),
+            else => |tag| return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement type {}s", .{ tag }),
         }
 
-        try self.types.put(ty, result_id);
+        try self.types.putNoClobber(ty, result_id);
         return result_id;
     }
 
@@ -225,11 +285,61 @@ pub const DeclGen = struct {
                 self.args.appendAssumeCapacity(arg_result_id);
             }
 
-            // TODO: Body
+            // TODO: This could probably be done in a better way...
+            const root_block_id = self.spv.allocResultId();
+            _ = try writeInstruction(&self.spv.fn_decls, .OpLabel, &[_]u32{root_block_id});
+            try self.genBody(func_payload.data.body);
 
             try writeInstruction(&self.spv.fn_decls, .OpFunctionEnd, &[_]u32{});
         } else {
             return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: generate decl type {}", .{ tv.ty.zigTypeTag() });
         }
     }
+
+    fn genBody(self: *DeclGen, body: ir.Body) !void {
+        for (body.instructions) |inst| {
+            const maybe_result_id = try self.genInst(inst);
+            if (maybe_result_id) |result_id|
+                try self.values.putNoClobber(inst, result_id);
+        }
+    }
+
+    fn genInst(self: *DeclGen, inst: *Inst) !?u32 {
+        return switch (inst.tag) {
+            .arg => self.genArg(),
+            // TODO: Breakpoints won't be supported in SPIR-V, but the compiler seems to insert them
+            // throughout the IR.
+            .breakpoint => null,
+            // TODO: What does this entail?
+            .dbg_stmt => null,
+            .ret => self.genRet(inst.castTag(.ret).?),
+            .retvoid => self.genRetVoid(),
+            .unreach => self.genUnreach(),
+            else => self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement inst {}", .{inst.tag}),
+        };
+    }
+
+    fn genArg(self: *DeclGen) u32 {
+        defer self.next_arg_index += 1;
+        return self.args.items[self.next_arg_index];
+    }
+
+    fn genRet(self: *DeclGen, inst: *Inst.UnOp) !?u32 {
+        const operand_id = try self.resolve(inst.operand);
+        // TODO: This instruction needs to be the last in a block. Is that guaranteed?
+        try writeInstruction(&self.spv.fn_decls, .OpReturnValue, &[_]u32{ operand_id });
+        return null;
+    }
+
+    fn genRetVoid(self: *DeclGen) !?u32 {
+        // TODO: This instruction needs to be the last in a block. Is that guaranteed?
+        try writeInstruction(&self.spv.fn_decls, .OpReturn, &[_]u32{});
+        return null;
+    }
+
+    fn genUnreach(self: *DeclGen) !?u32 {
+        // TODO: This instruction needs to be the last in a block. Is that guaranteed?
+        try writeInstruction(&self.spv.fn_decls, .OpUnreachable, &[_]u32{});
+        return null;
+    }
 };
diff --git a/src/link/SpirV.zig b/src/link/SpirV.zig
index 8bc0d9fe9f..96ed2d3e77 100644
--- a/src/link/SpirV.zig
+++ b/src/link/SpirV.zig
@@ -146,11 +146,14 @@ pub fn flushModule(self: *SpirV, comp: *Compilation) !void {
             .module = module,
             .spv = &spv,
             .args = std.ArrayList(u32).init(self.base.allocator),
+            .next_arg_index = undefined,
             .types = codegen.TypeMap.init(self.base.allocator),
+            .values = codegen.ValueMap.init(self.base.allocator),
             .decl = undefined,
             .error_msg = undefined,
         };
 
+        defer decl_gen.values.deinit();
         defer decl_gen.types.deinit();
         defer decl_gen.args.deinit();
 
@@ -160,6 +163,7 @@ pub fn flushModule(self: *SpirV, comp: *Compilation) !void {
                 continue;
 
             decl_gen.args.items.len = 0;
+            decl_gen.next_arg_index = 0;
             decl_gen.decl = decl;
             decl_gen.error_msg = null;
 
@@ -191,7 +195,7 @@ pub fn flushModule(self: *SpirV, comp: *Compilation) !void {
     // follows the SPIR-V logical module format!
     var all_buffers = [_]std.os.iovec_const{
         wordsToIovConst(binary.items),
-        wordsToIovConst(spv.types_and_globals.items),
+        wordsToIovConst(spv.types_globals_constants.items),
         wordsToIovConst(spv.fn_decls.items),
     };
 

From ae2e21639a958b0bc8c085c8eac6733aaa933457 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sat, 15 May 2021 14:04:41 +0200
Subject: [PATCH 08/21] SPIR-V: Some initial floating point constant generation

---
 src/codegen/spirv.zig | 47 +++++++++++++++++++++++++++++++++++++++----
 1 file changed, 43 insertions(+), 4 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index 8be42627ff..f4f437a064 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -143,6 +143,13 @@ pub const DeclGen = struct {
             32;
     }
 
+    /// Checks whether the type is "composite int", an integer consisting of multiple native integers. These are represented by
+    /// arrays of largestSupportedIntBits().
+    /// Asserts `ty` is an integer.
+    fn isCompositeInt(self: *DeclGen, ty: Type) bool {
+        return self.backingIntBits(ty) == null;
+    }
+
     /// Generate a constant representing `val`.
     /// TODO: Deduplication?
     fn genConstant(self: *DeclGen, ty: Type, val: Value) Error!u32 {
@@ -160,6 +167,37 @@ pub const DeclGen = struct {
                 const opcode: spec.Opcode = if (val.toBool()) .OpConstantTrue else .OpConstantFalse;
                 try writeInstruction(code, opcode, &[_]u32{ result_type_id, result_id });
             },
+            .Float => {
+                // At this point we are guaranteed that the target floating point type is supported, otherwise the function
+                // would have exited at getOrGenType(ty).
+
+                // f16 and f32 require one word of storage. f64 requires 2, low-order first.
+
+                switch (val.tag()) {
+                    .float_16 => try writeInstruction(code, .OpConstant, &[_]u32{
+                        result_type_id,
+                        result_id,
+                        @bitCast(u16, val.castTag(.float_16).?.data)
+                    }),
+                    .float_32 => try writeInstruction(code, .OpConstant, &[_]u32{
+                        result_type_id,
+                        result_id,
+                        @bitCast(u32, val.castTag(.float_32).?.data)
+                    }),
+                    .float_64 => {
+                        const float_bits = @bitCast(u64, val.castTag(.float_64).?.data);
+                        try writeInstruction(code, .OpConstant, &[_]u32{
+                            result_type_id,
+                            result_id,
+                            @truncate(u32, float_bits),
+                            @truncate(u32, float_bits >> 32),
+                        });
+                    },
+                    .float_128 => unreachable, // Filtered out in the call to getOrGenType.
+                    // TODO: What tags do we need to handle here anyway?
+                    else => return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: float constant generation of value {s}\n", .{ val.tag() }),
+                }
+            },
             else => return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: constant generation of type {s}\n", .{ ty.zigTypeTag() }),
         }
 
@@ -180,8 +218,10 @@ pub const DeclGen = struct {
             .Void => try writeInstruction(code, .OpTypeVoid, &[_]u32{ result_id }),
             .Bool => try writeInstruction(code, .OpTypeBool, &[_]u32{ result_id }),
             .Int => {
-                const backing_bits = self.backingIntBits(ty) orelse
-                    return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement fallback for {}", .{ ty });
+                const backing_bits = self.backingIntBits(ty) orelse {
+                    // Integers too big for any native type are represented as "composite integers": An array of largestSupportedIntBits.
+                    return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement composite ints {}", .{ ty });
+                };
 
                 // TODO: If backing_bits != int_info.bits, a duplicate type might be generated here.
                 try writeInstruction(code, .OpTypeInt, &[_]u32{
@@ -238,7 +278,7 @@ pub const DeclGen = struct {
                 // Although not 100% the same, Zig vectors map quite neatly to SPIR-V vectors (including many integer and float operations
                 // which work on them), so simply use those.
                 // Note: SPIR-V vectors only support bools, ints and floats, so pointer vectors need to be supported another way.
-                // "big integers" (larger than the largest supported native type) can probably be represented by an array of vectors.
+                // "composite integers" (larger than the largest supported native type) can probably be represented by an array of vectors.
 
                 // TODO: Vectors are not yet supported by the self-hosted compiler itself it seems.
                 return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement type Vector", .{});
@@ -310,7 +350,6 @@ pub const DeclGen = struct {
             // TODO: Breakpoints won't be supported in SPIR-V, but the compiler seems to insert them
             // throughout the IR.
             .breakpoint => null,
-            // TODO: What does this entail?
             .dbg_stmt => null,
             .ret => self.genRet(inst.castTag(.ret).?),
             .retvoid => self.genRetVoid(),

From 10678af8768a8d8cad7640837d0ec354dc8c07bc Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sun, 16 May 2021 13:09:32 +0200
Subject: [PATCH 09/21] SPIR-V: genBinOp setup

---
 src/codegen/spirv.zig | 118 +++++++++++++++++++++++++++++++++++++++---
 1 file changed, 111 insertions(+), 7 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index f4f437a064..3272de47ae 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -74,6 +74,44 @@ pub const DeclGen = struct {
         OutOfMemory
     };
 
+    /// This structure is used to return information about a type typically used for arithmetic operations.
+    /// These types may either be integers, floats, or a vector of these. Most scalar operations also work on vectors,
+    /// so we can easily represent those as arithmetic types.
+    /// If the type is a scalar, 'inner type' refers to the scalar type. Otherwise, if its a vector, it refers
+    /// to the vector's element type.
+    const ArithmeticTypeInfo = struct {
+        /// A classification of the inner type.
+        const Class = enum {
+            /// A regular, **native**, integer operation.
+            /// This is only returned when the backend supports this int as a native type (when
+            /// the relevant capability is enabled).
+            integer,
+
+            /// A regular float. These are all required to be natively supported. Floating points for
+            /// which the relevant capability is not enabled are not emulated.
+            float,
+
+            /// An integer of a 'strange' size (which' bit size is not the same as its backing type. **Note**: this
+            /// may **also** include power-of-2 integers for which the relevant capability is not enabled), but still
+            /// within the limits of the largest natively supported integer type.
+            strange_integer,
+
+            /// An integer with more bits than the largest natively supported integer type.
+            composite_integer,
+        };
+
+        /// The number of bits in the inner type.
+        /// Note: this is the actual number of bits of the type, not the size of the backing integer.
+        bits: u32,
+
+        /// Whether the type is a vector.
+        is_vector: bool,
+
+        /// A classification of the inner type. These four scenarios
+        /// will all have to be handled slightly different.
+        class: Class,
+    };
+
     fn fail(self: *DeclGen, src: LazySrcLoc, comptime format: []const u8, args: anytype) Error {
         @setCold(true);
         const src_loc = src.toSrcLocWithDecl(self.decl);
@@ -96,16 +134,14 @@ pub const DeclGen = struct {
     /// that size. In this case, multiple elements of the largest type should be used.
     /// The backing type will be chosen as the smallest supported integer larger or equal to it in number of bits.
     /// The result is valid to be used with OpTypeInt.
-    /// asserts `ty` is an integer.
     /// TODO: The extension SPV_INTEL_arbitrary_precision_integers allows any integer size (at least up to 32 bits).
     /// TODO: This probably needs an ABI-version as well (especially in combination with SPV_INTEL_arbitrary_precision_integers).
     /// TODO: Should the result of this function be cached?
-    fn backingIntBits(self: *DeclGen, ty: Type) ?u32 {
+    fn backingIntBits(self: *DeclGen, bits: u32) ?u32 {
         const target = self.module.getTarget();
-        const int_info = ty.intInfo(target);
 
         // TODO: Figure out what to do with u0/i0.
-        std.debug.assert(int_info.bits != 0);
+        std.debug.assert(bits != 0);
 
         // 8, 16 and 64-bit integers require the Int8, Int16 and Inr64 capabilities respectively.
         const ints = [_]struct{ bits: u32, feature: ?Target.spirv.Feature } {
@@ -121,7 +157,7 @@ pub const DeclGen = struct {
             else
                 true;
 
-            if (int_info.bits <= int.bits and has_feature) {
+            if (bits <= int.bits and has_feature) {
                 return int.bits;
             }
         }
@@ -150,6 +186,34 @@ pub const DeclGen = struct {
         return self.backingIntBits(ty) == null;
     }
 
+    fn arithmeticTypeInfo(self: *DeclGen, ty: Type) !ArithmeticTypeInfo {
+        const target = self.module.getTarget();
+
+        return switch (ty.zigTypeTag()) {
+            .Float => ArithmeticTypeInfo{ .bits = ty.floatBits(target), .is_vector = false, .class = .float },
+            .Int => blk: {
+                const int_info = ty.intInfo(target);
+                // TODO: Maybe it's useful to also return this value.
+                const maybe_backing_bits = self.backingIntBits(int_info.bits);
+                break :blk ArithmeticTypeInfo{
+                    .bits = int_info.bits,
+                    .is_vector = false,
+                    .class = if (maybe_backing_bits) |backing_bits|
+                            if (backing_bits == int_info.bits)
+                                ArithmeticTypeInfo.Class.integer
+                            else
+                                ArithmeticTypeInfo.Class.strange_integer
+                        else
+                            .composite_integer
+                };
+            },
+            // As of yet, there is no vector support in the self-hosted compiler.
+            .Vector => self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement arithmeticTypeInfo for Vector", .{}),
+            // TODO: For which types is this the case?
+            else => self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement arithmeticTypeInfo for {}", .{ty}),
+        };
+    }
+
     /// Generate a constant representing `val`.
     /// TODO: Deduplication?
     fn genConstant(self: *DeclGen, ty: Type, val: Value) Error!u32 {
@@ -218,7 +282,8 @@ pub const DeclGen = struct {
             .Void => try writeInstruction(code, .OpTypeVoid, &[_]u32{ result_id }),
             .Bool => try writeInstruction(code, .OpTypeBool, &[_]u32{ result_id }),
             .Int => {
-                const backing_bits = self.backingIntBits(ty) orelse {
+                const int_info = ty.intInfo(target);
+                const backing_bits = self.backingIntBits(int_info.bits) orelse {
                     // Integers too big for any native type are represented as "composite integers": An array of largestSupportedIntBits.
                     return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement composite ints {}", .{ ty });
                 };
@@ -227,7 +292,10 @@ pub const DeclGen = struct {
                 try writeInstruction(code, .OpTypeInt, &[_]u32{
                     result_id,
                     backing_bits,
-                    @boolToInt(ty.isSignedInt()),
+                    switch (int_info.signedness) {
+                        .unsigned => 0,
+                        .signed => 1,
+                    },
                 });
             },
             .Float => {
@@ -346,6 +414,7 @@ pub const DeclGen = struct {
 
     fn genInst(self: *DeclGen, inst: *Inst) !?u32 {
         return switch (inst.tag) {
+            .add => try self.genBinOp(inst.castTag(.add).?),
             .arg => self.genArg(),
             // TODO: Breakpoints won't be supported in SPIR-V, but the compiler seems to insert them
             // throughout the IR.
@@ -358,6 +427,41 @@ pub const DeclGen = struct {
         };
     }
 
+    fn genBinOp(self: *DeclGen, inst: *Inst.BinOp) !u32 {
+        // TODO: Will lhs and rhs have the same type?
+        const lhs_id = try self.resolve(inst.lhs);
+        const rhs_id = try self.resolve(inst.rhs);
+
+        const binop_result_id = self.spv.allocResultId();
+        const result_type_id = try self.getOrGenType(inst.base.ty);
+
+        // TODO: Is the result the same as the argument types?
+        // This is supposed to be the case for SPIR-V.
+        std.debug.assert(inst.base.ty.eql(inst.lhs.ty) and inst.base.ty.eql(inst.rhs.ty));
+
+        // Binary operations are generally applicable to both scalar and vector operations in SPIR-V, but int and float
+        // versions of operations require different opcodes.
+        const info = try self.arithmeticTypeInfo(inst.base.ty);
+
+        if (info.class == .composite_integer)
+            return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: binary operations for composite integers", .{});
+
+        // Fetch the integer and float opcodes for each operation.
+        // Doing it this way removes a bit of code clutter.
+        const opcodes: [2]spec.Opcode = switch (inst.base.tag) {
+            .add => .{.OpIAdd, .OpFAdd},
+            else => unreachable,
+        };
+
+        const opcode = if (info.class == .float) opcodes[1] else opcodes[0];
+        try writeInstruction(&self.spv.fn_decls, opcode, &[_]u32{ result_type_id, binop_result_id, lhs_id, rhs_id });
+
+        if (info.class != .strange_integer)
+            return binop_result_id;
+
+        return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: strange integer operation mask", .{});
+    }
+
     fn genArg(self: *DeclGen) u32 {
         defer self.next_arg_index += 1;
         return self.args.items[self.next_arg_index];

From 4735e95d1699c90e821655bdbe0afbb1044738ea Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sun, 16 May 2021 13:32:32 +0200
Subject: [PATCH 10/21] SPIR-V: More binary operations

---
 src/codegen/spirv.zig | 53 +++++++++++++++++++++++++++++++------------
 1 file changed, 39 insertions(+), 14 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index 3272de47ae..c8bebbb70b 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -4,6 +4,8 @@ const Target = std.Target;
 const log = std.log.scoped(.codegen);
 
 const spec = @import("spirv/spec.zig");
+const Opcode = spec.Opcode;
+
 const Module = @import("../Module.zig");
 const Decl = Module.Decl;
 const Type = @import("../type.zig").Type;
@@ -15,12 +17,12 @@ const Inst = ir.Inst;
 pub const TypeMap = std.HashMap(Type, u32, Type.hash, Type.eql, std.hash_map.default_max_load_percentage);
 pub const ValueMap = std.AutoHashMap(*Inst, u32);
 
-pub fn writeOpcode(code: *std.ArrayList(u32), opcode: spec.Opcode, arg_count: u32) !void {
+pub fn writeOpcode(code: *std.ArrayList(u32), opcode: Opcode, arg_count: u32) !void {
     const word_count = arg_count + 1;
     try code.append((word_count << 16) | @enumToInt(opcode));
 }
 
-pub fn writeInstruction(code: *std.ArrayList(u32), opcode: spec.Opcode, args: []const u32) !void {
+pub fn writeInstruction(code: *std.ArrayList(u32), opcode: Opcode, args: []const u32) !void {
     try writeOpcode(code, opcode, @intCast(u32, args.len));
     try code.appendSlice(args);
 }
@@ -102,11 +104,14 @@ pub const DeclGen = struct {
 
         /// The number of bits in the inner type.
         /// Note: this is the actual number of bits of the type, not the size of the backing integer.
-        bits: u32,
+        bits: u16,
 
         /// Whether the type is a vector.
         is_vector: bool,
 
+        /// Whether the inner type is signed. Only relevant for integers.
+        signedness: std.builtin.Signedness,
+
         /// A classification of the inner type. These four scenarios
         /// will all have to be handled slightly different.
         class: Class,
@@ -137,14 +142,14 @@ pub const DeclGen = struct {
     /// TODO: The extension SPV_INTEL_arbitrary_precision_integers allows any integer size (at least up to 32 bits).
     /// TODO: This probably needs an ABI-version as well (especially in combination with SPV_INTEL_arbitrary_precision_integers).
     /// TODO: Should the result of this function be cached?
-    fn backingIntBits(self: *DeclGen, bits: u32) ?u32 {
+    fn backingIntBits(self: *DeclGen, bits: u16) ?u16 {
         const target = self.module.getTarget();
 
         // TODO: Figure out what to do with u0/i0.
         std.debug.assert(bits != 0);
 
         // 8, 16 and 64-bit integers require the Int8, Int16 and Inr64 capabilities respectively.
-        const ints = [_]struct{ bits: u32, feature: ?Target.spirv.Feature } {
+        const ints = [_]struct{ bits: u16, feature: ?Target.spirv.Feature } {
             .{ .bits = 8, .feature = .Int8 },
             .{ .bits = 16, .feature = .Int16 },
             .{ .bits = 32, .feature = null },
@@ -171,7 +176,7 @@ pub const DeclGen = struct {
     /// In theory that could also be used, but since the spec says that it only guarantees support up to 32-bit ints there
     /// is no way of knowing whether those are actually supported.
     /// TODO: Maybe this should be cached?
-    fn largestSupportedIntBits(self: *DeclGen) u32 {
+    fn largestSupportedIntBits(self: *DeclGen) u16 {
         const target = self.module.getTarget();
         return if (Target.spirv.featureSetHas(target.cpu.features, .Int64))
             64
@@ -190,7 +195,12 @@ pub const DeclGen = struct {
         const target = self.module.getTarget();
 
         return switch (ty.zigTypeTag()) {
-            .Float => ArithmeticTypeInfo{ .bits = ty.floatBits(target), .is_vector = false, .class = .float },
+            .Float => ArithmeticTypeInfo{
+                .bits = ty.floatBits(target),
+                .is_vector = false,
+                .signedness = .signed, // I guess technically it is.
+                .class = .float
+            },
             .Int => blk: {
                 const int_info = ty.intInfo(target);
                 // TODO: Maybe it's useful to also return this value.
@@ -198,6 +208,7 @@ pub const DeclGen = struct {
                 break :blk ArithmeticTypeInfo{
                     .bits = int_info.bits,
                     .is_vector = false,
+                    .signedness = int_info.signedness,
                     .class = if (maybe_backing_bits) |backing_bits|
                             if (backing_bits == int_info.bits)
                                 ArithmeticTypeInfo.Class.integer
@@ -228,7 +239,7 @@ pub const DeclGen = struct {
 
         switch (ty.zigTypeTag()) {
             .Bool => {
-                const opcode: spec.Opcode = if (val.toBool()) .OpConstantTrue else .OpConstantFalse;
+                const opcode: Opcode = if (val.toBool()) .OpConstantTrue else .OpConstantFalse;
                 try writeInstruction(code, opcode, &[_]u32{ result_type_id, result_id });
             },
             .Float => {
@@ -414,7 +425,10 @@ pub const DeclGen = struct {
 
     fn genInst(self: *DeclGen, inst: *Inst) !?u32 {
         return switch (inst.tag) {
-            .add => try self.genBinOp(inst.castTag(.add).?),
+            .add, .addwrap => try self.genBinOp(inst.castTag(.add).?),
+            .sub, .subwrap => try self.genBinOp(inst.castTag(.sub).?),
+            .mul, .mulwrap => try self.genBinOp(inst.castTag(.mul).?),
+            .div => try self.genBinOp(inst.castTag(.div).?),
             .arg => self.genArg(),
             // TODO: Breakpoints won't be supported in SPIR-V, but the compiler seems to insert them
             // throughout the IR.
@@ -446,16 +460,27 @@ pub const DeclGen = struct {
         if (info.class == .composite_integer)
             return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: binary operations for composite integers", .{});
 
-        // Fetch the integer and float opcodes for each operation.
-        // Doing it this way removes a bit of code clutter.
-        const opcodes: [2]spec.Opcode = switch (inst.base.tag) {
-            .add => .{.OpIAdd, .OpFAdd},
+        const is_float = info.class == .float;
+        const is_signed = info.signedness == .signed;
+        // **Note**: All these operations must be valid for vectors of floats and integers as well!
+        const opcode = switch (inst.base.tag) {
+            // The regular integer operations are all defined for wrapping. Since theyre only relevant for integers,
+            // we can just switch on both cases here.
+            .add, .addwrap => if (is_float) Opcode.OpFAdd else Opcode.OpIAdd,
+            .sub, .subwrap => if (is_float) Opcode.OpFSub else Opcode.OpISub,
+            .mul, .mulwrap => if (is_float) Opcode.OpFMul else Opcode.OpIMul,
+            // TODO: Trap if divisor is 0?
+            // TODO: Figure out of OpSDiv for unsigned/OpUDiv for signed does anything useful.
+            .div => if (is_float) Opcode.OpFDiv else if (is_signed) Opcode.OpSDiv else Opcode.OpUDiv,
+
             else => unreachable,
         };
 
-        const opcode = if (info.class == .float) opcodes[1] else opcodes[0];
         try writeInstruction(&self.spv.fn_decls, opcode, &[_]u32{ result_type_id, binop_result_id, lhs_id, rhs_id });
 
+        // TODO: Trap on overflow? Probably going to be annoying.
+        // TODO: Look into NoSignedWrap/NoUnsignedWrap extensions.
+
         if (info.class != .strange_integer)
             return binop_result_id;
 

From f14000c7e1bd9032185b3c0a2a28a73eec6c48f7 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sun, 16 May 2021 13:39:17 +0200
Subject: [PATCH 11/21] SPIR-V: More bitwise binary operations

---
 src/codegen/spirv.zig | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index c8bebbb70b..88d6f74a2d 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -84,7 +84,7 @@ pub const DeclGen = struct {
     const ArithmeticTypeInfo = struct {
         /// A classification of the inner type.
         const Class = enum {
-            /// A regular, **native**, integer operation.
+            /// A regular, **native**, integer.
             /// This is only returned when the backend supports this int as a native type (when
             /// the relevant capability is enabled).
             integer,
@@ -112,7 +112,7 @@ pub const DeclGen = struct {
         /// Whether the inner type is signed. Only relevant for integers.
         signedness: std.builtin.Signedness,
 
-        /// A classification of the inner type. These four scenarios
+        /// A classification of the inner type. These scenarios
         /// will all have to be handled slightly different.
         class: Class,
     };
@@ -149,6 +149,7 @@ pub const DeclGen = struct {
         std.debug.assert(bits != 0);
 
         // 8, 16 and 64-bit integers require the Int8, Int16 and Inr64 capabilities respectively.
+        // 32-bit integers are always supported (see spec, 2.16.1, Data rules).
         const ints = [_]struct{ bits: u16, feature: ?Target.spirv.Feature } {
             .{ .bits = 8, .feature = .Int8 },
             .{ .bits = 16, .feature = .Int16 },
@@ -198,8 +199,8 @@ pub const DeclGen = struct {
             .Float => ArithmeticTypeInfo{
                 .bits = ty.floatBits(target),
                 .is_vector = false,
-                .signedness = .signed, // I guess technically it is.
-                .class = .float
+                .signedness = .signed, // Technically, but doesn't matter for this class.
+                .class = .float,
             },
             .Int => blk: {
                 const int_info = ty.intInfo(target);
@@ -315,6 +316,7 @@ pub const DeclGen = struct {
                 const bits = ty.floatBits(target);
                 const supported = switch (bits) {
                     16 => Target.spirv.featureSetHas(target.cpu.features, .Float16),
+                    // 32-bit floats are always supported (see spec, 2.16.1, Data rules).
                     32 => true,
                     64 => Target.spirv.featureSetHas(target.cpu.features, .Float64),
                     else => false,
@@ -358,6 +360,8 @@ pub const DeclGen = struct {
                 // which work on them), so simply use those.
                 // Note: SPIR-V vectors only support bools, ints and floats, so pointer vectors need to be supported another way.
                 // "composite integers" (larger than the largest supported native type) can probably be represented by an array of vectors.
+                // TODO: The SPIR-V spec mentions that vector sizes may be quite restricted! look into which we can use, and whether OpTypeVector
+                // is adequate at all for this.
 
                 // TODO: Vectors are not yet supported by the self-hosted compiler itself it seems.
                 return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: implement type Vector", .{});
@@ -429,6 +433,9 @@ pub const DeclGen = struct {
             .sub, .subwrap => try self.genBinOp(inst.castTag(.sub).?),
             .mul, .mulwrap => try self.genBinOp(inst.castTag(.mul).?),
             .div => try self.genBinOp(inst.castTag(.div).?),
+            .bit_and => try self.genBinOp(inst.castTag(.bit_and).?),
+            .bit_or => try self.genBinOp(inst.castTag(.bit_or).?),
+            .xor => try self.genBinOp(inst.castTag(.xor).?),
             .arg => self.genArg(),
             // TODO: Breakpoints won't be supported in SPIR-V, but the compiler seems to insert them
             // throughout the IR.
@@ -472,7 +479,10 @@ pub const DeclGen = struct {
             // TODO: Trap if divisor is 0?
             // TODO: Figure out of OpSDiv for unsigned/OpUDiv for signed does anything useful.
             .div => if (is_float) Opcode.OpFDiv else if (is_signed) Opcode.OpSDiv else Opcode.OpUDiv,
-
+            // Only integer versions for these.
+            .bit_and => Opcode.OpBitwiseAnd,
+            .bit_or => Opcode.OpBitwiseOr,
+            .xor => Opcode.OpBitwiseXor,
             else => unreachable,
         };
 

From 585122b1ac51f9ac23bae537dfc40bbae1d7cb3c Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sun, 16 May 2021 14:46:58 +0200
Subject: [PATCH 12/21] SPIR-V: comparison and equality operations

---
 src/codegen/spirv.zig | 39 ++++++++++++++++++++++++++++++++++++---
 1 file changed, 36 insertions(+), 3 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index 88d6f74a2d..14511e287f 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -84,6 +84,9 @@ pub const DeclGen = struct {
     const ArithmeticTypeInfo = struct {
         /// A classification of the inner type.
         const Class = enum {
+            /// A boolean.
+            bool,
+
             /// A regular, **native**, integer.
             /// This is only returned when the backend supports this int as a native type (when
             /// the relevant capability is enabled).
@@ -196,6 +199,12 @@ pub const DeclGen = struct {
         const target = self.module.getTarget();
 
         return switch (ty.zigTypeTag()) {
+            .Bool => ArithmeticTypeInfo{
+                .bits = 1, // Doesn't matter for this class.
+                .is_vector = false,
+                .signedness = .unsigned, // Technically, but doesn't matter for this class.
+                .class = .bool,
+            },
             .Float => ArithmeticTypeInfo{
                 .bits = ty.floatBits(target),
                 .is_vector = false,
@@ -436,6 +445,12 @@ pub const DeclGen = struct {
             .bit_and => try self.genBinOp(inst.castTag(.bit_and).?),
             .bit_or => try self.genBinOp(inst.castTag(.bit_or).?),
             .xor => try self.genBinOp(inst.castTag(.xor).?),
+            .cmp_eq => try self.genBinOp(inst.castTag(.cmp_eq).?),
+            .cmp_neq => try self.genBinOp(inst.castTag(.cmp_neq).?),
+            .cmp_gt => try self.genBinOp(inst.castTag(.cmp_gt).?),
+            .cmp_gte => try self.genBinOp(inst.castTag(.cmp_gte).?),
+            .cmp_lt => try self.genBinOp(inst.castTag(.cmp_lt).?),
+            .cmp_lte => try self.genBinOp(inst.castTag(.cmp_lte).?),
             .arg => self.genArg(),
             // TODO: Breakpoints won't be supported in SPIR-V, but the compiler seems to insert them
             // throughout the IR.
@@ -458,18 +473,23 @@ pub const DeclGen = struct {
 
         // TODO: Is the result the same as the argument types?
         // This is supposed to be the case for SPIR-V.
-        std.debug.assert(inst.base.ty.eql(inst.lhs.ty) and inst.base.ty.eql(inst.rhs.ty));
+        std.debug.assert(inst.rhs.ty.eql(inst.lhs.ty));
+        std.debug.assert(inst.base.ty.tag() == .bool or inst.base.ty.eql(inst.lhs.ty));
 
         // Binary operations are generally applicable to both scalar and vector operations in SPIR-V, but int and float
         // versions of operations require different opcodes.
-        const info = try self.arithmeticTypeInfo(inst.base.ty);
+        // For operations which produce bools, the information of inst.base.ty is not useful, so just pick either operand
+        // instead.
+        const info = try self.arithmeticTypeInfo(inst.lhs.ty);
 
         if (info.class == .composite_integer)
             return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: binary operations for composite integers", .{});
 
+        const is_bool = info.class == .bool;
         const is_float = info.class == .float;
         const is_signed = info.signedness == .signed;
-        // **Note**: All these operations must be valid for vectors of floats and integers as well!
+        // **Note**: All these operations must be valid for vectors of floats, integers and bools as well!
+        // For floating points, we generally want ordered operations (which return false if either operand is nan).
         const opcode = switch (inst.base.tag) {
             // The regular integer operations are all defined for wrapping. Since theyre only relevant for integers,
             // we can just switch on both cases here.
@@ -478,11 +498,24 @@ pub const DeclGen = struct {
             .mul, .mulwrap => if (is_float) Opcode.OpFMul else Opcode.OpIMul,
             // TODO: Trap if divisor is 0?
             // TODO: Figure out of OpSDiv for unsigned/OpUDiv for signed does anything useful.
+            //  => Those are probably for divTrunc and divFloor, though the compiler does not yet generate those.
+            //  => TODO: Figure out how those work on the SPIR-V side.
+            //  => TODO: Test these.
             .div => if (is_float) Opcode.OpFDiv else if (is_signed) Opcode.OpSDiv else Opcode.OpUDiv,
             // Only integer versions for these.
             .bit_and => Opcode.OpBitwiseAnd,
             .bit_or => Opcode.OpBitwiseOr,
             .xor => Opcode.OpBitwiseXor,
+            // Int/bool/float -> bool operations.
+            .cmp_eq => if (is_float) Opcode.OpFOrdEqual else if (is_bool) Opcode.OpLogicalEqual else Opcode.OpIEqual,
+            .cmp_neq => if (is_float) Opcode.OpFOrdNotEqual else if (is_bool) Opcode.OpLogicalNotEqual else Opcode.OpINotEqual,
+            // Int/float -> bool operations.
+            // TODO: Verify that these OpFOrd type operations produce the right value.
+            // TODO: Is there a more fundamental difference between OpU and OpS operations here than just the type?
+            .cmp_gt => if (is_float) Opcode.OpFOrdGreaterThan else if (is_signed) Opcode.OpSGreaterThan else Opcode.OpUGreaterThan,
+            .cmp_gte => if (is_float) Opcode.OpFOrdGreaterThanEqual else if (is_signed) Opcode.OpSGreaterThanEqual else Opcode.OpUGreaterThanEqual,
+            .cmp_lt => if (is_float) Opcode.OpFOrdLessThan else if (is_signed) Opcode.OpSLessThan else Opcode.OpULessThan,
+            .cmp_lte => if (is_float) Opcode.OpFOrdLessThanEqual else if (is_signed) Opcode.OpSLessThanEqual else Opcode.OpULessThanEqual,
             else => unreachable,
         };
 

From 489b3ef7d47c877aa7e761ddf00763bfe1dc03a7 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sun, 16 May 2021 14:52:11 +0200
Subject: [PATCH 13/21] SPIR-V: bool binary operations

---
 src/codegen/spirv.zig | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index 14511e287f..949e9fc6a2 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -451,6 +451,8 @@ pub const DeclGen = struct {
             .cmp_gte => try self.genBinOp(inst.castTag(.cmp_gte).?),
             .cmp_lt => try self.genBinOp(inst.castTag(.cmp_lt).?),
             .cmp_lte => try self.genBinOp(inst.castTag(.cmp_lte).?),
+            .bool_and => try self.genBinOp(inst.castTag(.bool_and).?),
+            .bool_or => try self.genBinOp(inst.castTag(.bool_or).?),
             .arg => self.genArg(),
             // TODO: Breakpoints won't be supported in SPIR-V, but the compiler seems to insert them
             // throughout the IR.
@@ -468,7 +470,7 @@ pub const DeclGen = struct {
         const lhs_id = try self.resolve(inst.lhs);
         const rhs_id = try self.resolve(inst.rhs);
 
-        const binop_result_id = self.spv.allocResultId();
+        const result_id = self.spv.allocResultId();
         const result_type_id = try self.getOrGenType(inst.base.ty);
 
         // TODO: Is the result the same as the argument types?
@@ -516,16 +518,19 @@ pub const DeclGen = struct {
             .cmp_gte => if (is_float) Opcode.OpFOrdGreaterThanEqual else if (is_signed) Opcode.OpSGreaterThanEqual else Opcode.OpUGreaterThanEqual,
             .cmp_lt => if (is_float) Opcode.OpFOrdLessThan else if (is_signed) Opcode.OpSLessThan else Opcode.OpULessThan,
             .cmp_lte => if (is_float) Opcode.OpFOrdLessThanEqual else if (is_signed) Opcode.OpSLessThanEqual else Opcode.OpULessThanEqual,
+            // Bool -> bool operations.
+            .bool_and => Opcode.OpLogicalAnd,
+            .bool_or => Opcode.OpLogicalOr,
             else => unreachable,
         };
 
-        try writeInstruction(&self.spv.fn_decls, opcode, &[_]u32{ result_type_id, binop_result_id, lhs_id, rhs_id });
+        try writeInstruction(&self.spv.fn_decls, opcode, &[_]u32{ result_type_id, result_id, lhs_id, rhs_id });
 
         // TODO: Trap on overflow? Probably going to be annoying.
         // TODO: Look into NoSignedWrap/NoUnsignedWrap extensions.
 
         if (info.class != .strange_integer)
-            return binop_result_id;
+            return result_id;
 
         return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: strange integer operation mask", .{});
     }

From 880473dc3f08e2f8c0cef85777d50e25e4bcb062 Mon Sep 17 00:00:00 2001
From: Robin Voetter <robin@voetter.nl>
Date: Sun, 16 May 2021 14:55:09 +0200
Subject: [PATCH 14/21] SPIR-V: Unary not operation

---
 src/codegen/spirv.zig | 22 +++++++++++++++++++++-
 1 file changed, 21 insertions(+), 1 deletion(-)

diff --git a/src/codegen/spirv.zig b/src/codegen/spirv.zig
index 949e9fc6a2..7992a7a465 100644
--- a/src/codegen/spirv.zig
+++ b/src/codegen/spirv.zig
@@ -453,6 +453,7 @@ pub const DeclGen = struct {
             .cmp_lte => try self.genBinOp(inst.castTag(.cmp_lte).?),
             .bool_and => try self.genBinOp(inst.castTag(.bool_and).?),
             .bool_or => try self.genBinOp(inst.castTag(.bool_or).?),
+            .not => try self.genUnOp(inst.castTag(.not).?),
             .arg => self.genArg(),
             // TODO: Breakpoints won't be supported in SPIR-V, but the compiler seems to insert them
             // throughout the IR.
@@ -527,7 +528,7 @@ pub const DeclGen = struct {
         try writeInstruction(&self.spv.fn_decls, opcode, &[_]u32{ result_type_id, result_id, lhs_id, rhs_id });
 
         // TODO: Trap on overflow? Probably going to be annoying.
-        // TODO: Look into NoSignedWrap/NoUnsignedWrap extensions.
+        // TODO: Look into SPV_KHR_no_integer_wrap_decoration which provides NoSignedWrap/NoUnsignedWrap.
 
         if (info.class != .strange_integer)
             return result_id;
@@ -535,6 +536,25 @@ pub const DeclGen = struct {
         return self.fail(.{.node_offset = 0}, "TODO: SPIR-V backend: strange integer operation mask", .{});
     }
 
+    fn genUnOp(self: *DeclGen, inst: *Inst.UnOp) !u32 {
+        const operand_id = try self.resolve(inst.operand);
+
+        const result_id = self.spv.allocResultId();
+        const result_type_id = try self.getOrGenType(inst.base.ty);
+
+        const info = try self.arithmeticTypeInfo(inst.operand.ty);
+
+        const opcode = switch (inst.base.tag) {
+            // Bool -> bool
+            .not => Opcode.OpLogicalNot,
+            else => unreachable,
+        };
+
+        try writeInstruction(&self.spv.fn_decls, opcode, &[_]u32{ result_type_id, result_id, operand_id });
+
+        return result_id;
+    }
+
     fn genArg(self: *DeclGen) u32 {
         defer self.next_arg_index += 1;
         return self.args.items[self.next_arg_index];

From 5042a476822af59ef005df2e7ff9f994907cceb4 Mon Sep 17 00:00:00 2001
From: Jonathan Marler <johnnymarler@gmail.com>
Date: Sun, 16 May 2021 10:37:29 -0600
Subject: [PATCH 15/21] add missing EBADF error code for openat

---
 lib/std/os.zig | 1 +
 1 file changed, 1 insertion(+)

diff --git a/lib/std/os.zig b/lib/std/os.zig
index 0ee20d7c8e..5df2c441cd 100644
--- a/lib/std/os.zig
+++ b/lib/std/os.zig
@@ -1244,6 +1244,7 @@ pub fn openatZ(dir_fd: fd_t, file_path: [*:0]const u8, flags: u32, mode: mode_t)
 
             EFAULT => unreachable,
             EINVAL => unreachable,
+            EBADF => unreachable,
             EACCES => return error.AccessDenied,
             EFBIG => return error.FileTooBig,
             EOVERFLOW => return error.FileTooBig,

From 8467373bb84fc7b0690c0e4d024d2cb298931cd9 Mon Sep 17 00:00:00 2001
From: Aiz672 <aizawey672@gmail.com>
Date: Sat, 15 May 2021 14:58:53 +0700
Subject: [PATCH 16/21] Remove `isIntegerNumber` and `isFloatingNumber`

---
 lib/std/meta/trait.zig | 14 --------------
 1 file changed, 14 deletions(-)

diff --git a/lib/std/meta/trait.zig b/lib/std/meta/trait.zig
index 2bf2e6a0a5..b9bb6c332d 100644
--- a/lib/std/meta/trait.zig
+++ b/lib/std/meta/trait.zig
@@ -298,20 +298,6 @@ pub fn isNumber(comptime T: type) bool {
     };
 }
 
-pub fn isIntegerNumber(comptime T: type) bool {
-    return switch (@typeInfo(T)) {
-        .Int, .ComptimeInt => true,
-        else => false,
-    };
-}
-
-pub fn isFloatingNumber(comptime T: type) bool {
-    return switch (@typeInfo(T)) {
-        .Float, .ComptimeFloat => true,
-        else => false,
-    };
-}
-
 test "std.meta.trait.isNumber" {
     const NotANumber = struct {
         number: u8,

From 04d95ea4192a7f70e7c11b8ee67d237cf38da9b7 Mon Sep 17 00:00:00 2001
From: Michael Dusan <michael.dusan@gmail.com>
Date: Mon, 17 May 2021 09:10:47 -0400
Subject: [PATCH 17/21] ci linux: bump qemu-5.2.0.1

- apply patch for qemu-user syscall do_sendrecvmsg_locked
- see #8750
---
 ci/azure/linux_script | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ci/azure/linux_script b/ci/azure/linux_script
index 64407b79b7..2ca7ed9f75 100755
--- a/ci/azure/linux_script
+++ b/ci/azure/linux_script
@@ -20,7 +20,7 @@ cd $HOME
 wget -nv "https://ziglang.org/deps/$CACHE_BASENAME.tar.xz"
 tar xf "$CACHE_BASENAME.tar.xz"
 
-QEMUBASE="qemu-linux-x86_64-5.2.0"
+QEMUBASE="qemu-linux-x86_64-5.2.0.1"
 wget -nv "https://ziglang.org/deps/$QEMUBASE.tar.xz"
 tar xf "$QEMUBASE.tar.xz"
 export PATH="$(pwd)/$QEMUBASE/bin:$PATH"

From 5414bd48edd460ae8667c811e13aa9b5d9fab919 Mon Sep 17 00:00:00 2001
From: aiz <78422626+aizawey672@users.noreply.github.com>
Date: Tue, 18 May 2021 02:57:51 +0700
Subject: [PATCH 18/21] std.math.Complex: Change `new()` to `init()`

---
 lib/std/math/complex.zig       | 27 ++++++++--------
 lib/std/math/complex/abs.zig   |  2 +-
 lib/std/math/complex/acos.zig  |  4 +--
 lib/std/math/complex/acosh.zig |  4 +--
 lib/std/math/complex/arg.zig   |  2 +-
 lib/std/math/complex/asin.zig  |  8 ++---
 lib/std/math/complex/asinh.zig |  6 ++--
 lib/std/math/complex/atan.zig  | 20 ++++++------
 lib/std/math/complex/atanh.zig |  6 ++--
 lib/std/math/complex/conj.zig  |  4 +--
 lib/std/math/complex/cos.zig   |  4 +--
 lib/std/math/complex/cosh.zig  | 56 +++++++++++++++++-----------------
 lib/std/math/complex/exp.zig   | 28 ++++++++---------
 lib/std/math/complex/ldexp.zig |  4 +--
 lib/std/math/complex/log.zig   |  4 +--
 lib/std/math/complex/pow.zig   |  4 +--
 lib/std/math/complex/proj.zig  |  6 ++--
 lib/std/math/complex/sin.zig   |  6 ++--
 lib/std/math/complex/sinh.zig  | 56 +++++++++++++++++-----------------
 lib/std/math/complex/sqrt.zig  | 32 +++++++++----------
 lib/std/math/complex/tan.zig   |  6 ++--
 lib/std/math/complex/tanh.zig  | 24 +++++++--------
 lib/std/zig/parser_test.zig    | 12 ++++----
 23 files changed, 164 insertions(+), 161 deletions(-)

diff --git a/lib/std/math/complex.zig b/lib/std/math/complex.zig
index abac923cdd..614d9d0ae1 100644
--- a/lib/std/math/complex.zig
+++ b/lib/std/math/complex.zig
@@ -38,9 +38,12 @@ pub fn Complex(comptime T: type) type {
 
         /// Imaginary part.
         im: T,
+        
+        /// Deprecated, use init()
+        pub const new = init;
 
         /// Create a new Complex number from the given real and imaginary parts.
-        pub fn new(re: T, im: T) Self {
+        pub fn init(re: T, im: T) Self {
             return Self{
                 .re = re,
                 .im = im,
@@ -110,32 +113,32 @@ pub fn Complex(comptime T: type) type {
 const epsilon = 0.0001;
 
 test "complex.add" {
-    const a = Complex(f32).new(5, 3);
-    const b = Complex(f32).new(2, 7);
+    const a = Complex(f32).init(5, 3);
+    const b = Complex(f32).init(2, 7);
     const c = a.add(b);
 
     try testing.expect(c.re == 7 and c.im == 10);
 }
 
 test "complex.sub" {
-    const a = Complex(f32).new(5, 3);
-    const b = Complex(f32).new(2, 7);
+    const a = Complex(f32).init(5, 3);
+    const b = Complex(f32).init(2, 7);
     const c = a.sub(b);
 
     try testing.expect(c.re == 3 and c.im == -4);
 }
 
 test "complex.mul" {
-    const a = Complex(f32).new(5, 3);
-    const b = Complex(f32).new(2, 7);
+    const a = Complex(f32).init(5, 3);
+    const b = Complex(f32).init(2, 7);
     const c = a.mul(b);
 
     try testing.expect(c.re == -11 and c.im == 41);
 }
 
 test "complex.div" {
-    const a = Complex(f32).new(5, 3);
-    const b = Complex(f32).new(2, 7);
+    const a = Complex(f32).init(5, 3);
+    const b = Complex(f32).init(2, 7);
     const c = a.div(b);
 
     try testing.expect(math.approxEqAbs(f32, c.re, @as(f32, 31) / 53, epsilon) and
@@ -143,14 +146,14 @@ test "complex.div" {
 }
 
 test "complex.conjugate" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = a.conjugate();
 
     try testing.expect(c.re == 5 and c.im == -3);
 }
 
 test "complex.reciprocal" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = a.reciprocal();
 
     try testing.expect(math.approxEqAbs(f32, c.re, @as(f32, 5) / 34, epsilon) and
@@ -158,7 +161,7 @@ test "complex.reciprocal" {
 }
 
 test "complex.magnitude" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = a.magnitude();
 
     try testing.expect(math.approxEqAbs(f32, c, 5.83095, epsilon));
diff --git a/lib/std/math/complex/abs.zig b/lib/std/math/complex/abs.zig
index a2678d21db..6890d15f8a 100644
--- a/lib/std/math/complex/abs.zig
+++ b/lib/std/math/complex/abs.zig
@@ -18,7 +18,7 @@ pub fn abs(z: anytype) @TypeOf(z.re) {
 const epsilon = 0.0001;
 
 test "complex.cabs" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = abs(a);
     try testing.expect(math.approxEqAbs(f32, c, 5.83095, epsilon));
 }
diff --git a/lib/std/math/complex/acos.zig b/lib/std/math/complex/acos.zig
index 72abea47fe..3d02ad6358 100644
--- a/lib/std/math/complex/acos.zig
+++ b/lib/std/math/complex/acos.zig
@@ -13,13 +13,13 @@ const Complex = cmath.Complex;
 pub fn acos(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
     const q = cmath.asin(z);
-    return Complex(T).new(@as(T, math.pi) / 2 - q.re, -q.im);
+    return Complex(T).init(@as(T, math.pi) / 2 - q.re, -q.im);
 }
 
 const epsilon = 0.0001;
 
 test "complex.cacos" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = acos(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 0.546975, epsilon));
diff --git a/lib/std/math/complex/acosh.zig b/lib/std/math/complex/acosh.zig
index 4f76dea01a..c239936d47 100644
--- a/lib/std/math/complex/acosh.zig
+++ b/lib/std/math/complex/acosh.zig
@@ -13,13 +13,13 @@ const Complex = cmath.Complex;
 pub fn acosh(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
     const q = cmath.acos(z);
-    return Complex(T).new(-q.im, q.re);
+    return Complex(T).init(-q.im, q.re);
 }
 
 const epsilon = 0.0001;
 
 test "complex.cacosh" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = acosh(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 2.452914, epsilon));
diff --git a/lib/std/math/complex/arg.zig b/lib/std/math/complex/arg.zig
index c583b9e360..8a79daa073 100644
--- a/lib/std/math/complex/arg.zig
+++ b/lib/std/math/complex/arg.zig
@@ -18,7 +18,7 @@ pub fn arg(z: anytype) @TypeOf(z.re) {
 const epsilon = 0.0001;
 
 test "complex.carg" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = arg(a);
     try testing.expect(math.approxEqAbs(f32, c, 0.540420, epsilon));
 }
diff --git a/lib/std/math/complex/asin.zig b/lib/std/math/complex/asin.zig
index 7ce200fae2..f2c377df56 100644
--- a/lib/std/math/complex/asin.zig
+++ b/lib/std/math/complex/asin.zig
@@ -15,17 +15,17 @@ pub fn asin(z: anytype) Complex(@TypeOf(z.re)) {
     const x = z.re;
     const y = z.im;
 
-    const p = Complex(T).new(1.0 - (x - y) * (x + y), -2.0 * x * y);
-    const q = Complex(T).new(-y, x);
+    const p = Complex(T).init(1.0 - (x - y) * (x + y), -2.0 * x * y);
+    const q = Complex(T).init(-y, x);
     const r = cmath.log(q.add(cmath.sqrt(p)));
 
-    return Complex(T).new(r.im, -r.re);
+    return Complex(T).init(r.im, -r.re);
 }
 
 const epsilon = 0.0001;
 
 test "complex.casin" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = asin(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 1.023822, epsilon));
diff --git a/lib/std/math/complex/asinh.zig b/lib/std/math/complex/asinh.zig
index 821218acf7..b1083c81bf 100644
--- a/lib/std/math/complex/asinh.zig
+++ b/lib/std/math/complex/asinh.zig
@@ -12,15 +12,15 @@ const Complex = cmath.Complex;
 /// Returns the hyperbolic arc-sine of z.
 pub fn asinh(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
-    const q = Complex(T).new(-z.im, z.re);
+    const q = Complex(T).init(-z.im, z.re);
     const r = cmath.asin(q);
-    return Complex(T).new(r.im, -r.re);
+    return Complex(T).init(r.im, -r.re);
 }
 
 const epsilon = 0.0001;
 
 test "complex.casinh" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = asinh(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 2.459831, epsilon));
diff --git a/lib/std/math/complex/atan.zig b/lib/std/math/complex/atan.zig
index d9a95c8dc1..b26b1e7ef3 100644
--- a/lib/std/math/complex/atan.zig
+++ b/lib/std/math/complex/atan.zig
@@ -50,14 +50,14 @@ fn atan32(z: Complex(f32)) Complex(f32) {
 
     if ((x == 0.0) and (y > 1.0)) {
         // overflow
-        return Complex(f32).new(maxnum, maxnum);
+        return Complex(f32).init(maxnum, maxnum);
     }
 
     const x2 = x * x;
     var a = 1.0 - x2 - (y * y);
     if (a == 0.0) {
         // overflow
-        return Complex(f32).new(maxnum, maxnum);
+        return Complex(f32).init(maxnum, maxnum);
     }
 
     var t = 0.5 * math.atan2(f32, 2.0 * x, a);
@@ -67,12 +67,12 @@ fn atan32(z: Complex(f32)) Complex(f32) {
     a = x2 + t * t;
     if (a == 0.0) {
         // overflow
-        return Complex(f32).new(maxnum, maxnum);
+        return Complex(f32).init(maxnum, maxnum);
     }
 
     t = y + 1.0;
     a = (x2 + (t * t)) / a;
-    return Complex(f32).new(w, 0.25 * math.ln(a));
+    return Complex(f32).init(w, 0.25 * math.ln(a));
 }
 
 fn redupif64(x: f64) f64 {
@@ -99,14 +99,14 @@ fn atan64(z: Complex(f64)) Complex(f64) {
 
     if ((x == 0.0) and (y > 1.0)) {
         // overflow
-        return Complex(f64).new(maxnum, maxnum);
+        return Complex(f64).init(maxnum, maxnum);
     }
 
     const x2 = x * x;
     var a = 1.0 - x2 - (y * y);
     if (a == 0.0) {
         // overflow
-        return Complex(f64).new(maxnum, maxnum);
+        return Complex(f64).init(maxnum, maxnum);
     }
 
     var t = 0.5 * math.atan2(f64, 2.0 * x, a);
@@ -116,18 +116,18 @@ fn atan64(z: Complex(f64)) Complex(f64) {
     a = x2 + t * t;
     if (a == 0.0) {
         // overflow
-        return Complex(f64).new(maxnum, maxnum);
+        return Complex(f64).init(maxnum, maxnum);
     }
 
     t = y + 1.0;
     a = (x2 + (t * t)) / a;
-    return Complex(f64).new(w, 0.25 * math.ln(a));
+    return Complex(f64).init(w, 0.25 * math.ln(a));
 }
 
 const epsilon = 0.0001;
 
 test "complex.catan32" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = atan(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 1.423679, epsilon));
@@ -135,7 +135,7 @@ test "complex.catan32" {
 }
 
 test "complex.catan64" {
-    const a = Complex(f64).new(5, 3);
+    const a = Complex(f64).init(5, 3);
     const c = atan(a);
 
     try testing.expect(math.approxEqAbs(f64, c.re, 1.423679, epsilon));
diff --git a/lib/std/math/complex/atanh.zig b/lib/std/math/complex/atanh.zig
index 420f401f17..41232570b5 100644
--- a/lib/std/math/complex/atanh.zig
+++ b/lib/std/math/complex/atanh.zig
@@ -12,15 +12,15 @@ const Complex = cmath.Complex;
 /// Returns the hyperbolic arc-tangent of z.
 pub fn atanh(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
-    const q = Complex(T).new(-z.im, z.re);
+    const q = Complex(T).init(-z.im, z.re);
     const r = cmath.atan(q);
-    return Complex(T).new(r.im, -r.re);
+    return Complex(T).init(r.im, -r.re);
 }
 
 const epsilon = 0.0001;
 
 test "complex.catanh" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = atanh(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 0.146947, epsilon));
diff --git a/lib/std/math/complex/conj.zig b/lib/std/math/complex/conj.zig
index 960295830a..ea4ba7356c 100644
--- a/lib/std/math/complex/conj.zig
+++ b/lib/std/math/complex/conj.zig
@@ -12,11 +12,11 @@ const Complex = cmath.Complex;
 /// Returns the complex conjugate of z.
 pub fn conj(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
-    return Complex(T).new(z.re, -z.im);
+    return Complex(T).init(z.re, -z.im);
 }
 
 test "complex.conj" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = a.conjugate();
 
     try testing.expect(c.re == 5 and c.im == -3);
diff --git a/lib/std/math/complex/cos.zig b/lib/std/math/complex/cos.zig
index 2de3735d12..0760485bb0 100644
--- a/lib/std/math/complex/cos.zig
+++ b/lib/std/math/complex/cos.zig
@@ -12,14 +12,14 @@ const Complex = cmath.Complex;
 /// Returns the cosine of z.
 pub fn cos(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
-    const p = Complex(T).new(-z.im, z.re);
+    const p = Complex(T).init(-z.im, z.re);
     return cmath.cosh(p);
 }
 
 const epsilon = 0.0001;
 
 test "complex.ccos" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = cos(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 2.855815, epsilon));
diff --git a/lib/std/math/complex/cosh.zig b/lib/std/math/complex/cosh.zig
index 2514e72bdb..7f82fb98d4 100644
--- a/lib/std/math/complex/cosh.zig
+++ b/lib/std/math/complex/cosh.zig
@@ -40,55 +40,55 @@ fn cosh32(z: Complex(f32)) Complex(f32) {
 
     if (ix < 0x7f800000 and iy < 0x7f800000) {
         if (iy == 0) {
-            return Complex(f32).new(math.cosh(x), y);
+            return Complex(f32).init(math.cosh(x), y);
         }
         // small x: normal case
         if (ix < 0x41100000) {
-            return Complex(f32).new(math.cosh(x) * math.cos(y), math.sinh(x) * math.sin(y));
+            return Complex(f32).init(math.cosh(x) * math.cos(y), math.sinh(x) * math.sin(y));
         }
 
         // |x|>= 9, so cosh(x) ~= exp(|x|)
         if (ix < 0x42b17218) {
             // x < 88.7: exp(|x|) won't overflow
             const h = math.exp(math.fabs(x)) * 0.5;
-            return Complex(f32).new(math.copysign(f32, h, x) * math.cos(y), h * math.sin(y));
+            return Complex(f32).init(math.copysign(f32, h, x) * math.cos(y), h * math.sin(y));
         }
         // x < 192.7: scale to avoid overflow
         else if (ix < 0x4340b1e7) {
-            const v = Complex(f32).new(math.fabs(x), y);
+            const v = Complex(f32).init(math.fabs(x), y);
             const r = ldexp_cexp(v, -1);
-            return Complex(f32).new(r.re, r.im * math.copysign(f32, 1, x));
+            return Complex(f32).init(r.re, r.im * math.copysign(f32, 1, x));
         }
         // x >= 192.7: result always overflows
         else {
             const h = 0x1p127 * x;
-            return Complex(f32).new(h * h * math.cos(y), h * math.sin(y));
+            return Complex(f32).init(h * h * math.cos(y), h * math.sin(y));
         }
     }
 
     if (ix == 0 and iy >= 0x7f800000) {
-        return Complex(f32).new(y - y, math.copysign(f32, 0, x * (y - y)));
+        return Complex(f32).init(y - y, math.copysign(f32, 0, x * (y - y)));
     }
 
     if (iy == 0 and ix >= 0x7f800000) {
         if (hx & 0x7fffff == 0) {
-            return Complex(f32).new(x * x, math.copysign(f32, 0, x) * y);
+            return Complex(f32).init(x * x, math.copysign(f32, 0, x) * y);
         }
-        return Complex(f32).new(x, math.copysign(f32, 0, (x + x) * y));
+        return Complex(f32).init(x, math.copysign(f32, 0, (x + x) * y));
     }
 
     if (ix < 0x7f800000 and iy >= 0x7f800000) {
-        return Complex(f32).new(y - y, x * (y - y));
+        return Complex(f32).init(y - y, x * (y - y));
     }
 
     if (ix >= 0x7f800000 and (hx & 0x7fffff) == 0) {
         if (iy >= 0x7f800000) {
-            return Complex(f32).new(x * x, x * (y - y));
+            return Complex(f32).init(x * x, x * (y - y));
         }
-        return Complex(f32).new((x * x) * math.cos(y), x * math.sin(y));
+        return Complex(f32).init((x * x) * math.cos(y), x * math.sin(y));
     }
 
-    return Complex(f32).new((x * x) * (y - y), (x + x) * (y - y));
+    return Complex(f32).init((x * x) * (y - y), (x + x) * (y - y));
 }
 
 fn cosh64(z: Complex(f64)) Complex(f64) {
@@ -108,61 +108,61 @@ fn cosh64(z: Complex(f64)) Complex(f64) {
     // nearly non-exceptional case where x, y are finite
     if (ix < 0x7ff00000 and iy < 0x7ff00000) {
         if (iy | ly == 0) {
-            return Complex(f64).new(math.cosh(x), x * y);
+            return Complex(f64).init(math.cosh(x), x * y);
         }
         // small x: normal case
         if (ix < 0x40360000) {
-            return Complex(f64).new(math.cosh(x) * math.cos(y), math.sinh(x) * math.sin(y));
+            return Complex(f64).init(math.cosh(x) * math.cos(y), math.sinh(x) * math.sin(y));
         }
 
         // |x|>= 22, so cosh(x) ~= exp(|x|)
         if (ix < 0x40862e42) {
             // x < 710: exp(|x|) won't overflow
             const h = math.exp(math.fabs(x)) * 0.5;
-            return Complex(f64).new(h * math.cos(y), math.copysign(f64, h, x) * math.sin(y));
+            return Complex(f64).init(h * math.cos(y), math.copysign(f64, h, x) * math.sin(y));
         }
         // x < 1455: scale to avoid overflow
         else if (ix < 0x4096bbaa) {
-            const v = Complex(f64).new(math.fabs(x), y);
+            const v = Complex(f64).init(math.fabs(x), y);
             const r = ldexp_cexp(v, -1);
-            return Complex(f64).new(r.re, r.im * math.copysign(f64, 1, x));
+            return Complex(f64).init(r.re, r.im * math.copysign(f64, 1, x));
         }
         // x >= 1455: result always overflows
         else {
             const h = 0x1p1023;
-            return Complex(f64).new(h * h * math.cos(y), h * math.sin(y));
+            return Complex(f64).init(h * h * math.cos(y), h * math.sin(y));
         }
     }
 
     if (ix | lx == 0 and iy >= 0x7ff00000) {
-        return Complex(f64).new(y - y, math.copysign(f64, 0, x * (y - y)));
+        return Complex(f64).init(y - y, math.copysign(f64, 0, x * (y - y)));
     }
 
     if (iy | ly == 0 and ix >= 0x7ff00000) {
         if ((hx & 0xfffff) | lx == 0) {
-            return Complex(f64).new(x * x, math.copysign(f64, 0, x) * y);
+            return Complex(f64).init(x * x, math.copysign(f64, 0, x) * y);
         }
-        return Complex(f64).new(x * x, math.copysign(f64, 0, (x + x) * y));
+        return Complex(f64).init(x * x, math.copysign(f64, 0, (x + x) * y));
     }
 
     if (ix < 0x7ff00000 and iy >= 0x7ff00000) {
-        return Complex(f64).new(y - y, x * (y - y));
+        return Complex(f64).init(y - y, x * (y - y));
     }
 
     if (ix >= 0x7ff00000 and (hx & 0xfffff) | lx == 0) {
         if (iy >= 0x7ff00000) {
-            return Complex(f64).new(x * x, x * (y - y));
+            return Complex(f64).init(x * x, x * (y - y));
         }
-        return Complex(f64).new(x * x * math.cos(y), x * math.sin(y));
+        return Complex(f64).init(x * x * math.cos(y), x * math.sin(y));
     }
 
-    return Complex(f64).new((x * x) * (y - y), (x + x) * (y - y));
+    return Complex(f64).init((x * x) * (y - y), (x + x) * (y - y));
 }
 
 const epsilon = 0.0001;
 
 test "complex.ccosh32" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = cosh(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, -73.467300, epsilon));
@@ -170,7 +170,7 @@ test "complex.ccosh32" {
 }
 
 test "complex.ccosh64" {
-    const a = Complex(f64).new(5, 3);
+    const a = Complex(f64).init(5, 3);
     const c = cosh(a);
 
     try testing.expect(math.approxEqAbs(f64, c.re, -73.467300, epsilon));
diff --git a/lib/std/math/complex/exp.zig b/lib/std/math/complex/exp.zig
index a7eb3d51b5..fa74300dea 100644
--- a/lib/std/math/complex/exp.zig
+++ b/lib/std/math/complex/exp.zig
@@ -39,25 +39,25 @@ fn exp32(z: Complex(f32)) Complex(f32) {
     const hy = @bitCast(u32, y) & 0x7fffffff;
     // cexp(x + i0) = exp(x) + i0
     if (hy == 0) {
-        return Complex(f32).new(math.exp(x), y);
+        return Complex(f32).init(math.exp(x), y);
     }
 
     const hx = @bitCast(u32, x);
     // cexp(0 + iy) = cos(y) + isin(y)
     if ((hx & 0x7fffffff) == 0) {
-        return Complex(f32).new(math.cos(y), math.sin(y));
+        return Complex(f32).init(math.cos(y), math.sin(y));
     }
 
     if (hy >= 0x7f800000) {
         // cexp(finite|nan +- i inf|nan) = nan + i nan
         if ((hx & 0x7fffffff) != 0x7f800000) {
-            return Complex(f32).new(y - y, y - y);
+            return Complex(f32).init(y - y, y - y);
         } // cexp(-inf +- i inf|nan) = 0 + i0
         else if (hx & 0x80000000 != 0) {
-            return Complex(f32).new(0, 0);
+            return Complex(f32).init(0, 0);
         } // cexp(+inf +- i inf|nan) = inf + i nan
         else {
-            return Complex(f32).new(x, y - y);
+            return Complex(f32).init(x, y - y);
         }
     }
 
@@ -70,7 +70,7 @@ fn exp32(z: Complex(f32)) Complex(f32) {
     // - x = nan
     else {
         const exp_x = math.exp(x);
-        return Complex(f32).new(exp_x * math.cos(y), exp_x * math.sin(y));
+        return Complex(f32).init(exp_x * math.cos(y), exp_x * math.sin(y));
     }
 }
 
@@ -87,7 +87,7 @@ fn exp64(z: Complex(f64)) Complex(f64) {
 
     // cexp(x + i0) = exp(x) + i0
     if (hy | ly == 0) {
-        return Complex(f64).new(math.exp(x), y);
+        return Complex(f64).init(math.exp(x), y);
     }
 
     const fx = @bitCast(u64, x);
@@ -96,19 +96,19 @@ fn exp64(z: Complex(f64)) Complex(f64) {
 
     // cexp(0 + iy) = cos(y) + isin(y)
     if ((hx & 0x7fffffff) | lx == 0) {
-        return Complex(f64).new(math.cos(y), math.sin(y));
+        return Complex(f64).init(math.cos(y), math.sin(y));
     }
 
     if (hy >= 0x7ff00000) {
         // cexp(finite|nan +- i inf|nan) = nan + i nan
         if (lx != 0 or (hx & 0x7fffffff) != 0x7ff00000) {
-            return Complex(f64).new(y - y, y - y);
+            return Complex(f64).init(y - y, y - y);
         } // cexp(-inf +- i inf|nan) = 0 + i0
         else if (hx & 0x80000000 != 0) {
-            return Complex(f64).new(0, 0);
+            return Complex(f64).init(0, 0);
         } // cexp(+inf +- i inf|nan) = inf + i nan
         else {
-            return Complex(f64).new(x, y - y);
+            return Complex(f64).init(x, y - y);
         }
     }
 
@@ -121,14 +121,14 @@ fn exp64(z: Complex(f64)) Complex(f64) {
     // - x = nan
     else {
         const exp_x = math.exp(x);
-        return Complex(f64).new(exp_x * math.cos(y), exp_x * math.sin(y));
+        return Complex(f64).init(exp_x * math.cos(y), exp_x * math.sin(y));
     }
 }
 
 const epsilon = 0.0001;
 
 test "complex.cexp32" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = exp(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, -146.927917, epsilon));
@@ -136,7 +136,7 @@ test "complex.cexp32" {
 }
 
 test "complex.cexp64" {
-    const a = Complex(f64).new(5, 3);
+    const a = Complex(f64).init(5, 3);
     const c = exp(a);
 
     try testing.expect(math.approxEqAbs(f64, c.re, -146.927917, epsilon));
diff --git a/lib/std/math/complex/ldexp.zig b/lib/std/math/complex/ldexp.zig
index 3ae0382fe3..9f31aae549 100644
--- a/lib/std/math/complex/ldexp.zig
+++ b/lib/std/math/complex/ldexp.zig
@@ -48,7 +48,7 @@ fn ldexp_cexp32(z: Complex(f32), expt: i32) Complex(f32) {
     const half_expt2 = exptf - half_expt1;
     const scale2 = @bitCast(f32, (0x7f + half_expt2) << 23);
 
-    return Complex(f32).new(math.cos(z.im) * exp_x * scale1 * scale2, math.sin(z.im) * exp_x * scale1 * scale2);
+    return Complex(f32).init(math.cos(z.im) * exp_x * scale1 * scale2, math.sin(z.im) * exp_x * scale1 * scale2);
 }
 
 fn frexp_exp64(x: f64, expt: *i32) f64 {
@@ -78,7 +78,7 @@ fn ldexp_cexp64(z: Complex(f64), expt: i32) Complex(f64) {
     const half_expt2 = exptf - half_expt1;
     const scale2 = @bitCast(f64, (0x3ff + half_expt2) << 20);
 
-    return Complex(f64).new(
+    return Complex(f64).init(
         math.cos(z.im) * exp_x * scale1 * scale2,
         math.sin(z.im) * exp_x * scale1 * scale2,
     );
diff --git a/lib/std/math/complex/log.zig b/lib/std/math/complex/log.zig
index 7f8f649953..e59870d556 100644
--- a/lib/std/math/complex/log.zig
+++ b/lib/std/math/complex/log.zig
@@ -15,13 +15,13 @@ pub fn log(z: anytype) Complex(@TypeOf(z.re)) {
     const r = cmath.abs(z);
     const phi = cmath.arg(z);
 
-    return Complex(T).new(math.ln(r), phi);
+    return Complex(T).init(math.ln(r), phi);
 }
 
 const epsilon = 0.0001;
 
 test "complex.clog" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = log(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 1.763180, epsilon));
diff --git a/lib/std/math/complex/pow.zig b/lib/std/math/complex/pow.zig
index 0d3a45e6d2..092c8c2422 100644
--- a/lib/std/math/complex/pow.zig
+++ b/lib/std/math/complex/pow.zig
@@ -19,8 +19,8 @@ pub fn pow(comptime T: type, z: T, c: T) T {
 const epsilon = 0.0001;
 
 test "complex.cpow" {
-    const a = Complex(f32).new(5, 3);
-    const b = Complex(f32).new(2.3, -1.3);
+    const a = Complex(f32).init(5, 3);
+    const b = Complex(f32).init(2.3, -1.3);
     const c = pow(Complex(f32), a, b);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 58.049110, epsilon));
diff --git a/lib/std/math/complex/proj.zig b/lib/std/math/complex/proj.zig
index 260816481b..8527be2293 100644
--- a/lib/std/math/complex/proj.zig
+++ b/lib/std/math/complex/proj.zig
@@ -14,16 +14,16 @@ pub fn proj(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
 
     if (math.isInf(z.re) or math.isInf(z.im)) {
-        return Complex(T).new(math.inf(T), math.copysign(T, 0, z.re));
+        return Complex(T).init(math.inf(T), math.copysign(T, 0, z.re));
     }
 
-    return Complex(T).new(z.re, z.im);
+    return Complex(T).init(z.re, z.im);
 }
 
 const epsilon = 0.0001;
 
 test "complex.cproj" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = proj(a);
 
     try testing.expect(c.re == 5 and c.im == 3);
diff --git a/lib/std/math/complex/sin.zig b/lib/std/math/complex/sin.zig
index 68551b8596..39b5f584ac 100644
--- a/lib/std/math/complex/sin.zig
+++ b/lib/std/math/complex/sin.zig
@@ -12,15 +12,15 @@ const Complex = cmath.Complex;
 /// Returns the sine of z.
 pub fn sin(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
-    const p = Complex(T).new(-z.im, z.re);
+    const p = Complex(T).init(-z.im, z.re);
     const q = cmath.sinh(p);
-    return Complex(T).new(q.im, -q.re);
+    return Complex(T).init(q.im, -q.re);
 }
 
 const epsilon = 0.0001;
 
 test "complex.csin" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = sin(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, -9.654126, epsilon));
diff --git a/lib/std/math/complex/sinh.zig b/lib/std/math/complex/sinh.zig
index ea09f8e17d..b673da3d2f 100644
--- a/lib/std/math/complex/sinh.zig
+++ b/lib/std/math/complex/sinh.zig
@@ -40,55 +40,55 @@ fn sinh32(z: Complex(f32)) Complex(f32) {
 
     if (ix < 0x7f800000 and iy < 0x7f800000) {
         if (iy == 0) {
-            return Complex(f32).new(math.sinh(x), y);
+            return Complex(f32).init(math.sinh(x), y);
         }
         // small x: normal case
         if (ix < 0x41100000) {
-            return Complex(f32).new(math.sinh(x) * math.cos(y), math.cosh(x) * math.sin(y));
+            return Complex(f32).init(math.sinh(x) * math.cos(y), math.cosh(x) * math.sin(y));
         }
 
         // |x|>= 9, so cosh(x) ~= exp(|x|)
         if (ix < 0x42b17218) {
             // x < 88.7: exp(|x|) won't overflow
             const h = math.exp(math.fabs(x)) * 0.5;
-            return Complex(f32).new(math.copysign(f32, h, x) * math.cos(y), h * math.sin(y));
+            return Complex(f32).init(math.copysign(f32, h, x) * math.cos(y), h * math.sin(y));
         }
         // x < 192.7: scale to avoid overflow
         else if (ix < 0x4340b1e7) {
-            const v = Complex(f32).new(math.fabs(x), y);
+            const v = Complex(f32).init(math.fabs(x), y);
             const r = ldexp_cexp(v, -1);
-            return Complex(f32).new(r.re * math.copysign(f32, 1, x), r.im);
+            return Complex(f32).init(r.re * math.copysign(f32, 1, x), r.im);
         }
         // x >= 192.7: result always overflows
         else {
             const h = 0x1p127 * x;
-            return Complex(f32).new(h * math.cos(y), h * h * math.sin(y));
+            return Complex(f32).init(h * math.cos(y), h * h * math.sin(y));
         }
     }
 
     if (ix == 0 and iy >= 0x7f800000) {
-        return Complex(f32).new(math.copysign(f32, 0, x * (y - y)), y - y);
+        return Complex(f32).init(math.copysign(f32, 0, x * (y - y)), y - y);
     }
 
     if (iy == 0 and ix >= 0x7f800000) {
         if (hx & 0x7fffff == 0) {
-            return Complex(f32).new(x, y);
+            return Complex(f32).init(x, y);
         }
-        return Complex(f32).new(x, math.copysign(f32, 0, y));
+        return Complex(f32).init(x, math.copysign(f32, 0, y));
     }
 
     if (ix < 0x7f800000 and iy >= 0x7f800000) {
-        return Complex(f32).new(y - y, x * (y - y));
+        return Complex(f32).init(y - y, x * (y - y));
     }
 
     if (ix >= 0x7f800000 and (hx & 0x7fffff) == 0) {
         if (iy >= 0x7f800000) {
-            return Complex(f32).new(x * x, x * (y - y));
+            return Complex(f32).init(x * x, x * (y - y));
         }
-        return Complex(f32).new(x * math.cos(y), math.inf_f32 * math.sin(y));
+        return Complex(f32).init(x * math.cos(y), math.inf_f32 * math.sin(y));
     }
 
-    return Complex(f32).new((x * x) * (y - y), (x + x) * (y - y));
+    return Complex(f32).init((x * x) * (y - y), (x + x) * (y - y));
 }
 
 fn sinh64(z: Complex(f64)) Complex(f64) {
@@ -107,61 +107,61 @@ fn sinh64(z: Complex(f64)) Complex(f64) {
 
     if (ix < 0x7ff00000 and iy < 0x7ff00000) {
         if (iy | ly == 0) {
-            return Complex(f64).new(math.sinh(x), y);
+            return Complex(f64).init(math.sinh(x), y);
         }
         // small x: normal case
         if (ix < 0x40360000) {
-            return Complex(f64).new(math.sinh(x) * math.cos(y), math.cosh(x) * math.sin(y));
+            return Complex(f64).init(math.sinh(x) * math.cos(y), math.cosh(x) * math.sin(y));
         }
 
         // |x|>= 22, so cosh(x) ~= exp(|x|)
         if (ix < 0x40862e42) {
             // x < 710: exp(|x|) won't overflow
             const h = math.exp(math.fabs(x)) * 0.5;
-            return Complex(f64).new(math.copysign(f64, h, x) * math.cos(y), h * math.sin(y));
+            return Complex(f64).init(math.copysign(f64, h, x) * math.cos(y), h * math.sin(y));
         }
         // x < 1455: scale to avoid overflow
         else if (ix < 0x4096bbaa) {
-            const v = Complex(f64).new(math.fabs(x), y);
+            const v = Complex(f64).init(math.fabs(x), y);
             const r = ldexp_cexp(v, -1);
-            return Complex(f64).new(r.re * math.copysign(f64, 1, x), r.im);
+            return Complex(f64).init(r.re * math.copysign(f64, 1, x), r.im);
         }
         // x >= 1455: result always overflows
         else {
             const h = 0x1p1023 * x;
-            return Complex(f64).new(h * math.cos(y), h * h * math.sin(y));
+            return Complex(f64).init(h * math.cos(y), h * h * math.sin(y));
         }
     }
 
     if (ix | lx == 0 and iy >= 0x7ff00000) {
-        return Complex(f64).new(math.copysign(f64, 0, x * (y - y)), y - y);
+        return Complex(f64).init(math.copysign(f64, 0, x * (y - y)), y - y);
     }
 
     if (iy | ly == 0 and ix >= 0x7ff00000) {
         if ((hx & 0xfffff) | lx == 0) {
-            return Complex(f64).new(x, y);
+            return Complex(f64).init(x, y);
         }
-        return Complex(f64).new(x, math.copysign(f64, 0, y));
+        return Complex(f64).init(x, math.copysign(f64, 0, y));
     }
 
     if (ix < 0x7ff00000 and iy >= 0x7ff00000) {
-        return Complex(f64).new(y - y, x * (y - y));
+        return Complex(f64).init(y - y, x * (y - y));
     }
 
     if (ix >= 0x7ff00000 and (hx & 0xfffff) | lx == 0) {
         if (iy >= 0x7ff00000) {
-            return Complex(f64).new(x * x, x * (y - y));
+            return Complex(f64).init(x * x, x * (y - y));
         }
-        return Complex(f64).new(x * math.cos(y), math.inf_f64 * math.sin(y));
+        return Complex(f64).init(x * math.cos(y), math.inf_f64 * math.sin(y));
     }
 
-    return Complex(f64).new((x * x) * (y - y), (x + x) * (y - y));
+    return Complex(f64).init((x * x) * (y - y), (x + x) * (y - y));
 }
 
 const epsilon = 0.0001;
 
 test "complex.csinh32" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = sinh(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, -73.460617, epsilon));
@@ -169,7 +169,7 @@ test "complex.csinh32" {
 }
 
 test "complex.csinh64" {
-    const a = Complex(f64).new(5, 3);
+    const a = Complex(f64).init(5, 3);
     const c = sinh(a);
 
     try testing.expect(math.approxEqAbs(f64, c.re, -73.460617, epsilon));
diff --git a/lib/std/math/complex/sqrt.zig b/lib/std/math/complex/sqrt.zig
index 25c486f838..07fed152fa 100644
--- a/lib/std/math/complex/sqrt.zig
+++ b/lib/std/math/complex/sqrt.zig
@@ -32,15 +32,15 @@ fn sqrt32(z: Complex(f32)) Complex(f32) {
     const y = z.im;
 
     if (x == 0 and y == 0) {
-        return Complex(f32).new(0, y);
+        return Complex(f32).init(0, y);
     }
     if (math.isInf(y)) {
-        return Complex(f32).new(math.inf(f32), y);
+        return Complex(f32).init(math.inf(f32), y);
     }
     if (math.isNan(x)) {
         // raise invalid if y is not nan
         const t = (y - y) / (y - y);
-        return Complex(f32).new(x, t);
+        return Complex(f32).init(x, t);
     }
     if (math.isInf(x)) {
         // sqrt(inf + i nan)    = inf + nan i
@@ -48,9 +48,9 @@ fn sqrt32(z: Complex(f32)) Complex(f32) {
         // sqrt(-inf + i nan)   = nan +- inf i
         // sqrt(-inf + iy)      = 0 + inf i
         if (math.signbit(x)) {
-            return Complex(f32).new(math.fabs(x - y), math.copysign(f32, x, y));
+            return Complex(f32).init(math.fabs(x - y), math.copysign(f32, x, y));
         } else {
-            return Complex(f32).new(x, math.copysign(f32, y - y, y));
+            return Complex(f32).init(x, math.copysign(f32, y - y, y));
         }
     }
 
@@ -62,13 +62,13 @@ fn sqrt32(z: Complex(f32)) Complex(f32) {
 
     if (dx >= 0) {
         const t = math.sqrt((dx + math.hypot(f64, dx, dy)) * 0.5);
-        return Complex(f32).new(
+        return Complex(f32).init(
             @floatCast(f32, t),
             @floatCast(f32, dy / (2.0 * t)),
         );
     } else {
         const t = math.sqrt((-dx + math.hypot(f64, dx, dy)) * 0.5);
-        return Complex(f32).new(
+        return Complex(f32).init(
             @floatCast(f32, math.fabs(y) / (2.0 * t)),
             @floatCast(f32, math.copysign(f64, t, y)),
         );
@@ -83,15 +83,15 @@ fn sqrt64(z: Complex(f64)) Complex(f64) {
     var y = z.im;
 
     if (x == 0 and y == 0) {
-        return Complex(f64).new(0, y);
+        return Complex(f64).init(0, y);
     }
     if (math.isInf(y)) {
-        return Complex(f64).new(math.inf(f64), y);
+        return Complex(f64).init(math.inf(f64), y);
     }
     if (math.isNan(x)) {
         // raise invalid if y is not nan
         const t = (y - y) / (y - y);
-        return Complex(f64).new(x, t);
+        return Complex(f64).init(x, t);
     }
     if (math.isInf(x)) {
         // sqrt(inf + i nan)    = inf + nan i
@@ -99,9 +99,9 @@ fn sqrt64(z: Complex(f64)) Complex(f64) {
         // sqrt(-inf + i nan)   = nan +- inf i
         // sqrt(-inf + iy)      = 0 + inf i
         if (math.signbit(x)) {
-            return Complex(f64).new(math.fabs(x - y), math.copysign(f64, x, y));
+            return Complex(f64).init(math.fabs(x - y), math.copysign(f64, x, y));
         } else {
-            return Complex(f64).new(x, math.copysign(f64, y - y, y));
+            return Complex(f64).init(x, math.copysign(f64, y - y, y));
         }
     }
 
@@ -118,10 +118,10 @@ fn sqrt64(z: Complex(f64)) Complex(f64) {
     var result: Complex(f64) = undefined;
     if (x >= 0) {
         const t = math.sqrt((x + math.hypot(f64, x, y)) * 0.5);
-        result = Complex(f64).new(t, y / (2.0 * t));
+        result = Complex(f64).init(t, y / (2.0 * t));
     } else {
         const t = math.sqrt((-x + math.hypot(f64, x, y)) * 0.5);
-        result = Complex(f64).new(math.fabs(y) / (2.0 * t), math.copysign(f64, t, y));
+        result = Complex(f64).init(math.fabs(y) / (2.0 * t), math.copysign(f64, t, y));
     }
 
     if (scale) {
@@ -135,7 +135,7 @@ fn sqrt64(z: Complex(f64)) Complex(f64) {
 const epsilon = 0.0001;
 
 test "complex.csqrt32" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = sqrt(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 2.327117, epsilon));
@@ -143,7 +143,7 @@ test "complex.csqrt32" {
 }
 
 test "complex.csqrt64" {
-    const a = Complex(f64).new(5, 3);
+    const a = Complex(f64).init(5, 3);
     const c = sqrt(a);
 
     try testing.expect(math.approxEqAbs(f64, c.re, 2.3271175190399496, epsilon));
diff --git a/lib/std/math/complex/tan.zig b/lib/std/math/complex/tan.zig
index ca9d4ce7e9..0ee34dfcc2 100644
--- a/lib/std/math/complex/tan.zig
+++ b/lib/std/math/complex/tan.zig
@@ -12,15 +12,15 @@ const Complex = cmath.Complex;
 /// Returns the tanget of z.
 pub fn tan(z: anytype) Complex(@TypeOf(z.re)) {
     const T = @TypeOf(z.re);
-    const q = Complex(T).new(-z.im, z.re);
+    const q = Complex(T).init(-z.im, z.re);
     const r = cmath.tanh(q);
-    return Complex(T).new(r.im, -r.re);
+    return Complex(T).init(r.im, -r.re);
 }
 
 const epsilon = 0.0001;
 
 test "complex.ctan" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = tan(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, -0.002708233, epsilon));
diff --git a/lib/std/math/complex/tanh.zig b/lib/std/math/complex/tanh.zig
index 096bdaac06..cb5e7b3f6d 100644
--- a/lib/std/math/complex/tanh.zig
+++ b/lib/std/math/complex/tanh.zig
@@ -36,22 +36,22 @@ fn tanh32(z: Complex(f32)) Complex(f32) {
     if (ix >= 0x7f800000) {
         if (ix & 0x7fffff != 0) {
             const r = if (y == 0) y else x * y;
-            return Complex(f32).new(x, r);
+            return Complex(f32).init(x, r);
         }
         const xx = @bitCast(f32, hx - 0x40000000);
         const r = if (math.isInf(y)) y else math.sin(y) * math.cos(y);
-        return Complex(f32).new(xx, math.copysign(f32, 0, r));
+        return Complex(f32).init(xx, math.copysign(f32, 0, r));
     }
 
     if (!math.isFinite(y)) {
         const r = if (ix != 0) y - y else x;
-        return Complex(f32).new(r, y - y);
+        return Complex(f32).init(r, y - y);
     }
 
     // x >= 11
     if (ix >= 0x41300000) {
         const exp_mx = math.exp(-math.fabs(x));
-        return Complex(f32).new(math.copysign(f32, 1, x), 4 * math.sin(y) * math.cos(y) * exp_mx * exp_mx);
+        return Complex(f32).init(math.copysign(f32, 1, x), 4 * math.sin(y) * math.cos(y) * exp_mx * exp_mx);
     }
 
     // Kahan's algorithm
@@ -61,7 +61,7 @@ fn tanh32(z: Complex(f32)) Complex(f32) {
     const rho = math.sqrt(1 + s * s);
     const den = 1 + beta * s * s;
 
-    return Complex(f32).new((beta * rho * s) / den, t / den);
+    return Complex(f32).init((beta * rho * s) / den, t / den);
 }
 
 fn tanh64(z: Complex(f64)) Complex(f64) {
@@ -78,23 +78,23 @@ fn tanh64(z: Complex(f64)) Complex(f64) {
     if (ix >= 0x7ff00000) {
         if ((ix & 0x7fffff) | lx != 0) {
             const r = if (y == 0) y else x * y;
-            return Complex(f64).new(x, r);
+            return Complex(f64).init(x, r);
         }
 
         const xx = @bitCast(f64, (@as(u64, hx - 0x40000000) << 32) | lx);
         const r = if (math.isInf(y)) y else math.sin(y) * math.cos(y);
-        return Complex(f64).new(xx, math.copysign(f64, 0, r));
+        return Complex(f64).init(xx, math.copysign(f64, 0, r));
     }
 
     if (!math.isFinite(y)) {
         const r = if (ix != 0) y - y else x;
-        return Complex(f64).new(r, y - y);
+        return Complex(f64).init(r, y - y);
     }
 
     // x >= 22
     if (ix >= 0x40360000) {
         const exp_mx = math.exp(-math.fabs(x));
-        return Complex(f64).new(math.copysign(f64, 1, x), 4 * math.sin(y) * math.cos(y) * exp_mx * exp_mx);
+        return Complex(f64).init(math.copysign(f64, 1, x), 4 * math.sin(y) * math.cos(y) * exp_mx * exp_mx);
     }
 
     // Kahan's algorithm
@@ -104,13 +104,13 @@ fn tanh64(z: Complex(f64)) Complex(f64) {
     const rho = math.sqrt(1 + s * s);
     const den = 1 + beta * s * s;
 
-    return Complex(f64).new((beta * rho * s) / den, t / den);
+    return Complex(f64).init((beta * rho * s) / den, t / den);
 }
 
 const epsilon = 0.0001;
 
 test "complex.ctanh32" {
-    const a = Complex(f32).new(5, 3);
+    const a = Complex(f32).init(5, 3);
     const c = tanh(a);
 
     try testing.expect(math.approxEqAbs(f32, c.re, 0.999913, epsilon));
@@ -118,7 +118,7 @@ test "complex.ctanh32" {
 }
 
 test "complex.ctanh64" {
-    const a = Complex(f64).new(5, 3);
+    const a = Complex(f64).init(5, 3);
     const c = tanh(a);
 
     try testing.expect(math.approxEqAbs(f64, c.re, 0.999913, epsilon));
diff --git a/lib/std/zig/parser_test.zig b/lib/std/zig/parser_test.zig
index 03586bc777..b2ce8df73f 100644
--- a/lib/std/zig/parser_test.zig
+++ b/lib/std/zig/parser_test.zig
@@ -1608,13 +1608,13 @@ test "zig fmt: if-else with comment before else" {
         \\comptime {
         \\    // cexp(finite|nan +- i inf|nan) = nan + i nan
         \\    if ((hx & 0x7fffffff) != 0x7f800000) {
-        \\        return Complex(f32).new(y - y, y - y);
+        \\        return Complex(f32).init(y - y, y - y);
         \\    } // cexp(-inf +- i inf|nan) = 0 + i0
         \\    else if (hx & 0x80000000 != 0) {
-        \\        return Complex(f32).new(0, 0);
+        \\        return Complex(f32).init(0, 0);
         \\    } // cexp(+inf +- i inf|nan) = inf + i nan
         \\    else {
-        \\        return Complex(f32).new(x, y - y);
+        \\        return Complex(f32).init(x, y - y);
         \\    }
         \\}
         \\
@@ -2267,16 +2267,16 @@ test "zig fmt: line comment between if block and else keyword" {
         \\test "aoeu" {
         \\    // cexp(finite|nan +- i inf|nan) = nan + i nan
         \\    if ((hx & 0x7fffffff) != 0x7f800000) {
-        \\        return Complex(f32).new(y - y, y - y);
+        \\        return Complex(f32).init(y - y, y - y);
         \\    }
         \\    // cexp(-inf +- i inf|nan) = 0 + i0
         \\    else if (hx & 0x80000000 != 0) {
-        \\        return Complex(f32).new(0, 0);
+        \\        return Complex(f32).init(0, 0);
         \\    }
         \\    // cexp(+inf +- i inf|nan) = inf + i nan
         \\    // another comment
         \\    else {
-        \\        return Complex(f32).new(x, y - y);
+        \\        return Complex(f32).init(x, y - y);
         \\    }
         \\}
         \\

From fe1a166589db0f2371429c93e1e1e622c19378f1 Mon Sep 17 00:00:00 2001
From: LemonBoy <thatlemon@gmail.com>
Date: Mon, 17 May 2021 17:41:42 +0200
Subject: [PATCH 19/21] translate-c: Add `@truncate` where needed

Make getLimitedValue API much easier to use with zig.
Fixes the compilation on 32bit hosts.
---
 src/clang.zig       | 5 ++++-
 src/translate_c.zig | 4 ++--
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/src/clang.zig b/src/clang.zig
index 256dbda1e1..b9e152aef4 100644
--- a/src/clang.zig
+++ b/src/clang.zig
@@ -1,3 +1,4 @@
+const std = @import("std");
 pub const builtin = @import("builtin");
 
 pub const SourceLocation = extern struct {
@@ -115,7 +116,9 @@ pub const APFloatBaseSemantics = extern enum {
 };
 
 pub const APInt = opaque {
-    pub const getLimitedValue = ZigClangAPInt_getLimitedValue;
+    pub fn getLimitedValue(self: *const APInt, comptime T: type) T {
+        return @truncate(T, ZigClangAPInt_getLimitedValue(self, std.math.maxInt(T)));
+    }
     extern fn ZigClangAPInt_getLimitedValue(*const APInt, limit: u64) u64;
 };
 
diff --git a/src/translate_c.zig b/src/translate_c.zig
index 348e284db3..19aec279ec 100644
--- a/src/translate_c.zig
+++ b/src/translate_c.zig
@@ -2341,7 +2341,7 @@ fn transInitListExprArray(
     assert(@ptrCast(*const clang.Type, arr_type).isConstantArrayType());
     const const_arr_ty = @ptrCast(*const clang.ConstantArrayType, arr_type);
     const size_ap_int = const_arr_ty.getSize();
-    const all_count = size_ap_int.getLimitedValue(math.maxInt(usize));
+    const all_count = size_ap_int.getLimitedValue(usize);
     const leftover_count = all_count - init_count;
 
     if (all_count == 0) {
@@ -4266,7 +4266,7 @@ fn transType(c: *Context, scope: *Scope, ty: *const clang.Type, source_loc: clan
             const const_arr_ty = @ptrCast(*const clang.ConstantArrayType, ty);
 
             const size_ap_int = const_arr_ty.getSize();
-            const size = size_ap_int.getLimitedValue(math.maxInt(usize));
+            const size = size_ap_int.getLimitedValue(usize);
             const elem_type = try transType(c, scope, const_arr_ty.getElementType().getTypePtr(), source_loc);
 
             return Tag.array_type.create(c.arena, .{ .len = size, .elem_type = elem_type });

From b7eab32f42040ecb80501111adb05bc8369625e8 Mon Sep 17 00:00:00 2001
From: LemonBoy <thatlemon@gmail.com>
Date: Sun, 16 May 2021 11:51:39 +0200
Subject: [PATCH 20/21] std: Allocate tlscsprng memory as needed

Let mmap allocate a block of memory that's wide enough to use with
MADV_WIPEONFORK, madvise granularity is the current system page size
(using a static buffer of mem.page_size bytes would be wrong, that's the
minimum page size).

As a result, we don't zero some random chunk of memory every time we
fork the process.

Fixes #7609
---
 lib/std/crypto/tlcsprng.zig | 117 +++++++++++++++++++++++-------------
 1 file changed, 75 insertions(+), 42 deletions(-)

diff --git a/lib/std/crypto/tlcsprng.zig b/lib/std/crypto/tlcsprng.zig
index 115a7ab882..7c43125920 100644
--- a/lib/std/crypto/tlcsprng.zig
+++ b/lib/std/crypto/tlcsprng.zig
@@ -12,6 +12,7 @@
 const std = @import("std");
 const root = @import("root");
 const mem = std.mem;
+const os = std.os;
 
 /// We use this as a layer of indirection because global const pointers cannot
 /// point to thread-local variables.
@@ -42,16 +43,12 @@ const maybe_have_wipe_on_fork = std.Target.current.os.isAtLeast(.linux, .{
     .minor = 14,
 }) orelse true;
 
-const WipeMe = struct {
-    init_state: enum { uninitialized, initialized, failed },
+const Context = struct {
+    init_state: enum(u8) { uninitialized = 0, initialized, failed },
     gimli: std.crypto.core.Gimli,
 };
-const wipe_align = if (maybe_have_wipe_on_fork) mem.page_size else @alignOf(WipeMe);
 
-threadlocal var wipe_me: WipeMe align(wipe_align) = .{
-    .gimli = undefined,
-    .init_state = .uninitialized,
-};
+threadlocal var wipe_mem: []align(mem.page_size) u8 = &[_]u8{};
 
 fn tlsCsprngFill(_: *const std.rand.Random, buffer: []u8) void {
     if (std.builtin.link_libc and @hasDecl(std.c, "arc4random_buf")) {
@@ -64,35 +61,69 @@ fn tlsCsprngFill(_: *const std.rand.Random, buffer: []u8) void {
     if (comptime std.meta.globalOption("crypto_always_getrandom", bool) orelse false) {
         return fillWithOsEntropy(buffer);
     }
-    switch (wipe_me.init_state) {
+
+    if (wipe_mem.len == 0) {
+        // Not initialized yet.
+        if (want_fork_safety and maybe_have_wipe_on_fork) {
+            // Allocate a per-process page, madvise operates with page
+            // granularity.
+            wipe_mem = os.mmap(
+                null,
+                @sizeOf(Context),
+                os.PROT_READ | os.PROT_WRITE,
+                os.MAP_PRIVATE | os.MAP_ANONYMOUS,
+                -1,
+                0,
+            ) catch |err| {
+                // Could not allocate memory for the local state, fall back to
+                // the OS syscall.
+                return fillWithOsEntropy(buffer);
+            };
+            // The memory is already zero-initialized.
+        } else {
+            // Use a static thread-local buffer.
+            const S = struct {
+                threadlocal var buf: Context align(mem.page_size) = .{
+                    .init_state = .uninitialized,
+                    .gimli = undefined,
+                };
+            };
+            wipe_mem = mem.asBytes(&S.buf);
+        }
+    }
+    const ctx = @ptrCast(*Context, wipe_mem.ptr);
+
+    switch (ctx.init_state) {
         .uninitialized => {
-            if (want_fork_safety) {
-                if (maybe_have_wipe_on_fork) {
-                    if (std.os.madvise(
-                        @ptrCast([*]align(mem.page_size) u8, &wipe_me),
-                        @sizeOf(@TypeOf(wipe_me)),
-                        std.os.MADV_WIPEONFORK,
-                    )) |_| {
-                        return initAndFill(buffer);
-                    } else |_| if (std.Thread.use_pthreads) {
-                        return setupPthreadAtforkAndFill(buffer);
-                    } else {
-                        // Since we failed to set up fork safety, we fall back to always
-                        // calling getrandom every time.
-                        wipe_me.init_state = .failed;
-                        return fillWithOsEntropy(buffer);
-                    }
-                } else if (std.Thread.use_pthreads) {
-                    return setupPthreadAtforkAndFill(buffer);
-                } else {
-                    // We have no mechanism to provide fork safety, but we want fork safety,
-                    // so we fall back to calling getrandom every time.
-                    wipe_me.init_state = .failed;
-                    return fillWithOsEntropy(buffer);
-                }
-            } else {
+            if (!want_fork_safety) {
                 return initAndFill(buffer);
             }
+
+            if (maybe_have_wipe_on_fork) wof: {
+                // Qemu user-mode emulation ignores any valid/invalid madvise
+                // hint and returns success. Check if this is the case by
+                // passing bogus parameters, we expect EINVAL as result.
+                if (os.madvise(wipe_mem.ptr, 0, 0xffffffff)) |_| {
+                    break :wof;
+                } else |_| {}
+
+                os.madvise(
+                    wipe_mem.ptr,
+                    wipe_mem.len,
+                    os.MADV_WIPEONFORK,
+                ) catch |_| {
+                    return initAndFill(buffer);
+                };
+            }
+
+            if (std.Thread.use_pthreads) {
+                return setupPthreadAtforkAndFill(buffer);
+            }
+
+            // Since we failed to set up fork safety, we fall back to always
+            // calling getrandom every time.
+            ctx.init_state = .failed;
+            return fillWithOsEntropy(buffer);
         },
         .initialized => {
             return fillWithCsprng(buffer);
@@ -110,7 +141,8 @@ fn tlsCsprngFill(_: *const std.rand.Random, buffer: []u8) void {
 fn setupPthreadAtforkAndFill(buffer: []u8) void {
     const failed = std.c.pthread_atfork(null, null, childAtForkHandler) != 0;
     if (failed) {
-        wipe_me.init_state = .failed;
+        const ctx = @ptrCast(*Context, wipe_mem.ptr);
+        ctx.init_state = .failed;
         return fillWithOsEntropy(buffer);
     } else {
         return initAndFill(buffer);
@@ -118,21 +150,21 @@ fn setupPthreadAtforkAndFill(buffer: []u8) void {
 }
 
 fn childAtForkHandler() callconv(.C) void {
-    const wipe_slice = @ptrCast([*]u8, &wipe_me)[0..@sizeOf(@TypeOf(wipe_me))];
-    std.crypto.utils.secureZero(u8, wipe_slice);
+    std.crypto.utils.secureZero(u8, wipe_mem);
 }
 
 fn fillWithCsprng(buffer: []u8) void {
+    const ctx = @ptrCast(*Context, wipe_mem.ptr);
     if (buffer.len != 0) {
-        wipe_me.gimli.squeeze(buffer);
+        ctx.gimli.squeeze(buffer);
     } else {
-        wipe_me.gimli.permute();
+        ctx.gimli.permute();
     }
-    mem.set(u8, wipe_me.gimli.toSlice()[0..std.crypto.core.Gimli.RATE], 0);
+    mem.set(u8, ctx.gimli.toSlice()[0..std.crypto.core.Gimli.RATE], 0);
 }
 
 fn fillWithOsEntropy(buffer: []u8) void {
-    std.os.getrandom(buffer) catch @panic("getrandom() failed to provide entropy");
+    os.getrandom(buffer) catch @panic("getrandom() failed to provide entropy");
 }
 
 fn initAndFill(buffer: []u8) void {
@@ -147,11 +179,12 @@ fn initAndFill(buffer: []u8) void {
         fillWithOsEntropy(&seed);
     }
 
-    wipe_me.gimli = std.crypto.core.Gimli.init(seed);
+    const ctx = @ptrCast(*Context, wipe_mem.ptr);
+    ctx.gimli = std.crypto.core.Gimli.init(seed);
 
     // This is at the end so that accidental recursive dependencies result
     // in stack overflows instead of invalid random data.
-    wipe_me.init_state = .initialized;
+    ctx.init_state = .initialized;
 
     return fillWithCsprng(buffer);
 }

From 65cee0b3fd5f9b3f83b79cc8fd1b64d13f4dd0c4 Mon Sep 17 00:00:00 2001
From: joachimschmidt557 <joachim.schmidt557@outlook.com>
Date: Sun, 16 May 2021 12:54:51 +0800
Subject: [PATCH 21/21] stage2 ARM: correct spilling in genArmMul as well

---
 src/codegen.zig     | 69 ++++++++++++++++++++++++++++++++++-----------
 test/stage2/arm.zig | 49 ++++++++++++++++++++++++++++++++
 2 files changed, 101 insertions(+), 17 deletions(-)

diff --git a/src/codegen.zig b/src/codegen.zig
index 9d344bf1d0..379a0f9b7d 100644
--- a/src/codegen.zig
+++ b/src/codegen.zig
@@ -1563,28 +1563,63 @@ fn Function(comptime arch: std.Target.Cpu.Arch) type {
             const lhs = try self.resolveInst(op_lhs);
             const rhs = try self.resolveInst(op_rhs);
 
+            const lhs_is_register = lhs == .register;
+            const rhs_is_register = rhs == .register;
+            const reuse_lhs = lhs_is_register and self.reuseOperand(inst, 0, lhs);
+            const reuse_rhs = !reuse_lhs and rhs_is_register and self.reuseOperand(inst, 1, rhs);
+
             // Destination must be a register
             // LHS must be a register
             // RHS must be a register
             var dst_mcv: MCValue = undefined;
-            var lhs_mcv: MCValue = undefined;
-            var rhs_mcv: MCValue = undefined;
-            if (self.reuseOperand(inst, 0, lhs)) {
-                // LHS is the destination
-                lhs_mcv = if (lhs != .register) try self.copyToNewRegister(inst, lhs) else lhs;
-                rhs_mcv = if (rhs != .register) try self.copyToNewRegister(inst, rhs) else rhs;
-                dst_mcv = lhs_mcv;
-            } else if (self.reuseOperand(inst, 1, rhs)) {
-                // RHS is the destination
-                lhs_mcv = if (lhs != .register) try self.copyToNewRegister(inst, lhs) else lhs;
-                rhs_mcv = if (rhs != .register) try self.copyToNewRegister(inst, rhs) else rhs;
-                dst_mcv = rhs_mcv;
+            var lhs_mcv: MCValue = lhs;
+            var rhs_mcv: MCValue = rhs;
+
+            // Allocate registers for operands and/or destination
+            const branch = &self.branch_stack.items[self.branch_stack.items.len - 1];
+            if (reuse_lhs) {
+                // Allocate 0 or 1 registers
+                if (!rhs_is_register) {
+                    rhs_mcv = MCValue{ .register = try self.register_manager.allocReg(op_rhs, &.{lhs.register}) };
+                    branch.inst_table.putAssumeCapacity(op_rhs, rhs_mcv);
+                }
+                dst_mcv = lhs;
+            } else if (reuse_rhs) {
+                // Allocate 0 or 1 registers
+                if (!lhs_is_register) {
+                    lhs_mcv = MCValue{ .register = try self.register_manager.allocReg(op_lhs, &.{rhs.register}) };
+                    branch.inst_table.putAssumeCapacity(op_lhs, lhs_mcv);
+                }
+                dst_mcv = rhs;
             } else {
-                // TODO save 1 copy instruction by directly allocating the destination register
-                // LHS is the destination
-                lhs_mcv = try self.copyToNewRegister(inst, lhs);
-                rhs_mcv = if (rhs != .register) try self.copyToNewRegister(inst, rhs) else rhs;
-                dst_mcv = lhs_mcv;
+                // Allocate 1 or 2 registers
+                if (lhs_is_register and rhs_is_register) {
+                    dst_mcv = MCValue{ .register = try self.register_manager.allocReg(inst, &.{ lhs.register, rhs.register }) };
+                } else if (lhs_is_register) {
+                    // Move RHS to register
+                    dst_mcv = MCValue{ .register = try self.register_manager.allocReg(inst, &.{lhs.register}) };
+                    rhs_mcv = dst_mcv;
+                } else if (rhs_is_register) {
+                    // Move LHS to register
+                    dst_mcv = MCValue{ .register = try self.register_manager.allocReg(inst, &.{rhs.register}) };
+                    lhs_mcv = dst_mcv;
+                } else {
+                    // Move LHS and RHS to register
+                    const regs = try self.register_manager.allocRegs(2, .{ inst, op_rhs }, &.{});
+                    lhs_mcv = MCValue{ .register = regs[0] };
+                    rhs_mcv = MCValue{ .register = regs[1] };
+                    dst_mcv = lhs_mcv;
+
+                    branch.inst_table.putAssumeCapacity(op_rhs, rhs_mcv);
+                }
+            }
+
+            // Move the operands to the newly allocated registers
+            if (!lhs_is_register) {
+                try self.genSetReg(op_lhs.src, op_lhs.ty, lhs_mcv.register, lhs);
+            }
+            if (!rhs_is_register) {
+                try self.genSetReg(op_rhs.src, op_rhs.ty, rhs_mcv.register, rhs);
             }
 
             writeInt(u32, try self.code.addManyAsArray(4), Instruction.mul(.al, dst_mcv.register, lhs_mcv.register, rhs_mcv.register).toU32());
diff --git a/test/stage2/arm.zig b/test/stage2/arm.zig
index 31b3c06dcd..02033005d1 100644
--- a/test/stage2/arm.zig
+++ b/test/stage2/arm.zig
@@ -510,5 +510,54 @@ pub fn addCases(ctx: *TestContext) !void {
         ,
             "",
         );
+
+        case.addCompareOutput(
+            \\export fn _start() noreturn {
+            \\    assert(addMul(3, 4) == 357747496);
+            \\    exit();
+            \\}
+            \\
+            \\fn addMul(a: u32, b: u32) u32 {
+            \\    const x: u32 = blk: {
+            \\        const c = a + b; // 7
+            \\        const d = a + c; // 10
+            \\        const e = d + b; // 14
+            \\        const f = d + e; // 24
+            \\        const g = e + f; // 38
+            \\        const h = f + g; // 62
+            \\        const i = g + h; // 100
+            \\        const j = i + d; // 110
+            \\        const k = i + j; // 210
+            \\        const l = k + c; // 217
+            \\        const m = l * d; // 2170     
+            \\        const n = m + e; // 2184     
+            \\        const o = n * f; // 52416    
+            \\        const p = o + g; // 52454    
+            \\        const q = p * h; // 3252148  
+            \\        const r = q + i; // 3252248  
+            \\        const s = r * j; // 357747280
+            \\        const t = s + k; // 357747490
+            \\        break :blk t;
+            \\    };
+            \\    const y = x + a; // 357747493
+            \\    const z = y + a; // 357747496
+            \\    return z;
+            \\}
+            \\
+            \\fn assert(ok: bool) void {
+            \\    if (!ok) unreachable;
+            \\}
+            \\
+            \\fn exit() noreturn {
+            \\    asm volatile ("svc #0"
+            \\        :
+            \\        : [number] "{r7}" (1),
+            \\          [arg1] "{r0}" (0)
+            \\        : "memory"
+            \\    );
+            \\    unreachable;
+            \\}            
+            ,
+            "",);
     }
 }