Merge pull request #11235 from joachimschmidt557/stage2-riscv

stage2 RISCV64: remove MCValue.embedded_in_code
This commit is contained in:
Joachim Schmidt 2022-03-21 12:36:47 +01:00 committed by GitHub
commit b48d8cce52
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 262 additions and 71 deletions

View File

@ -111,11 +111,6 @@ const MCValue = union(enum) {
/// A pointer-sized integer that fits in a register.
/// If the type is a pointer, this is the pointer address in virtual address space.
immediate: u64,
/// The constant was emitted into the code, at this offset.
/// If the type is a pointer, it means the pointer address is embedded in the code.
embedded_in_code: usize,
/// The value is a pointer to a constant which was emitted into the code, at this offset.
ptr_embedded_in_code: usize,
/// The value is in a target-specific register.
register: Register,
/// The value is in memory at a hard-coded address.
@ -129,7 +124,7 @@ const MCValue = union(enum) {
fn isMemory(mcv: MCValue) bool {
return switch (mcv) {
.embedded_in_code, .memory, .stack_offset => true,
.memory, .stack_offset => true,
else => false,
};
}
@ -148,10 +143,8 @@ const MCValue = union(enum) {
.dead => unreachable,
.immediate,
.embedded_in_code,
.memory,
.ptr_stack_offset,
.ptr_embedded_in_code,
.undef,
=> false,
@ -416,14 +409,17 @@ fn gen(self: *Self) !void {
});
// exitlude jumps
if (self.exitlude_jump_relocs.items.len == 1) {
// There is only one relocation. Hence,
// this relocation must be at the end of
// the code. Therefore, we can just delete
// the space initially reserved for the
// jump
self.mir_instructions.len -= 1;
} else for (self.exitlude_jump_relocs.items) |jmp_reloc| {
if (self.exitlude_jump_relocs.items.len > 0 and
self.exitlude_jump_relocs.items[self.exitlude_jump_relocs.items.len - 1] == self.mir_instructions.len - 2)
{
// If the last Mir instruction (apart from the
// dbg_epilogue_begin) is the last exitlude jump
// relocation (which would just jump one instruction
// further), it can be safely removed
self.mir_instructions.orderedRemove(self.exitlude_jump_relocs.pop());
}
for (self.exitlude_jump_relocs.items) |jmp_reloc| {
_ = jmp_reloc;
return self.fail("TODO add branches in RISCV64", .{});
}
@ -496,10 +492,10 @@ fn genBody(self: *Self, body: []const Air.Inst.Index) InnerError!void {
switch (air_tags[inst]) {
// zig fmt: off
.add, .ptr_add => try self.airAdd(inst),
.add, .ptr_add => try self.airBinOp(inst),
.addwrap => try self.airAddWrap(inst),
.add_sat => try self.airAddSat(inst),
.sub, .ptr_sub => try self.airSub(inst),
.sub, .ptr_sub => try self.airBinOp(inst),
.subwrap => try self.airSubWrap(inst),
.sub_sat => try self.airSubSat(inst),
.mul => try self.airMul(inst),
@ -927,9 +923,182 @@ fn airSlice(self: *Self, inst: Air.Inst.Index) !void {
return self.finishAir(inst, result, .{ bin_op.lhs, bin_op.rhs, .none });
}
fn airAdd(self: *Self, inst: Air.Inst.Index) !void {
/// Don't call this function directly. Use binOp instead.
///
/// Calling this function signals an intention to generate a Mir
/// instruction of the form
///
/// op dest, lhs, rhs
///
/// Asserts that generating an instruction of that form is possible.
fn binOpRegister(
self: *Self,
tag: Air.Inst.Tag,
maybe_inst: ?Air.Inst.Index,
lhs: MCValue,
rhs: MCValue,
lhs_ty: Type,
rhs_ty: Type,
) !MCValue {
const lhs_is_register = lhs == .register;
const rhs_is_register = rhs == .register;
if (lhs_is_register) self.register_manager.freezeRegs(&.{lhs.register});
if (rhs_is_register) self.register_manager.freezeRegs(&.{rhs.register});
const branch = &self.branch_stack.items[self.branch_stack.items.len - 1];
const lhs_reg = if (lhs_is_register) lhs.register else blk: {
const track_inst: ?Air.Inst.Index = if (maybe_inst) |inst| inst: {
const bin_op = self.air.instructions.items(.data)[inst].bin_op;
break :inst Air.refToIndex(bin_op.lhs).?;
} else null;
const reg = try self.register_manager.allocReg(track_inst);
self.register_manager.freezeRegs(&.{reg});
if (track_inst) |inst| branch.inst_table.putAssumeCapacity(inst, .{ .register = reg });
break :blk reg;
};
defer self.register_manager.unfreezeRegs(&.{lhs_reg});
const rhs_reg = if (rhs_is_register) rhs.register else blk: {
const track_inst: ?Air.Inst.Index = if (maybe_inst) |inst| inst: {
const bin_op = self.air.instructions.items(.data)[inst].bin_op;
break :inst Air.refToIndex(bin_op.rhs).?;
} else null;
const reg = try self.register_manager.allocReg(track_inst);
self.register_manager.freezeRegs(&.{reg});
if (track_inst) |inst| branch.inst_table.putAssumeCapacity(inst, .{ .register = reg });
break :blk reg;
};
defer self.register_manager.unfreezeRegs(&.{rhs_reg});
const dest_reg = if (maybe_inst) |inst| blk: {
const bin_op = self.air.instructions.items(.data)[inst].bin_op;
if (lhs_is_register and self.reuseOperand(inst, bin_op.lhs, 0, lhs)) {
break :blk lhs_reg;
} else if (rhs_is_register and self.reuseOperand(inst, bin_op.rhs, 1, rhs)) {
break :blk rhs_reg;
} else {
break :blk try self.register_manager.allocReg(inst);
}
} else try self.register_manager.allocReg(null);
if (!lhs_is_register) try self.genSetReg(lhs_ty, lhs_reg, lhs);
if (!rhs_is_register) try self.genSetReg(rhs_ty, rhs_reg, rhs);
const mir_tag: Mir.Inst.Tag = switch (tag) {
.add => .add,
.sub => .sub,
else => unreachable,
};
const mir_data: Mir.Inst.Data = switch (tag) {
.add,
.sub,
=> .{ .r_type = .{
.rd = dest_reg,
.rs1 = lhs_reg,
.rs2 = rhs_reg,
} },
else => unreachable,
};
_ = try self.addInst(.{
.tag = mir_tag,
.data = mir_data,
});
return MCValue{ .register = dest_reg };
}
/// For all your binary operation needs, this function will generate
/// the corresponding Mir instruction(s). Returns the location of the
/// result.
///
/// If the binary operation itself happens to be an Air instruction,
/// pass the corresponding index in the inst parameter. That helps
/// this function do stuff like reusing operands.
///
/// This function does not do any lowering to Mir itself, but instead
/// looks at the lhs and rhs and determines which kind of lowering
/// would be best suitable and then delegates the lowering to other
/// functions.
fn binOp(
self: *Self,
tag: Air.Inst.Tag,
maybe_inst: ?Air.Inst.Index,
lhs: MCValue,
rhs: MCValue,
lhs_ty: Type,
rhs_ty: Type,
) InnerError!MCValue {
switch (tag) {
// Arithmetic operations on integers and floats
.add,
.sub,
=> {
switch (lhs_ty.zigTypeTag()) {
.Float => return self.fail("TODO binary operations on floats", .{}),
.Vector => return self.fail("TODO binary operations on vectors", .{}),
.Int => {
assert(lhs_ty.eql(rhs_ty));
const int_info = lhs_ty.intInfo(self.target.*);
if (int_info.bits <= 64) {
// TODO immediate operands
return try self.binOpRegister(tag, maybe_inst, lhs, rhs, lhs_ty, rhs_ty);
} else {
return self.fail("TODO binary operations on int with bits > 64", .{});
}
},
else => unreachable,
}
},
.ptr_add,
.ptr_sub,
=> {
switch (lhs_ty.zigTypeTag()) {
.Pointer => {
const ptr_ty = lhs_ty;
const elem_ty = switch (ptr_ty.ptrSize()) {
.One => ptr_ty.childType().childType(), // ptr to array, so get array element type
else => ptr_ty.childType(),
};
const elem_size = elem_ty.abiSize(self.target.*);
if (elem_size == 1) {
const base_tag: Air.Inst.Tag = switch (tag) {
.ptr_add => .add,
.ptr_sub => .sub,
else => unreachable,
};
return try self.binOpRegister(base_tag, maybe_inst, lhs, rhs, lhs_ty, rhs_ty);
} else {
return self.fail("TODO ptr_add with elem_size > 1", .{});
}
},
else => unreachable,
}
},
else => unreachable,
}
}
fn airBinOp(self: *Self, inst: Air.Inst.Index) !void {
const tag = self.air.instructions.items(.tag)[inst];
const bin_op = self.air.instructions.items(.data)[inst].bin_op;
const result: MCValue = if (self.liveness.isUnused(inst)) .dead else return self.fail("TODO implement add for {}", .{self.target.cpu.arch});
const lhs = try self.resolveInst(bin_op.lhs);
const rhs = try self.resolveInst(bin_op.rhs);
const lhs_ty = self.air.typeOf(bin_op.lhs);
const rhs_ty = self.air.typeOf(bin_op.rhs);
const result: MCValue = if (self.liveness.isUnused(inst)) .dead else try self.binOp(tag, inst, lhs, rhs, lhs_ty, rhs_ty);
return self.finishAir(inst, result, .{ bin_op.lhs, bin_op.rhs, .none });
}
@ -945,12 +1114,6 @@ fn airAddSat(self: *Self, inst: Air.Inst.Index) !void {
return self.finishAir(inst, result, .{ bin_op.lhs, bin_op.rhs, .none });
}
fn airSub(self: *Self, inst: Air.Inst.Index) !void {
const bin_op = self.air.instructions.items(.data)[inst].bin_op;
const result: MCValue = if (self.liveness.isUnused(inst)) .dead else return self.fail("TODO implement sub for {}", .{self.target.cpu.arch});
return self.finishAir(inst, result, .{ bin_op.lhs, bin_op.rhs, .none });
}
fn airSubWrap(self: *Self, inst: Air.Inst.Index) !void {
const bin_op = self.air.instructions.items(.data)[inst].bin_op;
const result: MCValue = if (self.liveness.isUnused(inst)) .dead else return self.fail("TODO implement subwrap for {}", .{self.target.cpu.arch});
@ -1283,12 +1446,6 @@ fn load(self: *Self, dst_mcv: MCValue, ptr: MCValue, ptr_ty: Type) InnerError!vo
.dead => unreachable,
.immediate => |imm| try self.setRegOrMem(elem_ty, dst_mcv, .{ .memory = imm }),
.ptr_stack_offset => |off| try self.setRegOrMem(elem_ty, dst_mcv, .{ .stack_offset = off }),
.ptr_embedded_in_code => |off| {
try self.setRegOrMem(elem_ty, dst_mcv, .{ .embedded_in_code = off });
},
.embedded_in_code => {
return self.fail("TODO implement loading from MCValue.embedded_in_code", .{});
},
.register => {
return self.fail("TODO implement loading from MCValue.register", .{});
},
@ -1331,27 +1488,19 @@ fn airLoad(self: *Self, inst: Air.Inst.Index) !void {
return self.finishAir(inst, result, .{ ty_op.operand, .none, .none });
}
fn airStore(self: *Self, inst: Air.Inst.Index) !void {
const bin_op = self.air.instructions.items(.data)[inst].bin_op;
const ptr = try self.resolveInst(bin_op.lhs);
const value = try self.resolveInst(bin_op.rhs);
const elem_ty = self.air.typeOf(bin_op.rhs);
fn store(self: *Self, ptr: MCValue, value: MCValue, ptr_ty: Type, value_ty: Type) !void {
_ = ptr_ty;
switch (ptr) {
.none => unreachable,
.undef => unreachable,
.unreach => unreachable,
.dead => unreachable,
.immediate => |imm| {
try self.setRegOrMem(elem_ty, .{ .memory = imm }, value);
try self.setRegOrMem(value_ty, .{ .memory = imm }, value);
},
.ptr_stack_offset => |off| {
try self.genSetStack(elem_ty, off, value);
},
.ptr_embedded_in_code => |off| {
try self.setRegOrMem(elem_ty, .{ .embedded_in_code = off }, value);
},
.embedded_in_code => {
return self.fail("TODO implement storing to MCValue.embedded_in_code", .{});
try self.genSetStack(value_ty, off, value);
},
.register => {
return self.fail("TODO implement storing to MCValue.register", .{});
@ -1363,6 +1512,17 @@ fn airStore(self: *Self, inst: Air.Inst.Index) !void {
return self.fail("TODO implement storing to MCValue.stack_offset", .{});
},
}
}
fn airStore(self: *Self, inst: Air.Inst.Index) !void {
const bin_op = self.air.instructions.items(.data)[inst].bin_op;
const ptr = try self.resolveInst(bin_op.lhs);
const value = try self.resolveInst(bin_op.rhs);
const ptr_ty = self.air.typeOf(bin_op.lhs);
const value_ty = self.air.typeOf(bin_op.rhs);
try self.store(ptr, value, ptr_ty, value_ty);
return self.finishAir(inst, .dead, .{ bin_op.lhs, bin_op.rhs, .none });
}
@ -1508,7 +1668,6 @@ fn airCall(self: *Self, inst: Air.Inst.Index, modifier: std.builtin.CallOptions.
.immediate => unreachable,
.unreach => unreachable,
.dead => unreachable,
.embedded_in_code => unreachable,
.memory => unreachable,
.register => |reg| {
try self.register_manager.getReg(reg, null);
@ -1520,9 +1679,6 @@ fn airCall(self: *Self, inst: Air.Inst.Index, modifier: std.builtin.CallOptions.
.ptr_stack_offset => {
return self.fail("TODO implement calling with MCValue.ptr_stack_offset arg", .{});
},
.ptr_embedded_in_code => {
return self.fail("TODO implement calling with MCValue.ptr_embedded_in_code arg", .{});
},
}
}
@ -2052,7 +2208,6 @@ fn genSetReg(self: *Self, ty: Type, reg: Register, mcv: MCValue) InnerError!void
switch (mcv) {
.dead => unreachable,
.ptr_stack_offset => unreachable,
.ptr_embedded_in_code => unreachable,
.unreach, .none => return, // Nothing to do.
.undef => {
if (!self.wantSafety())
@ -2098,6 +2253,20 @@ fn genSetReg(self: *Self, ty: Type, reg: Register, mcv: MCValue) InnerError!void
return self.fail("TODO genSetReg 33-64 bit immediates for riscv64", .{}); // glhf
}
},
.register => |src_reg| {
// If the registers are the same, nothing to do.
if (src_reg.id() == reg.id())
return;
// mov reg, src_reg
_ = try self.addInst(.{
.tag = .mv,
.data = .{ .rr = .{
.rd = reg,
.rs = src_reg,
} },
});
},
.memory => |addr| {
// The value is in memory at a hard-coded address.
// If the type is a pointer, it means the pointer address is at this memory location.
@ -2321,27 +2490,6 @@ fn getResolvedInstValue(self: *Self, inst: Air.Inst.Index) MCValue {
}
}
/// If the MCValue is an immediate, and it does not fit within this type,
/// we put it in a register.
/// A potential opportunity for future optimization here would be keeping track
/// of the fact that the instruction is available both as an immediate
/// and as a register.
fn limitImmediateType(self: *Self, operand: Air.Inst.Ref, comptime T: type) !MCValue {
const mcv = try self.resolveInst(operand);
const ti = @typeInfo(T).Int;
switch (mcv) {
.immediate => |imm| {
// This immediate is unsigned.
const U = std.meta.Int(.unsigned, ti.bits - @boolToInt(ti.signedness == .signed));
if (imm >= math.maxInt(U)) {
return MCValue{ .register = try self.copyToTmpRegister(Type.initTag(.usize), mcv) };
}
},
else => {},
}
return mcv;
}
fn lowerDeclRef(self: *Self, tv: TypedValue, decl: *Module.Decl) InnerError!MCValue {
const ptr_bits = self.target.cpu.arch.ptrBitWidth();
const ptr_bytes: u64 = @divExact(ptr_bits, 8);

View File

@ -43,6 +43,9 @@ pub fn emitMir(
for (mir_tags) |tag, index| {
const inst = @intCast(u32, index);
switch (tag) {
.add => try emit.mirRType(inst),
.sub => try emit.mirRType(inst),
.addi => try emit.mirIType(inst),
.jalr => try emit.mirIType(inst),
.ld => try emit.mirIType(inst),
@ -56,6 +59,8 @@ pub fn emitMir(
.dbg_prologue_end => try emit.mirDebugPrologueEnd(),
.dbg_epilogue_begin => try emit.mirDebugEpilogueBegin(),
.mv => try emit.mirRR(inst),
.nop => try emit.mirNop(inst),
.ret => try emit.mirNop(inst),
@ -131,6 +136,17 @@ fn dbgAdvancePCAndLine(self: *Emit, line: u32, column: u32) !void {
}
}
fn mirRType(emit: *Emit, inst: Mir.Inst.Index) !void {
const tag = emit.mir.instructions.items(.tag)[inst];
const r_type = emit.mir.instructions.items(.data)[inst].r_type;
switch (tag) {
.add => try emit.writeInstruction(Instruction.add(r_type.rd, r_type.rs1, r_type.rs2)),
.sub => try emit.writeInstruction(Instruction.sub(r_type.rd, r_type.rs1, r_type.rs2)),
else => unreachable,
}
}
fn mirIType(emit: *Emit, inst: Mir.Inst.Index) !void {
const tag = emit.mir.instructions.items(.tag)[inst];
const i_type = emit.mir.instructions.items(.data)[inst].i_type;
@ -186,6 +202,15 @@ fn mirDebugEpilogueBegin(self: *Emit) !void {
}
}
fn mirRR(emit: *Emit, inst: Mir.Inst.Index) !void {
const tag = emit.mir.instructions.items(.tag)[inst];
const rr = emit.mir.instructions.items(.data)[inst].rr;
switch (tag) {
.mv => try emit.writeInstruction(Instruction.addi(rr.rd, rr.rs, 0)),
else => unreachable,
}
}
fn mirUType(emit: *Emit, inst: Mir.Inst.Index) !void {
const tag = emit.mir.instructions.items(.tag)[inst];
const u_type = emit.mir.instructions.items(.data)[inst].u_type;

View File

@ -24,6 +24,7 @@ pub const Inst = struct {
data: Data,
pub const Tag = enum(u16) {
add,
addi,
/// Pseudo-instruction: End of prologue
dbg_prologue_end,
@ -36,9 +37,11 @@ pub const Inst = struct {
jalr,
ld,
lui,
mv,
nop,
ret,
sd,
sub,
};
/// The position of an MIR instruction within the `Mir` instructions array.
@ -68,6 +71,13 @@ pub const Inst = struct {
///
/// Used by e.g. blr
reg: Register,
/// Two registers
///
/// Used by e.g. mv
rr: struct {
rd: Register,
rs: Register,
},
/// I-Type
///
/// Used by e.g. jalr
@ -76,6 +86,14 @@ pub const Inst = struct {
rs1: Register,
imm12: i12,
},
/// R-Type
///
/// Used by e.g. add
r_type: struct {
rd: Register,
rs1: Register,
rs2: Register,
},
/// U-Type
///
/// Used by e.g. lui