runs instructions that read other instructions' outputs

crates/cpu/src/reg_alloc.rs

@@ -415,6 +415,16 @@ pub fn reg_alloc(config: &CpuConfig) {
                 .element()
                 .HdlNone(),
         );
+        connect(
+            unit_forwarding_info.unit_reg_frees[unit_index],
+            HdlOption::map(
+                ReadyValid::firing_data(unit_free_regs_tracker.free_in[0]),
+                |value| {
+                    #[hdl]
+                    UnitOutRegNum::<_> { value }
+                },
+            ),
+        );
         #[hdl]
         if let HdlSome(output) = unit_to_reg_alloc.output {
             #[hdl]

crates/cpu/src/unit/unit_base.rs

@@ -20,6 +20,7 @@ use std::marker::PhantomData;
 #[hdl]
 pub struct UnitForwardingInfo<UnitNumWidth: Size, OutRegNumWidth: Size, UnitCount: Size> {
     pub unit_output_writes: ArrayType<HdlOption<UnitOutputWrite<OutRegNumWidth>>, UnitCount>,
+    pub unit_reg_frees: ArrayType<HdlOption<UnitOutRegNum<OutRegNumWidth>>, UnitCount>,
     pub _phantom: PhantomData<UnitNumWidth>,
 }
 
@@ -277,7 +278,12 @@ pub fn unit_base<
         HdlOption::is_some(in_flight_ops_summary.empty_op_index),
     );
 
-    let unit_output_writes = unit_to_reg_alloc.unit_forwarding_info.unit_output_writes;
+    #[hdl]
+    let UnitForwardingInfo::<_, _, _> {
+        unit_output_writes,
+        unit_reg_frees,
+        _phantom: _,
+    } = unit_to_reg_alloc.unit_forwarding_info;
     #[hdl]
     let read_src_regs = wire(mop_ty.src_regs_ty());
     connect(
@@ -287,6 +293,26 @@ pub fn unit_base<
     #[hdl]
     let read_src_values = wire();
     connect(read_src_values, [PRegValue::zeroed(); COMMON_MOP_SRC_LEN]);
+    #[hdl]
+    let input_src_regs = wire(mop_ty.src_regs_ty());
+    connect(
+        input_src_regs,
+        repeat(config.p_reg_num().const_zero().cast_to_bits(), ConstUsize),
+    );
+    #[hdl]
+    let input_src_regs_valid = wire();
+    connect(input_src_regs_valid, [true; COMMON_MOP_SRC_LEN]);
+    let mut unit_output_regs_valid: Vec<MemBuilder<Bool>> = (0..Expr::ty(unit_output_writes).len())
+        .map(|unit_index| {
+            let mut mem = memory_with_loc(
+                &format!("unit_{unit_index}_output_regs_valid"),
+                Bool,
+                SourceLocation::caller(),
+            );
+            mem.depth(1 << config.out_reg_num_width);
+            mem
+        })
+        .collect();
     for unit_index in 0..Expr::ty(unit_output_writes).len() {
         let mut unit_output_regs = memory_with_loc(
             &format!("unit_{unit_index}_output_regs"),
@@ -308,17 +334,60 @@ pub fn unit_base<
             }
         }
 
+        for src_index in 0..COMMON_MOP_SRC_LEN {
+            let read_port = unit_output_regs_valid[unit_index].new_read_port();
+            let p_reg_num = input_src_regs[src_index].cast_bits_to(config.p_reg_num());
+            connect_any(read_port.addr, p_reg_num.unit_out_reg.value);
+            connect(read_port.en, false);
+            connect(read_port.clk, cd.clk);
+            #[hdl]
+            if UnitNum::is_index(p_reg_num.unit_num, unit_index) {
+                connect(read_port.en, true);
+                connect(input_src_regs_valid[src_index], read_port.data);
+            }
+        }
+
         let write_port = unit_output_regs.new_write_port();
         connect_any(write_port.addr, 0u8);
         connect(write_port.en, false);
         connect(write_port.clk, cd.clk);
         connect(write_port.data, PRegValue::zeroed());
         connect(write_port.mask, splat_mask(PRegValue, true.to_expr()));
+        let ready_write_port = unit_output_regs_valid[unit_index].new_write_port();
+        connect_any(ready_write_port.addr, 0u8);
+        connect(ready_write_port.en, false);
+        connect(ready_write_port.clk, cd.clk);
+        connect(ready_write_port.data, true);
+        connect(ready_write_port.mask, true);
         #[hdl]
         if let HdlSome(unit_output_write) = unit_output_writes[unit_index] {
             connect_any(write_port.addr, unit_output_write.which.value);
             connect(write_port.data, unit_output_write.value);
             connect(write_port.en, true);
+            connect_any(ready_write_port.addr, unit_output_write.which.value);
+            connect(ready_write_port.en, true);
+            let p_reg_num = #[hdl]
+            PRegNum::<_, _> {
+                unit_num: config.unit_num().from_index(unit_index),
+                unit_out_reg: unit_output_write.which,
+            };
+            for src_index in 0..COMMON_MOP_SRC_LEN {
+                #[hdl]
+                if input_src_regs[src_index].cmp_eq(p_reg_num.cast_to_bits()) {
+                    connect(input_src_regs_valid[src_index], true);
+                }
+            }
+        }
+        let free_write_port = unit_output_regs_valid[unit_index].new_write_port();
+        connect_any(free_write_port.addr, 0u8);
+        connect(free_write_port.en, false);
+        connect(free_write_port.clk, cd.clk);
+        connect(free_write_port.data, false);
+        connect(free_write_port.mask, true);
+        #[hdl]
+        if let HdlSome(unit_reg_free) = unit_reg_frees[unit_index] {
+            connect_any(free_write_port.addr, unit_reg_free.value);
+            connect(free_write_port.en, true);
         }
     }
 
@@ -364,16 +433,8 @@ pub fn unit_base<
             SourceLocation::caller(),
             StaticType::TYPE,
         );
-        for src_index in 0..COMMON_MOP_SRC_LEN {
+        connect(src_ready_flags, input_src_regs_valid);
-            connect(
+        connect(input_src_regs, input_mop_src_regs);
-                src_ready_flags[src_index],
-                config
-                    .p_reg_num()
-                    .const_zero()
-                    .cast_to_bits()
-                    .cmp_eq(input_mop_src_regs[src_index]),
-            );
-        }
         #[hdl]
         if unit_to_reg_alloc.cancel_input.cmp_ne(HdlSome(
             #[hdl]
@@ -494,8 +555,11 @@ pub fn unit_base<
                     #[hdl]
                     if !in_flight_op_canceling[in_flight_op_index] {
                         #[hdl]
-                        if let InFlightOpState::Running = state {
+                        match state {
-                            connect(unit_to_reg_alloc.output, HdlSome(unit_output));
+                            InFlightOpState::Running | InFlightOpState::Ready => {
+                                connect(unit_to_reg_alloc.output, HdlSome(unit_output))
+                            }
+                            InFlightOpState::CanceledAndRunning => {}
                         }
                     }
                 }

crates/cpu/tests/reg_alloc.rs

@@ -47,12 +47,41 @@ fn test_reg_alloc() {
             ),
         },
     );
-    sim.write(
+    let insns = std::array::from_fn::<_, 4, _>(|i| {
-        fetch_decode_interface.decoded_insns[0].data,
+        MOp.AluBranch(MOp.AluBranch.AddSubI(
-        HdlSome(
             #[hdl]
-            FetchedDecodedMOp {
+            AddSubMOp {
-                mop: MOp.AluBranch(MOp.AluBranch.AddSub(
+                alu_common: #[hdl]
+                AluCommonMOp {
+                    common: CommonMOp::new(
+                        0_hdl_u0,
+                        #[hdl]
+                        MOpDestReg {
+                            normal_regs: #[hdl]
+                            [
+                                #[hdl]
+                                MOpRegNum { value: i as u8 + 1 },
+                                MOpRegNum::const_zero(),
+                            ],
+                            flag_regs: #[hdl]
+                            [HdlSome(()), HdlNone()],
+                        },
+                        [0u8; 2],
+                        0x12345678u32.cast_to(SInt[COMMON_MOP_2_IMM_WIDTH]),
+                    ),
+                    output_integer_mode: OutputIntegerMode.DupLow32(),
+                },
+                invert_src0: false,
+                src1_is_carry_in: false,
+                invert_carry_in: false,
+                add_pc: false,
+            },
+        ))
+    })
+    .into_iter()
+    .chain(
+        [
+            MOp.AluBranch(MOp.AluBranch.AddSub(
                 #[hdl]
                 AddSubMOp {
                     alu_common: #[hdl]
@@ -70,28 +99,18 @@ fn test_reg_alloc() {
                                 flag_regs: #[hdl]
                                 [HdlSome(()), HdlNone()],
                             },
-                                [2u8, 3u8, 4u8],
+                            [1u8, 0, 0],
-                                0x1234.cast_to(SInt[COMMON_MOP_3_IMM_WIDTH]),
+                            1.cast_to(SInt[COMMON_MOP_3_IMM_WIDTH]),
                         ),
                         output_integer_mode: OutputIntegerMode.Full64(),
                     },
-                        invert_src0: true,
+                    invert_src0: false,
-                        src1_is_carry_in: true,
+                    src1_is_carry_in: false,
-                        invert_carry_in: true,
+                    invert_carry_in: false,
-                        add_pc: true,
+                    add_pc: false,
                 },
             )),
-                is_unrelated_pc: true,
+            MOp.AluBranch(MOp.AluBranch.Logical(
-                pc: 0x1000_hdl_u64,
-            },
-        ),
-    );
-    sim.write(
-        fetch_decode_interface.decoded_insns[1].data,
-        HdlSome(
-            #[hdl]
-            FetchedDecodedMOp {
-                mop: MOp.AluBranch(MOp.AluBranch.Logical(
                 #[hdl]
                 LogicalMOp {
                     alu_common: #[hdl]
@@ -107,7 +126,7 @@ fn test_reg_alloc() {
                                 ],
                                 flag_regs: [HdlNone(), HdlSome(())],
                             },
-                                [3u8, 4u8],
+                            [2u8, 4u8],
                             SInt[COMMON_MOP_2_IMM_WIDTH].zero(),
                         ),
                         output_integer_mode: OutputIntegerMode.Full64(),
@@ -115,12 +134,30 @@ fn test_reg_alloc() {
                     lut: 0b0110_hdl_u4,
                 },
             )),
-                is_unrelated_pc: false,
+        ]
-                pc: 0x1004_hdl_u64,
+        .into_iter()
+        .cycle(),
+    );
+    let mut insn_index = 0;
+    for cycle in 0..20 {
+        for fetch_index in 0..config.fetch_width.get() {
+            let mop = insns.clone().nth(insn_index).unwrap();
+            sim.write(
+                fetch_decode_interface.decoded_insns[fetch_index].data,
+                HdlSome(
+                    #[hdl]
+                    FetchedDecodedMOp {
+                        mop,
+                        is_unrelated_pc: insn_index == 0,
+                        pc: 0x1000u64 + 4 * insn_index as u64,
                     },
                 ),
             );
-    for cycle in 0..20 {
+            insn_index += 1;
+        }
+        if cycle == 0 {
+            insn_index = 0;
+        }
         sim.advance_time(SimDuration::from_nanos(500));
         sim.write_clock(sim.io().cd.clk, true);
         sim.advance_time(SimDuration::from_nanos(500));