Rearrange LABELS to remove 2 instructions from the decode hot path

Author: object-Object

python/src/mlogv32/preprocessor/app.py

@@ -263,7 +263,7 @@ def _render_template(
     check_unsaved_variables(worker_ast)
     print(
         f"""\
-Code size:
+Worker:
   Instructions: {len(list(n for n in worker_ast if isinstance(n, Statement)))} / 1000
   Bytes: {len(worker_code.encode())} / {1024 * 100}"""
     )
@@ -287,6 +287,14 @@ def _render_template(
         **config.inputs,
     )
 
+    controller_ast = parse_mlog(controller_code)
+    print(
+        f"""\
+Controller:
+  Instructions: {len(list(n for n in controller_ast if isinstance(n, Statement)))} / 1000
+  Bytes: {len(controller_code.encode())} / {1024 * 100}"""
+    )
+
     # preprocess other code snippets
 
     debugger_code, _, _ = _render_template(config.templates.debugger)

python/src/mlogv32/preprocessor/models.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import math
 from contextlib import contextmanager
 from contextvars import ContextVar
 from pathlib import Path
@@ -52,6 +53,8 @@ class Instruction(BaseModel):
         label: str
         count: int = Field(default=1, ge=1)
         up_to: int | None = None
+        align: int = Field(default=1, ge=1)
+        address: int | None = None
 
     class CSR(BaseModel):
         read: CSRLocation
@@ -65,9 +68,18 @@ def _resolve_instructions(cls, instructions: list[Instruction]):
         address = 0
 
         for i, value in enumerate(instructions):
+            value = value.model_copy()
+            if value.address is None:
+                value.address = math.ceil(address / value.align) * value.align
+            elif value.address < address:
+                raise ValueError(
+                    f"Instruction {i} is at address {value.address}, but the previous instruction was at address {address}"
+                )
+            address = value.address
+
             if value.up_to is not None:
-                value = value.model_copy()
                 assert value.up_to is not None
+                assert value.align == 1, "Align != 1 is not supported with up_to"
 
                 if value.up_to < address:
                     raise ValueError(
@@ -77,7 +89,7 @@ def _resolve_instructions(cls, instructions: list[Instruction]):
                 value.count = value.up_to - address + 1
 
             result.append(value)
-            address += value.count
+            address += value.count * value.align
 
         return result
 

src/cpu/controller.mlog.jinja

@@ -113,21 +113,22 @@ init_incr__not_null:
 
     # initialize label lookup table
 
-#% set ns = namespace(i=0, loop=0)
+#% set ns = namespace(loop=0)
 #% for instruction in instructions
+#% set ns.i = instruction.address
     # {{ instruction.label }}
 #%#- unroll the loop if it's equal or better in terms of code size, else generate a loop
 #% if instruction.count <= 4
 #% for _ in range(instruction.count)
     write {{labels[instruction.label]}} {{LABELS}} {{ns.i}}
-#% set ns.i = ns.i + 1
+#% set ns.i = ns.i + instruction.align
 #% endfor
 #% else
     set i {{ns.i}}
-#% set ns.i = ns.i + instruction.count
+#% set ns.i = ns.i + instruction.count * instruction.align
 init_labels__loop{{ns.loop}}:
     write {{labels[instruction.label]}} {{LABELS}} i
-    op add i i 1
+    op add i i {{instruction.align}}
     jump init_labels__loop{{ns.loop}} lessThan i {{ns.i}}
 #% set ns.loop = ns.loop + 1
 #% endif
@@ -389,6 +390,7 @@ set {{SINGLE_STEP_SWITCH}} null
 set {{POWER_SWITCH}} null
 set {{PAUSE_SWITCH}} null
 set {{labels[instruction.label]}} null
+set {{instruction.align}} null
 set {{loop.index0}} null
 set {{ns.i}} null
 set {{value}} null

src/cpu/cpu.yaml

@@ -34,8 +34,68 @@ inputs:
   UART_TX_WRITE: 511
 
 instructions:
+  # decoders
+  - label: decode_LOAD
+    align: 4
+  - label: decode_illegal_instruction # LOAD-FP
+    align: 4
+  - label: decode_custom-0
+    align: 4
+  - label: decode_MISC-MEM
+    align: 4
+  - label: decode_OP-IMM
+    align: 4
+  - label: decode_AUIPC
+    align: 4
+  - label: decode_illegal_instruction # OP-IMM-32
+    align: 4
+  - label: decode_illegal_instruction # 48b
+    align: 4
+  - label: decode_STORE
+    align: 4
+  - label: decode_illegal_instruction # STORE-FP
+    align: 4
+  - label: decode_illegal_instruction # custom-1
+    align: 4
+  - label: decode_AMO
+    align: 4
+  - label: decode_OP
+    align: 4
+  - label: decode_LUI
+    align: 4
+  - label: decode_illegal_instruction
+    align: 4
+    count: 10
+    # OP-32
+    # 64b
+    # MADD
+    # MSUB
+    # NMSUB
+    # NMADD
+    # OP-FP
+    # OP-V
+    # custom-2/rv128
+    # 48b
+  - label: decode_BRANCH
+    align: 4
+  - label: decode_JALR
+    align: 4
+  - label: decode_illegal_instruction # reserved
+    align: 4
+  - label: decode_JAL
+    align: 4
+  - label: decode_SYSTEM
+    align: 4
+  - label: decode_illegal_instruction
+    align: 4
+    count: 3
+    # OP-VE
+    # custom-3/rv128
+    # >=80b
+
   # runtime handlers
   - label: BEQ
+    address: 128
   - label: BNE
   - label: JAL
   - label: JALR
@@ -101,43 +161,8 @@ instructions:
   - label: SRAI
   - label: ILLEGAL_OP
   - label: MLOGSYS
-  # padding
   - label: ILLEGAL_OP
-    up_to: 127
-  # decoders, starting at index 128
-  - label: decode_LOAD
-  - label: decode_illegal_instruction # LOAD-FP
-  - label: decode_custom-0
-  - label: decode_MISC-MEM
-  - label: decode_OP-IMM
-  - label: decode_AUIPC
-  - label: decode_illegal_instruction # OP-IMM-32
-  - label: decode_illegal_instruction # 48b
-  - label: decode_STORE
-  - label: decode_illegal_instruction # STORE-FP
-  - label: decode_illegal_instruction # custom-1
-  - label: decode_AMO
-  - label: decode_OP
-  - label: decode_LUI
-  - label: decode_illegal_instruction # OP-32
-    count: 10
-    # 64b
-    # MADD
-    # MSUB
-    # NMSUB
-    # NMADD
-    # OP-FP
-    # OP-V
-    # custom-2/rv128
-    # 48b
-  - label: decode_BRANCH
-  - label: decode_JALR
-  - label: decode_illegal_instruction # reserved
-  - label: decode_JAL
-  - label: decode_SYSTEM
-  - label: decode_illegal_instruction # OP-VE
-  - label: decode_illegal_instruction # custom-3/rv128
-  - label: decode_illegal_instruction # >=80b
+    up_to: 255
 
 csrs:
   # -- unprivileged --

src/cpu/worker.mlog.jinja

@@ -323,7 +323,10 @@ main__no_breakpoint:
 
 #% do reset_locals()
     # jump to instruction handler
-    op add $index op_id 64 # op_id is in the range [-64, 63]
+    # op_id is in the range [-64, 63]
+    # and the instruction handlers in {{LABELS}} start at index 128
+    # so we need to add 64 + 128 = 192
+    op add $index op_id 192
     read @counter {{LABELS}} $index
 
 # this is a normal length for a label
@@ -1100,11 +1103,9 @@ decode:
     op shr $$funct3 value 12
     op and $$funct3 $$funct3 0b111
 
-    # jump table based on (opcode >> 2) since the low 2 bits don't matter
-    # the jump table is stored in LABELS by the controller, starting at index 128
-    op and $tmp value 0b1111111
-    op shr $tmp $tmp 2
-    op add $tmp $tmp 128
+    # jump table based on opcode[6:2] since the low 2 bits have already been checked
+    # the jump table is stored in LABELS by the controller, starting at index 0 at multiples of 4
+    op and $tmp value 0b1111100
     read @counter {{LABELS}} $tmp
 
 decode_LUI: