Move reset vector back to 0x0, add MLOGSYS instruction to initialize icache instead of doing it automatically at boot

Author: object-Object

README.md

@@ -6,9 +6,11 @@ RISC-V processor in Mindustry logic. Requires Mindustry build 149+.
 
 ## Architecture
 
-Physical memory consists of three sections. Two are directly accessible by code: ROM (rx) and RAM (rwx). The third section is an instruction cache which is 4x less dense than main memory. The instruction cache is updated at reset, and whenever an instruction writes to RAM that is covered by the icache. If executing from memory not covered by the icache, the processor manually fetches and decodes the instruction from main memory.
+Physical memory consists of three sections. Two are directly accessible by code: ROM (rx) and RAM (rwx). The third section is an instruction cache which is 4x less dense than main memory.
 
-Code begins executing at address `0x4`. Address `0x0` must contain the size of the `.text` section (ie. `__etext`) to tell the processor how much data to decode from ROM. If this value is `0`, no ROM data will be decoded.
+The instruction cache should be initialized at boot using the MLOGSYS instruction. It is also updated whenever an instruction writes to RAM that is covered by the icache. If executing from memory not covered by the icache, the processor manually fetches and decodes the instruction from main memory.
+
+Code begins executing at address `0x00000000` (ie. the start of ROM).
 
 The main CPU code is generated from `src/main.mlog.jinja` using a custom Jinja-based preprocessor (`python/src/mlogv32/preprocessor`).
 
@@ -60,15 +62,16 @@ This non-standard extension adds instructions to control the mlogv32 processor's
 
 The MLOG instructions are encoded with an I-type instruction format using the _custom-0_ opcode. The zero-extended immediate is used as a minor opcode (funct12) for implementation reasons.
 
-The MLOGSYS instruction is used for simple system controls, including halt, `printchar`, `printflush`, `drawflush`, and sortKB/kbconv integration.
+The MLOGSYS instruction is used for simple system controls, including halt, `printchar`, `printflush`, `drawflush`, sortKB/kbconv integration, and icache initialization.
 
-| funct12 | rd                      | rs1     | name                       |
-| ------- | ----------------------- | ------- | -------------------------- |
-| 0       | `00000`                 | `00000` | Halt                       |
-| 1       | `00000`                 | char    | `printchar`                |
-| 2       | `00000`                 | `00000` | `printflush`               |
-| 3       | `00000`                 | `00000` | `drawflush`                |
-| 4       | char (0 if no new data) | `00000` | Read next char from sortKB |
+| funct12 | rd                      | rs1       | name                       |
+| ------- | ----------------------- | --------- | -------------------------- |
+| 0       | `00000`                 | `00000`   | Halt                       |
+| 1       | `00000`                 | char      | `printchar`                |
+| 2       | `00000`                 | `00000`   | `printflush`               |
+| 3       | `00000`                 | `00000`   | `drawflush`                |
+| 4       | char (0 if no new data) | `00000`   | Read next char from sortKB |
+| 5       | `00000`                 | \_\_etext | Decode ROM .text section   |
 
 The MLOGDRAW instruction is used for drawing graphics using the Mlog `draw` instruction. Arguments are passed to this instruction using registers a0 to a5 as necessary, and any return value is placed in _rd_. If _rd_ is specified as `00000` in the below table, no value will be written to `rd` in any case.
 

coremark/mlogv32/entry.s

@@ -2,6 +2,10 @@
 
 .global _start
 _start:
+    # initialize icache
+    la t0, __etext
+    .insn i CUSTOM_0, 0, zero, t0, 5
+
     # reset mtime/mcycle/minstret
     li t0, 0xf0000000 # mtime
     sw zero, 4(t0)

riscof/mlogv32/env/model_test.h

@@ -10,9 +10,11 @@
 #define RVMODEL_HALT                                              \
   .insn i CUSTOM_0, 0, zero, zero, 0;
 
+// initialize icache, .data, and .bss
 #define RVMODEL_BOOT \
   .option norelax; \
-  .word __etext; \
+  la t0, __etext; \
+  .insn i CUSTOM_0, 0, zero, t0, 5; \
   la t0, __sidata; \
   la t1, __sdata; \
   la t2, __edata; \

riscof/mlogv32/mlogv32_platform.yaml

@@ -1,4 +1,4 @@
 reset:
-  address: 0x4
+  address: 0x0
 nmi:
   label: nmi_vector # we don't have an nmi vector, but riscof gets mad if this is removed

rust/mlogv32/link.x

@@ -34,8 +34,6 @@ SECTIONS {
     .text : {
         __stext = .;
 
-        LONG(__etext);
-
         KEEP(*(.text.start));
         *(.text.reset);
         *(.text .text.*);

rust/mlogv32/src/lib.rs

@@ -25,10 +25,13 @@ pub mod prelude;
 #[global_allocator]
 static HEAP: Heap = Heap::empty();
 
-// init global/stack/frame pointers
+// init icache and global/stack/frame pointers
 #[rustfmt::skip]
 global_asm!("
 .section .text.start
+    la t0, __etext
+    .insn i CUSTOM_0, 0, zero, t0, 5
+
     la gp, __global_pointer$
     
     la t1, _stack_start

src/main.mlog.jinja

@@ -58,44 +58,9 @@ reset__find_lookup_start:
 
     set LOOKUP_START i
 
-    # read __etext from address 0, or default to ROM_SIZE if __etext is 0
-    set address 0
-    op add ret @counter 1
-    jump load_rom_word_unchecked always
-    set __etext result
-
-    jump reset__valid_etext lessThanEq __etext ROM_SIZE
-    print "Invalid __etext found at address 0x0, must be <={0}"
-    format ROM_SIZE
-    jump halt always
-reset__valid_etext:
-
-    # decode .text
-    set STATE "decoding"
-    set decode_address 4
-    set decode_var null
-
-reset__decode:
-    set address decode_address
-    op add ret @counter 1
-    jump load_rom_word_unchecked always
-    set instruction result
-
-    op add decode_ret @counter 1
-    jump decode always
-
-    op add decode_address decode_address 4
-    jump reset__decode lessThan decode_address __etext
-
-    set STATE "running"
-
     # clear LR/SC reservation set
     set reservation_set null
 
-    # start executing code at address 4 (address 0 is __etext)
-    set pc 4
-    set icache_var null
-
     # read initial CSR values
     read mtime {{CSRS}} "{{ 'time'|csr }}"
     read mtimeh {{CSRS}} "{{ 'timeh'|csr }}"
@@ -104,6 +69,12 @@ reset__decode:
     read minstret {{CSRS}} "{{ 'minstret'|csr }}"
     read minstreth {{CSRS}} "{{ 'minstreth'|csr }}"
 
+    # start executing code at address 0
+    set pc 0
+    set icache_var null
+    set __etext 0
+    set STATE "running"
+
     # schedule time update
     set last_time_update @time
     set last_cycle_update @tick
@@ -314,7 +285,6 @@ end_instruction_preserve_instret:
 # enable the reset switch, wait until it's disabled, then reset the processor
 # this must stay directly after end_instruction
 halt:
-    printflush {{ERROR_OUTPUT}}
     control enabled {{RESET_SWITCH}} true
     jump reset always
 
@@ -327,6 +297,7 @@ illegal_instruction:
 trap:
     print "Unhandled trap: "
     print mcause
+    printflush {{ERROR_OUTPUT}}
     jump halt always
 
 # helper functions
@@ -1712,8 +1683,8 @@ read_modify_write_csr__minstreth:
     jump end_instruction_with_result always
 
 MLOGSYS:
-    # I-type: arg1=rs1, arg2=imm, arg3=rd
-    jump illegal_instruction greaterThanEq arg2 5
+    # I-type: arg1=rs1, arg2=funct12, arg3=rd
+    jump illegal_instruction greaterThan arg2 5
 
     op mul jump arg2 2
     op add @counter @counter jump
@@ -1735,8 +1706,38 @@ MLOGSYS:
     jump end_instruction always
 
     # read sortKB
+    jump MLOGSYS__read_sortKB always
+    stop
+
+    # initialize icache
     # NOTE: if we add more instructions to MLOGSYS, this needs to move to its own label
 
+    # TODO: is this the right trap to use here?
+    jump illegal_instruction greaterThan rs1 ROM_SIZE
+
+    set __etext rs1
+
+    # decode .text
+    set STATE "decoding"
+    set decode_address 0
+    set decode_var null
+
+MLOGSYS__init_icache__loop:
+    set address decode_address
+    op add ret @counter 1
+    jump load_rom_word_unchecked always
+    set instruction result
+
+    op add decode_ret @counter 1
+    jump decode always
+
+    op add decode_address decode_address 4
+    jump MLOGSYS__init_icache__loop lessThan decode_address __etext
+
+    set STATE "running"
+    jump end_instruction always
+
+MLOGSYS__read_sortKB:
     # check the current ptr and return 0 if we've already reached it
     # NOTE: this will fail if the user has typed 63 characters since we last checked
     set result 0
@@ -1754,7 +1755,7 @@ MLOGSYS:
 
 MLOGDRAW:
     # I-type: arg1=rs1, arg2=imm, arg3=rd
-    jump illegal_instruction greaterThanEq arg2 16
+    jump illegal_instruction greaterThan arg2 15
 
     read a0 {{REGISTERS}} 10
     read a1 {{REGISTERS}} 11