From 5138223a23241abd05a61de004c4ed7446d168d3 Mon Sep 17 00:00:00 2001
From: Mikael Lund <Mlund@mac.com>
Date: Thu, 19 Sep 2024 05:30:25 +0200
Subject: [PATCH] [commodore] Implement KERNAL LOAD function as inline assembly
 (#373)

* Rewrite cbm_k_load w. inline asm

* Add doxygen documentation

* Remove empty line

* Update doxygen documentation
---
 mos-platform/commodore/CMakeLists.txt |  2 +-
 mos-platform/commodore/cbm.h          | 24 ++++++++++++++++++++-
 mos-platform/commodore/cbm_k_load.c   | 31 +++++++++++++++++++++++++++
 mos-platform/commodore/cbm_k_load.s   | 23 --------------------
 4 files changed, 55 insertions(+), 25 deletions(-)
 create mode 100644 mos-platform/commodore/cbm_k_load.c
 delete mode 100644 mos-platform/commodore/cbm_k_load.s

diff --git a/mos-platform/commodore/CMakeLists.txt b/mos-platform/commodore/CMakeLists.txt
index cd3ba7375..b97111086 100644
--- a/mos-platform/commodore/CMakeLists.txt
+++ b/mos-platform/commodore/CMakeLists.txt
@@ -38,7 +38,7 @@ add_platform_library(commodore-c
   cbm_k_getin.c
   cbm_k_iobase.s
   cbm_k_listen.c
-  cbm_k_load.s
+  cbm_k_load.c
   cbm_k_open.s
   cbm_k_readst.c
   cbm_k_save.s
diff --git a/mos-platform/commodore/cbm.h b/mos-platform/commodore/cbm.h
index 605e2324f..c0146905c 100644
--- a/mos-platform/commodore/cbm.h
+++ b/mos-platform/commodore/cbm.h
@@ -231,7 +231,29 @@ void cbm_k_clrch (void);
 unsigned char cbm_k_getin (void);
 unsigned cbm_k_iobase (void) __attribute__((leaf));
 void cbm_k_listen (unsigned char dev);
-void *cbm_k_load(unsigned char flag, void *startaddr) __attribute__((leaf));
+
+/**
+ * @brief CBM KERNAL function to load or verify file
+ *
+ * Should be called after `cbm_k_setlfs()` and `cbm_k_setnam()`.
+ * On Commander X16, the `flag` argument takes the following
+ * values:
+ *
+ * `flag` | Description
+ * ------ | ---------------
+ *   0    | Load at address
+ *   1    | Verify
+ *   2    | Loads into VRAM $00000 + address (cx16)
+ *   3    | Loads into VRAM $10000 + address (cx16)
+ *
+ * On other Commodore targets, a non-zero `flag` value triggers verify.
+ *
+ * @param flag Load or verify
+ * @param load_addr Load address pointer (if not verifying).
+ * @return Address of last byte loaded/verified or KERNAL error code
+ */
+void *cbm_k_load(unsigned char flag, void *load_addr);
+
 unsigned char cbm_k_open (void) __attribute__((leaf));
 unsigned char cbm_k_readst (void);
 unsigned char cbm_k_save(void *startaddr, void *endaddr_plusone) __attribute__((leaf));
diff --git a/mos-platform/commodore/cbm_k_load.c b/mos-platform/commodore/cbm_k_load.c
new file mode 100644
index 000000000..2c3723ec8
--- /dev/null
+++ b/mos-platform/commodore/cbm_k_load.c
@@ -0,0 +1,31 @@
+// Copyright 2024 LLVM-MOS Project
+// Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+// See https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+// information.
+
+#include <stdint.h>
+
+// Helper to (dis)assemble a 16-bit integer
+typedef union {
+  uint16_t value;
+  struct {
+    uint8_t lo, hi;
+  };
+} Word;
+
+void *cbm_k_load(const unsigned char flag, void *load_addr) {
+  Word result;
+  const Word addr = {.value = (uint16_t)(load_addr)};
+
+  __attribute__((leaf)) __asm__ volatile(
+      "    jsr __LOAD   \n"
+      "    bcc 1f       \n" // no error if carry clear
+      "    tax          \n" // else get error code from A
+      "    ldy #0       \n"
+      "1:               \n" // no error
+      : "=x"(result.lo), "=y"(result.hi)
+      : "a"(flag), "x"(addr.lo), "y"(addr.hi)
+      : "p");
+
+  return (void *)(result.value);
+}
diff --git a/mos-platform/commodore/cbm_k_load.s b/mos-platform/commodore/cbm_k_load.s
deleted file mode 100644
index 739728886..000000000
--- a/mos-platform/commodore/cbm_k_load.s
+++ /dev/null
@@ -1,23 +0,0 @@
-.include "imag.inc"
-
-.text
-
-;
-; void *cbm_k_load(unsigned char verifyflag, void *startaddr);	// if result address < 256, likely an error code
-;                                a           rc2/3
-;
-.global cbm_k_load
-cbm_k_load:
-	ldx __rc2
-	ldy __rc2+1
-	jsr __LOAD
-	bcc noerror
-	; store error code in A
-	sta __rc2
-	ldx #0
-	stx __rc2+1
-	rts
-noerror:
-	stx __rc2	; store end+1 low
-	sty __rc2+1	; store end+1 high
-	rts