libnum: add optimized num16 division for SH4AL-DSP

CMakeLists.txt

@@ -67,6 +67,7 @@ if(AZUR_PLATFORM STREQUAL gint)
   set(AZUR_PLATFORM_GENERIC TRUE)
   set(AZUR_TOOLKIT_GINT TRUE)
   set(AZUR_GRAPHICS_GINT_CG TRUE)
+  set(AZUR_ARCH_SH4ALDSP TRUE)
   set(AZUR_TERMINAL_NONE TRUE)
 
   # Link with gint and libprof (we only do this to get include directories)
@@ -84,6 +85,7 @@ if(AZUR_PLATFORM STREQUAL linux)
   set(AZUR_PLATFORM_GENERIC TRUE)
   set(AZUR_TOOLKIT_SDL TRUE)
   set(AZUR_GRAPHICS_OPENGL_3_3 TRUE)
+  set(AZUR_ARCH_GENERIC TRUE)
   set(AZUR_TERMINAL_ANSI TRUE)
 
   # Once again, we link only for the include paths
@@ -99,6 +101,7 @@ if(AZUR_PLATFORM STREQUAL emscripten)
   set(AZUR_PLATFORM_EMSCRIPTEN TRUE)
   set(AZUR_TOOLKIT_SDL TRUE)
   set(AZUR_GRAPHICS_OPENGL_ES_2_0 TRUE)
+  set(AZUR_ARCH_GENERIC TRUE)
   set(AZUR_TERMINAL_PLAIN TRUE)
 
   set(PORTS -sUSE_SDL=2 -sUSE_SDL_IMAGE=2 -sSDL2_IMAGE_FORMATS=["png"])

azur/include/azur/config.h.in

@@ -44,6 +44,14 @@
 #cmakedefine AZUR_GRAPHICS_GINT_CG
 
 
+/* CPU architecture (mostly useful for optimized SuperH stuff). */
+
+/* Generic/unknown CPU architecture; only C code is used. */
+#cmakedefine AZUR_ARCH_GENERIC
+/* SH4AL-DSP for the fx-CG. */
+#cmakedefine AZUR_ARCH_SH4ALDSP
+
+
 /* Input/output settings on terminal. */
 
 /* No terminal output entirely. */

libnum/CMakeLists.txt

@@ -7,16 +7,31 @@
 
 include(CTest)
 
-add_library(num STATIC
+set(SOURCES
   src/num.cpp
   src/str.cpp)
 
-target_include_directories(num PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}/include")
+if(DEFINED AZUR_ARCH_GENERIC)
+  set(LIBNUM_ARCH_GENERIC TRUE)
+elseif(DEFINED AZUR_ARCH_SH4ALDSP)
+  set(LIBNUM_ARCH_SH4ALDSP TRUE)
+  list(APPEND SOURCES
+    src/sh4aldsp/div_i32_i16_i16.s)
+endif()
+
+configure_file(include/num/config.h.in num/config.h)
+
+add_library(num STATIC ${SOURCES})
+target_include_directories(num PUBLIC
+  "${CMAKE_CURRENT_SOURCE_DIR}/include"
+  "${CMAKE_CURRENT_BINARY_DIR}")
 
 # Library file: libnum.a
 install(TARGETS num DESTINATION ${LIBDIR})
-# Headers: azur/*.h
+# Headers: num/config.h and num/*.h
 install(DIRECTORY include/ DESTINATION ${INCDIR})
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/num/config.h"
+        DESTINATION ${INCDIR}/num)
 
 #---
 # Unit tests

libnum/include/num/config.h.in

@@ -0,0 +1,15 @@
+//---------------------------------------------------------------------------//
+//  ,"  /\  ",    Azur: A game engine for CASIO fx-CG and PC                 //
+// |  _/__\_  |   Designed by Lephe' and the Planète Casio community.        //
+//  "._`\/'_."    License: MIT <https://opensource.org/licenses/MIT>         //
+//---------------------------------------------------------------------------//
+// num.config: Compile-time configuration options
+
+#pragma once
+
+/* CPU architecture. */
+
+/* Generic/unknown CPU architecture; only C code is used. */
+#cmakedefine LIBNUM_ARCH_GENERIC
+/* SH4AL-DSP for the fx-CG. */
+#cmakedefine LIBNUM_ARCH_SH4ALDSP

libnum/include/num/num.h

@@ -48,9 +48,8 @@
 
 #pragma once
 
-#include <cstdint>
+#include <num/primitives.h>
 #include <cstddef>
-
 #include <type_traits>
 #include <concepts>
 
@@ -219,8 +218,8 @@ struct num16
         v = (v * other.v) / 256;
         return *this;
     }
-    inline constexpr num16 &operator/=(num16 const &other) {
-        v = (v * 256) / other.v;
+    inline num16 &operator/=(num16 const &other) {
+        v = prim::div_i32_i16_i16(v * 256, other.v);
         return *this;
     }
     inline constexpr num16 &operator%=(num16 const &other) {

libnum/include/num/primitives.h

@@ -0,0 +1,37 @@
+//---------------------------------------------------------------------------//
+//  ,"  /\  ",    Azur: A game engine for CASIO fx-CG and PC                 //
+// |  _/__\_  |   Designed by Lephe' and the Planète Casio community.        //
+//  "._`\/'_."    License: MIT <https://opensource.org/licenses/MIT>         //
+//---------------------------------------------------------------------------//
+// num.primitives: Platform-specific optimized primitives
+//
+// This header provides a generic interface to optimized primitives such as
+// unusual-size divisions. These are used in libnum functions for a performance
+// boost.
+//
+// The following functions are defined:
+// - div_X_Y_Z: Division of an X by a Y with a result of type Z (where X,Y,Z is
+//              a signed integer type iN or an unsigned integer type uN)
+//---
+
+#pragma once
+
+#include <num/config.h>
+#include <cstdint>
+
+namespace libnum::prim {
+
+#ifdef LIBNUM_ARCH_SH4ALDSP
+
+extern int div_i32_i16_i16(int32_t num, int16_t denum);
+
+#else
+
+static inline constexpr int div_i32_i16_i16(int32_t num, int16_t denum)
+{
+    return num / denum;
+}
+
+#endif
+
+} /* namespace libnum::prim */

libnum/src/sh4aldsp/div_i32_i16_i16.s

@@ -0,0 +1,54 @@
+# --------------------------------------------------------------------------- #
+#  ,"  /\  ",    Azur: A game engine for CASIO fx-CG and PC                   #
+# |  _/__\_  |   Designed by Lephe' and the Planète Casio community.          #
+#  "._`\/'_."    License: MIT <https://opensource.org/licenses/MIT>           #
+# --------------------------------------------------------------------------- #
+# SH4AL-DSP optimized i32 / i16 -> i16 division.
+#
+# This simply uses the CPU's ability to divide without rotation for 16 bit
+# divisors without all the boilerplate than libgcc's __sdivisi3 requires since
+# it assumes 32-bit inputs. Used for num16 division.
+# ---
+
+.global __ZN6libnum4prim15div_i32_i16_i16Els
+
+# libnum::prim::div_i32_i16_i16(long, short)
+__ZN6libnum4prim15div_i32_i16_i16Els:
+	shll16	r5
+	mov	#0, r2
+
+	mov	r4, r3
+	rotcl	r3
+
+	subc	r2, r4
+
+	div0s	r5, r4
+
+	div1	r5, r4
+	div1	r5, r4
+	div1	r5, r4
+	div1	r5, r4
+
+	div1	r5, r4
+	div1	r5, r4
+	div1	r5, r4
+	div1	r5, r4
+
+	div1	r5, r4
+	div1	r5, r4
+	div1	r5, r4
+	div1	r5, r4
+
+	div1	r5, r4
+	div1	r5, r4
+	div1	r5, r4
+	div1	r5, r4
+
+	exts.w	r4, r4
+
+	rotcl	r4
+
+	addc	r2, r4
+
+	rts
+	exts.w	r4, r0