Fix unplugged controller memory writes, add some missing math routines, and slightly weaken DMA alignment requirements (#117)

---------

Co-authored-by: LittleCube <littlecubehax@gmail.com>

librecomp/src/cont.cpp

@@ -50,10 +50,12 @@ extern "C" void osContGetReadData_recomp(uint8_t* rdram, recomp_context* ctx) {
     osContGetReadData(dummy_data);
 
     for (int controller = 0; controller < MAXCONTROLLERS; controller++) {
-        MEM_H(6 * controller + 0, data) = dummy_data[controller].button;
+        if (dummy_data[controller].err_no == 0) {
-        MEM_B(6 * controller + 2, data) = dummy_data[controller].stick_x;
+            MEM_H(6 * controller + 0, data) = dummy_data[controller].button;
-        MEM_B(6 * controller + 3, data) = dummy_data[controller].stick_y;
+            MEM_B(6 * controller + 2, data) = dummy_data[controller].stick_x;
-        MEM_B(6 * controller + 4, data) = dummy_data[controller].err_no;
+            MEM_B(6 * controller + 3, data) = dummy_data[controller].stick_y;
+            MEM_B(6 * controller + 4, data) = dummy_data[controller].err_no;
+        }
     }
 }
 

librecomp/src/math_routines.cpp

@@ -80,3 +80,35 @@ extern "C" void __ull_to_f_recomp(uint8_t * rdram, recomp_context * ctx) {
 
     ctx->f0.fl = ret;
 }
+
+extern "C" void __ull_rshift_recomp(uint8_t * rdram, recomp_context * ctx) {
+    uint64_t a = (ctx->r4 << 32) | ((ctx->r5 << 0) & 0xFFFFFFFFu);
+    uint64_t b = (ctx->r6 << 32) | ((ctx->r7 << 0) & 0xFFFFFFFFu);
+    uint64_t ret = a >> b;
+
+    ctx->r2 = (int32_t)(ret >> 32);
+    ctx->r3 = (int32_t)(ret >> 0);
+}
+
+extern "C" void __ll_to_f_recomp(uint8_t * rdram, recomp_context * ctx) {
+    int64_t a = (ctx->r4 << 32) | ((ctx->r5 << 0) & 0xFFFFFFFFu);
+    float ret = (float)a;
+
+    ctx->f0.fl = ret;
+}
+
+extern "C" void __f_to_ll_recomp(uint8_t * rdram, recomp_context * ctx) {
+    int64_t ret = (int64_t)ctx->f12.fl;
+
+    ctx->r2 = (int32_t)(ret >> 32);
+    ctx->r3 = (int32_t)(ret >> 0);
+}
+
+extern "C" void __ll_lshift_recomp(uint8_t * rdram, recomp_context * ctx) {
+    uint64_t a = (ctx->r4 << 32) | ((ctx->r5 << 0) & 0xFFFFFFFFu);
+    uint64_t b = (ctx->r6 << 32) | ((ctx->r7 << 0) & 0xFFFFFFFFu);
+    uint64_t ret = a << b;
+
+    ctx->r2 = (int32_t)(ret >> 32);
+    ctx->r3 = (int32_t)(ret >> 0);
+}

librecomp/src/pi.cpp

@@ -67,7 +67,6 @@ void recomp::do_rom_read(uint8_t* rdram, gpr ram_address, uint32_t physical_addr
     // TODO handle misaligned DMA
     assert((physical_addr & 0x1) == 0 && "Only PI DMA from aligned ROM addresses is currently supported");
     assert((ram_address & 0x7) == 0 && "Only PI DMA to aligned RDRAM addresses is currently supported");
-    assert((num_bytes & 0x1) == 0 && "Only PI DMA with aligned sizes is currently supported");
     uint8_t* rom_addr = rom.data() + physical_addr - recomp::rom_base;
     for (size_t i = 0; i < num_bytes; i++) {
         MEM_B(i, ram_address) = *rom_addr;