From ee2e81eed256f8b115ab54c226979ac7b264cd8b Mon Sep 17 00:00:00 2001
From: Tharo <17233964+Thar0@users.noreply.github.com>
Date: Sat, 11 May 2024 18:02:37 +0100
Subject: [PATCH] Implement doubleword multiply and divide, and conversions
 between doubleword and single/double precision floats

---
 src/main.cpp          |  4 ----
 src/recompilation.cpp | 34 ++++++++++++++++++++++++++++++++++
 2 files changed, 34 insertions(+), 4 deletions(-)

diff --git a/src/main.cpp b/src/main.cpp
index 3cdf1dd..10231ca 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -141,10 +141,6 @@ std::unordered_set<std::string> reimplemented_funcs{
     "__osInitialize_kmc",
     "__osInitialize_isv",
     "__osRdbSend",
-    // libgcc math routines (these throw off the recompiler)
-    "__udivdi3",
-    "__divdi3",
-    "__umoddi3",
     // ido math routines
     "__ull_div",
     "__ll_div",
diff --git a/src/recompilation.cpp b/src/recompilation.cpp
index e525278..b656f05 100644
--- a/src/recompilation.cpp
+++ b/src/recompilation.cpp
@@ -386,16 +386,28 @@ bool process_instruction(const RecompPort::Context& context, const RecompPort::C
     case InstrId::cpu_mult:
         print_line("result = S64(S32({}{})) * S64(S32({}{})); lo = S32(result >> 0); hi = S32(result >> 32)", ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt);
         break;
+    case InstrId::cpu_dmult:
+        print_line("MUL128_S({}{}, {}{}, hi, lo)", ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt);
+        break;
     case InstrId::cpu_multu:
         print_line("result = U64(U32({}{})) * U64(U32({}{})); lo = S32(result >> 0); hi = S32(result >> 32)", ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt);
         break;
+    case InstrId::cpu_dmultu:
+        print_line("MUL128_U({}{}, {}{}, hi, lo)", ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt);
+        break;
     case InstrId::cpu_div:
         // Cast to 64-bits before division to prevent artihmetic exception for s32(0x80000000) / -1
         print_line("lo = S32(S64(S32({}{})) / S64(S32({}{}))); hi = S32(S64(S32({}{})) % S64(S32({}{})))", ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt, ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt);
         break;
+    case InstrId::cpu_ddiv:
+        print_line("lo = S64({}{}) / S64({}{}); hi = S64({}{}) % S64({}{})", ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt, ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt);
+        break;
     case InstrId::cpu_divu:
         print_line("lo = S32(U32({}{}) / U32({}{})); hi = S32(U32({}{}) % U32({}{}))", ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt, ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt);
         break;
+    case InstrId::cpu_ddivu:
+        print_line("lo = U64({}{}) / U64({}{}); hi = U64({}{}) % U64({}{})", ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt, ctx_gpr_prefix(rs), rs, ctx_gpr_prefix(rt), rt);
+        break;
     case InstrId::cpu_mflo:
         print_line("{}{} = lo", ctx_gpr_prefix(rd), rd);
         break;
@@ -929,6 +941,28 @@ bool process_instruction(const RecompPort::Context& context, const RecompPort::C
         print_line("NAN_CHECK(ctx->f{}.d)", fs);
         print_line("ctx->f{}.fl = CVT_S_D(ctx->f{}.d)", fd, fs);
         break;
+    case InstrId::cpu_cvt_d_l:
+        print_line("CHECK_FR(ctx, {})", fd);
+        print_line("CHECK_FR(ctx, {})", fs);
+        print_line("ctx->f{}.d = CVT_D_L(ctx->f{}.u64)", fd, fs);
+        break;
+    case InstrId::cpu_cvt_l_d:
+        print_line("CHECK_FR(ctx, {})", fd);
+        print_line("CHECK_FR(ctx, {})", fs);
+        print_line("NAN_CHECK(ctx->f{}.d)", fs);
+        print_line("ctx->f{}.u64 = CVT_L_D(ctx->f{}.d)", fd, fs);
+        break;
+    case InstrId::cpu_cvt_s_l:
+        print_line("CHECK_FR(ctx, {})", fd);
+        print_line("CHECK_FR(ctx, {})", fs);
+        print_line("ctx->f{}.fl = CVT_S_L(ctx->f{}.u64)", fd, fs);
+        break;
+    case InstrId::cpu_cvt_l_s:
+        print_line("CHECK_FR(ctx, {})", fd);
+        print_line("CHECK_FR(ctx, {})", fs);
+        print_line("NAN_CHECK(ctx->f{}.fl)", fs);
+        print_line("ctx->f{}.u64 = CVT_L_S(ctx->f{}.fl)", fd, fs);
+        break;
     case InstrId::cpu_trunc_w_s:
         print_line("CHECK_FR(ctx, {})", fd);
         print_line("CHECK_FR(ctx, {})", fs);