From 847b75078639a12d98b44473650833d3fd488fe9 Mon Sep 17 00:00:00 2001
From: DeaTh-G <55578911+DeaTh-G@users.noreply.github.com>
Date: Sun, 13 Oct 2024 18:12:23 +0200
Subject: [PATCH 1/8] Add more instructions regarding Bakugan Battle Brawlers

---
 XenonRecomp/recompiler.cpp | 219 ++++++++++++++++++++++++++++++++++++-
 1 file changed, 218 insertions(+), 1 deletion(-)

diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index f860817..4146def 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -808,6 +808,10 @@ bool Recompiler::Recompile(
         // no op
         break;
 
+    case PPC_INST_DCBST:
+        // no op
+        break;
+
     case PPC_INST_DCBTST:
         // no op
         break;
@@ -1035,6 +1039,12 @@ bool Recompiler::Recompile(
         println("{}.u32);", r(insn.operands[2]));
         break;
 
+    case PPC_INST_LBZUX:
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u64 = PPC_LOAD_U8({});", r(insn.operands[0]), ea());
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_LD:
         print("\t{}.u64 = PPC_LOAD_U64(", r(insn.operands[0]));
         if (insn.operands[2] != 0)
@@ -1063,6 +1073,12 @@ bool Recompiler::Recompile(
         println("{}.u32);", r(insn.operands[2]));
         break;
 
+    case PPC_INST_LDUX:
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u64 = PPC_LOAD_U64({});", r(insn.operands[0]), ea());
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_LFD:
         printSetFlushMode(false);
         print("\t{}.u64 = PPC_LOAD_U64(", f(insn.operands[0]));
@@ -1071,6 +1087,13 @@ bool Recompiler::Recompile(
         println("{});", int32_t(insn.operands[1]));
         break;
 
+    case PPC_INST_LFDU:
+        printSetFlushMode(false);
+        println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u64 = PPC_LOAD_U64({});", r(insn.operands[0]), ea());
+        println("\t{}.u32 = {};", r(insn.operands[2]), ea());
+        break;
+
     case PPC_INST_LFDX:
         printSetFlushMode(false);
         print("\t{}.u64 = PPC_LOAD_U64(", f(insn.operands[0]));
@@ -1079,6 +1102,13 @@ bool Recompiler::Recompile(
         println("{}.u32);", r(insn.operands[2]));
         break;
 
+    case PPC_INST_LFDUX:
+        printSetFlushMode(false);
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u64 = PPC_LOAD_U64({});", r(insn.operands[0]), ea());
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_LFS:
         printSetFlushMode(false);
         print("\t{}.u32 = PPC_LOAD_U32(", temp());
@@ -1088,6 +1118,14 @@ bool Recompiler::Recompile(
         println("\t{}.f64 = double({}.f32);", f(insn.operands[0]), temp());
         break;
 
+    case PPC_INST_LFSU:
+        printSetFlushMode(false);
+        println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u32 = PPC_LOAD_U32({});", temp(), ea());
+        println("\t{}.u32 = {};", r(insn.operands[2]), ea());
+        println("\t{}.f64 = double({}.f32);", f(insn.operands[0]), temp());
+        break;
+
     case PPC_INST_LFSX:
         printSetFlushMode(false);
         print("\t{}.u32 = PPC_LOAD_U32(", temp());
@@ -1097,6 +1135,14 @@ bool Recompiler::Recompile(
         println("\t{}.f64 = double({}.f32);", f(insn.operands[0]), temp());
         break;
 
+    case PPC_INST_LFSUX:
+        printSetFlushMode(false);
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u32 = PPC_LOAD_U32({});", temp(), ea());
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        println("\t{}.f64 = double({}.f32);", f(insn.operands[0]), temp());
+        break;
+
     case PPC_INST_LHA:
         print("\t{}.s64 = int16_t(PPC_LOAD_U16(", r(insn.operands[0]));
         if (insn.operands[2] != 0)
@@ -1104,6 +1150,12 @@ bool Recompiler::Recompile(
         println("{}));", int32_t(insn.operands[1]));
         break;
 
+    case PPC_INST_LHAU:
+        print("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
+        print("\t{}.s64 = int16_t(PPC_LOAD_U16({}));", r(insn.operands[0]), ea());
+        print("\t{}.u32 = {};", r(insn.operands[2]), ea());
+        break;
+
     case PPC_INST_LHAX:
         print("\t{}.s64 = int16_t(PPC_LOAD_U16(", r(insn.operands[0]));
         if (insn.operands[1] != 0)
@@ -1118,6 +1170,12 @@ bool Recompiler::Recompile(
         println("{});", int32_t(insn.operands[1]));
         break;
 
+    case PPC_INST_LHZU:
+        println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u64 = PPC_LOAD_U16({});", r(insn.operands[0]), ea());
+        println("\t{}.u32 = {};", r(insn.operands[2]), ea());
+        break;
+
     case PPC_INST_LHZX:
         print("\t{}.u64 = PPC_LOAD_U16(", r(insn.operands[0]));
         if (insn.operands[1] != 0)
@@ -1125,6 +1183,12 @@ bool Recompiler::Recompile(
         println("{}.u32);", r(insn.operands[2]));
         break;
 
+    case PPC_INST_LHZUX:
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u64 = PPC_LOAD_U16({});", r(insn.operands[0]), ea());
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_LI:
         println("\t{}.s64 = {};", r(insn.operands[0]), int32_t(insn.operands[1]));
         break;
@@ -1137,6 +1201,7 @@ bool Recompiler::Recompile(
     case PPC_INST_LVEWX128:
     case PPC_INST_LVX:
     case PPC_INST_LVX128:
+    case PPC_INST_LVEHX:
         // NOTE: for endian swapping, we reverse the whole vector instead of individual elements.
         // this is accounted for in every instruction (eg. dp3 sums yzw instead of xyz)
         print("\t_mm_store_si128((__m128i*){}.u8, _mm_shuffle_epi8(_mm_load_si128((__m128i*)(base + ((", v(insn.operands[0]));
@@ -1232,6 +1297,12 @@ bool Recompiler::Recompile(
         println("{}.u32);", r(insn.operands[2]));
         break;
 
+    case PPC_INST_LWZUX:
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u64 = PPC_LOAD_U32({});", r(insn.operands[0]), ea());
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_MFCR:
         for (size_t i = 0; i < 32; i++)
         {
@@ -1493,6 +1564,12 @@ bool Recompiler::Recompile(
         println("{}.u32, {}.u8);", r(insn.operands[2]), r(insn.operands[0]));
         break;
 
+    case PPC_INST_STBUX:
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\tPPC_STORE_U8({}, {}.u8);", ea(), r(insn.operands[0]));
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_STD:
         print("{}", mmioStore() ? "\tPPC_MM_STORE_U64(" : "\tPPC_STORE_U64(");
         if (insn.operands[2] != 0)
@@ -1523,6 +1600,12 @@ bool Recompiler::Recompile(
         println("{}.u32, {}.u64);", r(insn.operands[2]), r(insn.operands[0]));
         break;
 
+    case PPC_INST_STDUX:
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\tPPC_STORE_U64({}, {}.u64);", ea(), r(insn.operands[0]));
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_STFD:
         printSetFlushMode(false);
         print("{}", mmioStore() ? "\tPPC_MM_STORE_U64(" : "\tPPC_STORE_U64(");
@@ -1531,6 +1614,13 @@ bool Recompiler::Recompile(
         println("{}, {}.u64);", int32_t(insn.operands[1]), f(insn.operands[0]));
         break;
 
+    case PPC_INST_STFDU:
+        printSetFlushMode(false);
+        println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
+        println("\tPPC_STORE_U64({}, {}.u64);", ea(), r(insn.operands[0]));
+        println("\t{}.u32 = {};", r(insn.operands[2]), ea());
+        break;
+
     case PPC_INST_STFDX:
         printSetFlushMode(false);
         print("{}", mmioStore() ? "\tPPC_MM_STORE_U64(" : "\tPPC_STORE_U64(");
@@ -1556,6 +1646,14 @@ bool Recompiler::Recompile(
         println("{}, {}.u32);", int32_t(insn.operands[1]), temp());
         break;
 
+    case PPC_INST_STFSU:
+        printSetFlushMode(false);
+        println("\t{}.f32 = float({}.f64);", temp(), f(insn.operands[0]));
+        println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
+        println("\tPPC_STORE_U32({}, {}.u32);", ea(), temp());
+        println("\t{}.u32 = {};", r(insn.operands[2]), ea());
+        break;
+
     case PPC_INST_STFSX:
         printSetFlushMode(false);
         println("\t{}.f32 = float({}.f64);", temp(), f(insn.operands[0]));
@@ -1565,6 +1663,14 @@ bool Recompiler::Recompile(
         println("{}.u32, {}.u32);", r(insn.operands[2]), temp());
         break;
 
+    case PPC_INST_STFSUX:
+        printSetFlushMode(false);
+        println("\t{}.f32 = float({}.f64);", temp(), f(insn.operands[0]));
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\tPPC_STORE_U32({}, {}.u32);", ea(), temp());
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_STH:
         print("{}", mmioStore() ? "\tPPC_MM_STORE_U16(" : "\tPPC_STORE_U16(");
         if (insn.operands[2] != 0)
@@ -1572,6 +1678,18 @@ bool Recompiler::Recompile(
         println("{}, {}.u16);", int32_t(insn.operands[1]), r(insn.operands[0]));
         break;
 
+    case PPC_INST_STHU:
+        println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
+        println("\tPPC_STORE_U16({}, {}.u16);", ea(), r(insn.operands[0]));
+        println("\t{}.u32 = {};", r(insn.operands[2]), ea());
+        break;
+
+    case PPC_INST_STHUX:
+        println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
+        println("\tPPC_STORE_U16({}, {}.u16);", ea(), r(insn.operands[0]));
+        println("\t{}.u32 = {};", r(insn.operands[1]), ea());
+        break;
+
     case PPC_INST_STHBRX:
         print("{}", mmioStore() ? "\tPPC_MM_STORE_U16(" : "\tPPC_STORE_U16(");
         if (insn.operands[1] != 0)
@@ -1740,10 +1858,18 @@ bool Recompiler::Recompile(
         println("\t_mm_store_ps({}.f32, _mm_add_ps(_mm_load_ps({}.f32), _mm_load_ps({}.f32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
+    case PPC_INST_VADDSBS:
+        println("\t_mm_store_si128((__m128i*){}.s8, _mm_adds_epi8(_mm_load_si128((__m128i*){}.s8), _mm_load_si128((__m128i*){}.s8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        break;
+
     case PPC_INST_VADDSHS:
         println("\t_mm_store_si128((__m128i*){}.s16, _mm_adds_epi16(_mm_load_si128((__m128i*){}.s16), _mm_load_si128((__m128i*){}.s16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
+    case PPC_INST_VADDSWS:
+        println("\t_mm_store_si128((__m128i*){}.s32, _mm_adds_epi32(_mm_load_si128((__m128i*){}.s32), _mm_load_si128((__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        break;
+
     case PPC_INST_VADDUBM:
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_add_epi8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
@@ -1785,6 +1911,10 @@ bool Recompiler::Recompile(
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_avg_epu8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
+    case PPC_INST_VAVGUH:
+        println("\t_mm_store_si128((__m128i*){}.u8, _mm_avg_epu16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        break;
+
     case PPC_INST_VCTSXS:
     case PPC_INST_VCFPSXWS128:
         printSetFlushMode(true);
@@ -1848,6 +1978,12 @@ bool Recompiler::Recompile(
             println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.u8), 0xFFFF);", cr(6), v(insn.operands[0]));
         break;
 
+    case PPC_INST_VCMPEQUH:
+        println("\t_mm_store_si128((__m128i*){}.u8, _mm_cmpeq_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.u16), 0xFFFF);", cr(6), v(insn.operands[0]));
+        break;
+
     case PPC_INST_VCMPEQUW:
     case PPC_INST_VCMPEQUW128:
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_cmpeq_epi32(_mm_load_si128((__m128i*){}.u32), _mm_load_si128((__m128i*){}.u32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
@@ -1873,10 +2009,26 @@ bool Recompiler::Recompile(
 
     case PPC_INST_VCMPGTUB:
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_cmpgt_epu8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.u8), 0xFFFF);", cr(6), v(insn.operands[0]));
         break;
 
     case PPC_INST_VCMPGTUH:
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_cmpgt_epu16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.u16), 0xFFFF);", cr(6), v(insn.operands[0]));
+        break;
+
+    case PPC_INST_VCMPGTSH:
+        println("\t_mm_store_si128((__m128i*){}.s8, _mm_cmpgt_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.s16), 0xFFFF);", cr(6), v(insn.operands[0]));
+        break;
+
+    case PPC_INST_VCMPGTSW:
+        println("\t_mm_store_si128((__m128i*){}.s8, _mm_cmpgt_epi32(_mm_load_si128((__m128i*){}.u32), _mm_load_si128((__m128i*){}.u32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.setFromMask(_mm_load_si128((__m128i*){}.s32), 0xFFFF);", cr(6), v(insn.operands[0]));
         break;
 
     case PPC_INST_VEXPTEFP:
@@ -1908,10 +2060,18 @@ bool Recompiler::Recompile(
         println("\t_mm_store_ps({}.f32, _mm_max_ps(_mm_load_ps({}.f32), _mm_load_ps({}.f32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
+    case PPC_INST_VMAXSH:
+        println("\t_mm_store_si128((__m128i*){}.u16, _mm_max_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        break;
+
     case PPC_INST_VMAXSW:
         println("\t_mm_store_si128((__m128i*){}.u32, _mm_max_epi32(_mm_load_si128((__m128i*){}.u32), _mm_load_si128((__m128i*){}.u32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
+    case PPC_INST_VMINSH:
+        println("\t_mm_store_si128((__m128i*){}.u16, _mm_max_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        break;
+
     case PPC_INST_VMINFP:
     case PPC_INST_VMINFP128:
         printSetFlushMode(true);
@@ -2042,11 +2202,26 @@ bool Recompiler::Recompile(
         }
         break;
 
+    case PPC_INST_VPKSHSS:
+    case PPC_INST_VPKSHSS128:
+        println("\t_mm_store_si128((__m128i*){}.u8, _mm_packs_epi16(_mm_load_si128((__m128i*){}.s16), _mm_load_si128((__m128i*){}.s16)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
+        break;
+
+    case PPC_INST_VPKSWSS:
+    case PPC_INST_VPKSWSS128:
+        println("\t_mm_store_si128((__m128i*){}.u8, _mm_packs_epi32(_mm_load_si128((__m128i*){}.s32), _mm_load_si128((__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
+        break;
+
     case PPC_INST_VPKSHUS:
     case PPC_INST_VPKSHUS128:
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_packus_epi16(_mm_load_si128((__m128i*){}.s16), _mm_load_si128((__m128i*){}.s16)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
         break;
 
+    case PPC_INST_VPKSWUS:
+    case PPC_INST_VPKSWUS128:
+        println("\t_mm_store_si128((__m128i*){}.u8, _mm_packus_epi32(_mm_load_si128((__m128i*){}.s32), _mm_load_si128((__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
+        break;
+
     case PPC_INST_VREFP:
     case PPC_INST_VREFP128:
         // TODO: see if we can use rcp safely
@@ -2088,6 +2263,7 @@ bool Recompiler::Recompile(
         break;
 
     case PPC_INST_VSEL:
+    case PPC_INST_VSEL128:
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_or_si128(_mm_andnot_si128(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)), _mm_and_si128(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8))));", v(insn.operands[0]), v(insn.operands[3]), v(insn.operands[1]), v(insn.operands[3]), v(insn.operands[2]));
         break;
 
@@ -2097,6 +2273,12 @@ bool Recompiler::Recompile(
             println("\t{}.u8[{}] = {}.u8[{}] << ({}.u8[{}] & 0x7);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i);
         break;
 
+    case PPC_INST_VSLH:
+        // TODO: vectorize
+        for (size_t i = 0; i < 8; i++)
+            println("\t{}.u16[{}] = {}.u16[{}] << ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i);
+        break;
+
     case PPC_INST_VSLDOI:
     case PPC_INST_VSLDOI128:
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_alignr_epi8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8), {}));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]), 16 - insn.operands[3]);
@@ -2130,6 +2312,10 @@ bool Recompiler::Recompile(
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_set1_epi8(char(0x{:X})));", v(insn.operands[0]), insn.operands[1]);
         break;
 
+    case PPC_INST_VSPLTISH:
+        println("\t_mm_store_si128((__m128i*){}.u16, _mm_set1_epi16(int(0x{:X})));", v(insn.operands[0]), insn.operands[1]);
+        break;
+
     case PPC_INST_VSPLTISW:
     case PPC_INST_VSPLTISW128:
         println("\t_mm_store_si128((__m128i*){}.u32, _mm_set1_epi32(int(0x{:X})));", v(insn.operands[0]), insn.operands[1]);
@@ -2149,6 +2335,18 @@ bool Recompiler::Recompile(
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_vsr(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
+    case PPC_INST_VSRAB:
+        // TODO: vectorize, ensure endianness is correct
+        for (size_t i = 0; i < 16; i++)
+            println("\t{}.s8[{}] = {}.s8[{}] >> ({}.u8[{}] & 0x7);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 4);
+        break;
+
+    case PPC_INST_VSRAH:
+        // TODO: vectorize, ensure endianness is correct
+        for (size_t i = 0; i < 8; i++)
+            println("\t{}.s16[{}] = {}.s16[{}] >> ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 4);
+        break;
+
     case PPC_INST_VSRAW:
     case PPC_INST_VSRAW128:
         // TODO: vectorize, ensure endianness is correct
@@ -2156,6 +2354,12 @@ bool Recompiler::Recompile(
             println("\t{}.s32[{}] = {}.s32[{}] >> ({}.u8[{}] & 0x1F);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 4);
         break;
 
+    case PPC_INST_VSRH:
+        // TODO: vectorize, ensure endianness is correct
+        for (size_t i = 0; i < 8; i++)
+            println("\t{}.u16[{}] = {}.u16[{}] >> ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 4);
+        break;
+
     case PPC_INST_VSRW:
     case PPC_INST_VSRW128:
         // TODO: vectorize, ensure endianness is correct
@@ -2169,6 +2373,15 @@ bool Recompiler::Recompile(
         println("\t_mm_store_ps({}.f32, _mm_sub_ps(_mm_load_ps({}.f32), _mm_load_ps({}.f32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
+    case PPC_INST_VSUBSHS:
+        // TODO: vectorize
+        for (size_t i = 0; i < 8; i++)
+        {
+            println("\t{}.s64 = int64_t({}.s16[{}]) - int64_t({}.s16[{}]);", temp(), v(insn.operands[1]), i, v(insn.operands[2]), i);
+            println("\t{}.s16[{}] = {}.s64 > SHRT_MAX ? SHRT_MAX : {}.s64 < SHRT_MIN ? SHRT_MIN : {}.s64;", v(insn.operands[0]), i, temp(), temp(), temp());
+        }
+        break;
+
     case PPC_INST_VSUBSWS:
         // TODO: vectorize
         for (size_t i = 0; i < 4; i++)
@@ -2182,8 +2395,12 @@ bool Recompiler::Recompile(
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_subs_epu8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
+    case PPC_INST_VSUBUBM:
+        println("\t_mm_store_si128((__m128i*){}.u8, _mm_sub_epi8(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        break;
+
     case PPC_INST_VSUBUHM:
-        println("\t_mm_store_si128((__m128i*){}.u8, _mm_sub_epi16(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        println("\t_mm_store_si128((__m128i*){}.u8, _mm_sub_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
     case PPC_INST_VUPKD3D128:

From f23d22bc7fb366d2dd82527113c4b5aef8383afe Mon Sep 17 00:00:00 2001
From: DeaTh-G <55578911+DeaTh-G@users.noreply.github.com>
Date: Sun, 13 Oct 2024 19:29:33 +0200
Subject: [PATCH 2/8] Fix indexing on certain instructions

---
 XenonRecomp/recompiler.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index 4146def..be02ed8 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -2338,13 +2338,13 @@ bool Recompiler::Recompile(
     case PPC_INST_VSRAB:
         // TODO: vectorize, ensure endianness is correct
         for (size_t i = 0; i < 16; i++)
-            println("\t{}.s8[{}] = {}.s8[{}] >> ({}.u8[{}] & 0x7);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 4);
+            println("\t{}.s8[{}] = {}.s8[{}] >> ({}.u8[{}] & 0x7);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i);
         break;
 
     case PPC_INST_VSRAH:
         // TODO: vectorize, ensure endianness is correct
         for (size_t i = 0; i < 8; i++)
-            println("\t{}.s16[{}] = {}.s16[{}] >> ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 4);
+            println("\t{}.s16[{}] = {}.s16[{}] >> ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 2);
         break;
 
     case PPC_INST_VSRAW:
@@ -2357,7 +2357,7 @@ bool Recompiler::Recompile(
     case PPC_INST_VSRH:
         // TODO: vectorize, ensure endianness is correct
         for (size_t i = 0; i < 8; i++)
-            println("\t{}.u16[{}] = {}.u16[{}] >> ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 4);
+            println("\t{}.u16[{}] = {}.u16[{}] >> ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 2);
         break;
 
     case PPC_INST_VSRW:

From f6193ebe430cb6ae1049c2b270333c1dc1bea349 Mon Sep 17 00:00:00 2001
From: DeaTh-G <55578911+DeaTh-G@users.noreply.github.com>
Date: Sun, 3 Nov 2024 15:42:24 +0100
Subject: [PATCH 3/8] add more basic instructions

---
 XenonRecomp/recompiler.cpp | 66 ++++++++++++++++++++++++++++++++++++--
 1 file changed, 64 insertions(+), 2 deletions(-)

diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index be02ed8..d7b311a 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -531,6 +531,13 @@ bool Recompiler::Recompile(
             println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
         break;
 
+    case PPC_INST_ADDC:
+        println("\t{}.ca = {}.u32 >= ~{}.u32;", xer(), r(insn.operands[2]), r(insn.operands[1]));
+        println("\t{}.u64 = {}.u64 + {}.u64;", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]));
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
+        break;
+
     case PPC_INST_ADDE:
         println("\t{}.u8 = ({}.u32 + {}.u32 < {}.u32) | ({}.u32 + {}.u32 + {}.ca < {}.ca);", temp(), r(insn.operands[1]), r(insn.operands[2]), r(insn.operands[1]), r(insn.operands[1]), r(insn.operands[2]), xer(), xer());
         println("\t{}.u64 = {}.u64 + {}.u64 + {}.ca;", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]), xer());
@@ -539,6 +546,14 @@ bool Recompiler::Recompile(
             println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
         break;
 
+    case PPC_INST_ADDME:
+        println("\t{}.u8 = ({}.u32 - 1 < {}.u32) | ({}.u32 - 1 + {}.ca < {}.ca);", temp(), r(insn.operands[1]), r(insn.operands[1]), r(insn.operands[1]), xer(), xer());
+        println("\t{}.u64 = {}.u64 - 1 + {}.ca;", r(insn.operands[0]), r(insn.operands[1]), xer());
+        println("\t{}.ca = {}.u8;", xer(), temp());
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
+        break;
+
     case PPC_INST_ADDI:
         print("\t{}.s64 = ", r(insn.operands[0]));
         if (insn.operands[1] != 0)
@@ -652,6 +667,14 @@ bool Recompiler::Recompile(
         println("\tif ({}.u32 == 0) goto loc_{:X};", ctr(), insn.operands[0]);
         break;
 
+    case PPC_INST_BDZF:
+    {
+        constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
+        println("\t--{}.u64;", ctr());
+        println("\tif ({}.u32 == 0 && !{}.{}) goto loc_{:X};", ctr(), cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], insn.operands[1]);
+        break;
+    }
+
     case PPC_INST_BDZLR:
         println("\t--{}.u64;", ctr());
         println("\tif ({}.u32 == 0) return;", ctr(), insn.operands[0]);
@@ -663,10 +686,20 @@ bool Recompiler::Recompile(
         break;
 
     case PPC_INST_BDNZF:
-        // NOTE: assuming eq here as a shortcut because all the instructions in the game do that
+    {
+        constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
         println("\t--{}.u64;", ctr());
-        println("\tif ({}.u32 != 0 && !{}.eq) goto loc_{:X};", ctr(), cr(insn.operands[0] / 4), insn.operands[1]);
+        println("\tif ({}.u32 != 0 && !{}.{}) goto loc_{:X};", ctr(), cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], insn.operands[1]);
         break;
+    }
+
+    case PPC_INST_BDNZT:
+    {
+        constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
+        println("\t--{}.u64;", ctr());
+        println("\tif ({}.u32 != 0 && {}.{}) goto loc_{:X};", ctr(), cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], insn.operands[1]);
+        break;
+    }
 
     case PPC_INST_BEQ:
         printConditionalBranch(false, "eq");
@@ -796,6 +829,20 @@ bool Recompiler::Recompile(
         println("\t{0}.u64 = {1}.u32 == 0 ? 32 : __builtin_clz({1}.u32);", r(insn.operands[0]), r(insn.operands[1]));
         break;
 
+    case PPC_INST_CROR:
+    {
+        constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
+        println("\t{}.{} = {}.{} | {}.{};", cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], cr(insn.operands[1] / 4), fields[insn.operands[1] % 4], cr(insn.operands[2] / 4), fields[insn.operands[2] % 4]);
+        break;
+    }
+
+    case PPC_INST_CRORC:
+    {
+        constexpr std::string_view fields[] = { "lt", "gt", "eq", "so" };
+        println("\t{}.{} = {}.{} | (~{}.{} & 1);", cr(insn.operands[0] / 4), fields[insn.operands[0] % 4], cr(insn.operands[1] / 4), fields[insn.operands[1] % 4], cr(insn.operands[2] / 4), fields[insn.operands[2] % 4]);
+        break;
+    }
+
     case PPC_INST_DB16CYC:
         // no op
         break;
@@ -856,6 +903,13 @@ bool Recompiler::Recompile(
         // no op
         break;
 
+    case PPC_INST_EQV:
+        println("\t{}.u64 = ~({}.u32 ^ {}.u32);", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]));
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
+
+        break;
+
     case PPC_INST_EXTSB:
         println("\t{}.s64 = {}.s8;", r(insn.operands[0]), r(insn.operands[1]));
         if (strchr(insn.opcode->name, '.'))
@@ -1823,6 +1877,14 @@ bool Recompiler::Recompile(
             println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
         break;
 
+    case PPC_INST_SUBFZE:
+        println("\t{}.u8 = (~{}.u32 < ~{}.u32) | (~{}.u32 + {}.ca < {}.ca);", temp(), r(insn.operands[1]), r(insn.operands[1]), r(insn.operands[1]), xer(), xer());
+        println("\t{}.u64 = ~{}.u64 + {}.ca;", r(insn.operands[0]), r(insn.operands[1]), xer());
+        println("\t{}.ca = {}.u8;", xer(), temp());
+        if (strchr(insn.opcode->name, '.'))
+            println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
+        break;
+
     case PPC_INST_SUBFIC:
         println("\t{}.ca = {}.u32 <= {};", xer(), r(insn.operands[1]), insn.operands[2]);
         println("\t{}.s64 = {} - {}.s64;", r(insn.operands[0]), int32_t(insn.operands[2]), r(insn.operands[1]));

From cea0b2fc38422543282b46ae709ef5714b165e0c Mon Sep 17 00:00:00 2001
From: DeaTh-G <55578911+DeaTh-G@users.noreply.github.com>
Date: Sun, 3 Nov 2024 19:16:26 +0100
Subject: [PATCH 4/8] Fix instruction implementations based on unit tests

---
 XenonRecomp/recompiler.cpp | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index d7b311a..bb453fb 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -532,7 +532,7 @@ bool Recompiler::Recompile(
         break;
 
     case PPC_INST_ADDC:
-        println("\t{}.ca = {}.u32 >= ~{}.u32;", xer(), r(insn.operands[2]), r(insn.operands[1]));
+        println("\t{}.ca = {}.u32 > ~{}.u32;", xer(), r(insn.operands[2]), r(insn.operands[1]));
         println("\t{}.u64 = {}.u64 + {}.u64;", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]));
         if (strchr(insn.opcode->name, '.'))
             println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
@@ -904,10 +904,9 @@ bool Recompiler::Recompile(
         break;
 
     case PPC_INST_EQV:
-        println("\t{}.u64 = ~({}.u32 ^ {}.u32);", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]));
+        println("\t{}.u64 = ~({}.u64 ^ {}.u64);", r(insn.operands[0]), r(insn.operands[1]), r(insn.operands[2]));
         if (strchr(insn.opcode->name, '.'))
             println("\t{}.compare<int32_t>({}.s32, 0, {});", cr(0), r(insn.operands[0]), xer());
-
         break;
 
     case PPC_INST_EXTSB:
@@ -2131,7 +2130,7 @@ bool Recompiler::Recompile(
         break;
 
     case PPC_INST_VMINSH:
-        println("\t_mm_store_si128((__m128i*){}.u16, _mm_max_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        println("\t_mm_store_si128((__m128i*){}.u16, _mm_min_epi16(_mm_load_si128((__m128i*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
     case PPC_INST_VMINFP:
@@ -2338,7 +2337,7 @@ bool Recompiler::Recompile(
     case PPC_INST_VSLH:
         // TODO: vectorize
         for (size_t i = 0; i < 8; i++)
-            println("\t{}.u16[{}] = {}.u16[{}] << ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i);
+            println("\t{}.u16[{}] = {}.u16[{}] << ({}.u8[{}] & 0xF);", v(insn.operands[0]), i, v(insn.operands[1]), i, v(insn.operands[2]), i * 2);
         break;
 
     case PPC_INST_VSLDOI:

From 1d452c60a8929156f3cb40dcfcc9c6388632e8e5 Mon Sep 17 00:00:00 2001
From: DeaTh-G <55578911+DeaTh-G@users.noreply.github.com>
Date: Sun, 3 Nov 2024 20:27:53 +0100
Subject: [PATCH 5/8] add vpkuhus implementation

---
 XenonRecomp/recompiler.cpp | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index bb453fb..c4ab869 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -2283,6 +2283,16 @@ bool Recompiler::Recompile(
         println("\t_mm_store_si128((__m128i*){}.u8, _mm_packus_epi32(_mm_load_si128((__m128i*){}.s32), _mm_load_si128((__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
         break;
 
+    case PPC_INST_VPKUHUS:
+    case PPC_INST_VPKUHUS128:
+        for (size_t i = 0; i < 8; i++)
+        {
+            println("\t{0}.u8[{1}] = {2}.u16[{1}] > UCHAR_MAX ? UCHAR_MAX : {2}.u16[{1}];", vTemp(), i, v(insn.operands[2]));
+            println("\t{0}.u8[{1}] = {2}.u16[{3}] > UCHAR_MAX ? UCHAR_MAX : {2}.u16[{3}];", vTemp(), i + 8, v(insn.operands[1]), i);
+        }
+        println("{} = {};", v(insn.operands[0]), vTemp());
+        break;
+
     case PPC_INST_VREFP:
     case PPC_INST_VREFP128:
         // TODO: see if we can use rcp safely

From a5d63829753be07e845a5546c69720259a5e8bde Mon Sep 17 00:00:00 2001
From: DeaTh-G <55578911+DeaTh-G@users.noreply.github.com>
Date: Sun, 3 Nov 2024 21:27:33 +0100
Subject: [PATCH 6/8] add remaining altivec instructions

---
 XenonRecomp/recompiler.cpp | 18 ++++++++++++++++++
 XenonUtils/ppc_context.h   | 13 +++++++++++++
 2 files changed, 31 insertions(+)

diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index c4ab869..801d80f 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -1986,6 +1986,16 @@ bool Recompiler::Recompile(
             println("_mm_load_ps({}.f32)));", v(insn.operands[1]));
         break;
 
+    case PPC_INST_VCTUXS:
+    case PPC_INST_VCFPUXWS128:
+        printSetFlushMode(true);
+        print("\t_mm_store_si128((__m128i*){}.u32, _mm_vctuxs(", v(insn.operands[0]));
+        if (insn.operands[2] != 0)
+            println("_mm_mul_ps(_mm_load_ps({}.f32), _mm_set1_ps({}))));", v(insn.operands[1]), 1u << insn.operands[2]);
+        else
+            println("_mm_load_ps({}.f32)));", v(insn.operands[1]));
+        break;
+
     case PPC_INST_VCFSX:
     case PPC_INST_VCSXWFP128:
     {
@@ -2325,6 +2335,14 @@ bool Recompiler::Recompile(
         break;
     }
 
+    case PPC_INST_VRLH:
+        for (size_t i = 0; i < 8; i++)
+        {
+            println("\t{0}.u16[{1}] = ({2}.u16[{1}] << ({3}.u16[{1}] & 0xF)) | ({2}.u16[{1}] >> (16 - ({3}.u16[{1}] & 0xF)));", vTemp(), i, v(insn.operands[1]), v(insn.operands[2]));
+        }
+        println("{} = {};", v(insn.operands[0]), vTemp());
+        break;
+
     case PPC_INST_VRSQRTEFP:
     case PPC_INST_VRSQRTEFP128:
         // TODO: see if we can use rsqrt safely
diff --git a/XenonUtils/ppc_context.h b/XenonUtils/ppc_context.h
index c1091d1..d7ddbd5 100644
--- a/XenonUtils/ppc_context.h
+++ b/XenonUtils/ppc_context.h
@@ -645,6 +645,19 @@ inline __m128i _mm_vctsxs(__m128 src1)
     return _mm_andnot_si128(_mm_castps_si128(xmm2), _mm_castps_si128(dest));
 }
 
+inline __m128i _mm_vctuxs(__m128 src1)
+{
+    __m128 xmm0 = _mm_max_ps(src1, _mm_set1_epi32(0));
+    __m128 xmm1 = _mm_cmpge_ps(xmm0, _mm_set1_ps((float)0x80000000));
+    __m128 xmm2 = _mm_sub_ps(xmm0, _mm_set1_ps((float)0x80000000));
+    xmm0 = _mm_blendv_ps(xmm0, xmm2, xmm1);
+    __m128i dest = _mm_cvttps_epi32(xmm0);
+    xmm0 = _mm_cmpeq_epi32(dest, _mm_set1_epi32(INT_MIN));
+    xmm1 = _mm_and_si128(xmm1, _mm_set1_epi32(INT_MIN));
+    dest = _mm_add_epi32(dest, xmm1);
+    return _mm_or_si128(dest, xmm0);
+}
+
 inline __m128i _mm_vsr(__m128i a, __m128i b)
 {
     b = _mm_srli_epi64(_mm_slli_epi64(b, 61), 61);

From 830be1f69a9749322645ea51f935e7eb71d8c7cd Mon Sep 17 00:00:00 2001
From: DeaTh-G <hatvongeorge@gmail.com>
Date: Thu, 14 Nov 2024 17:34:45 +0100
Subject: [PATCH 7/8] fix vaddsws implementation

---
 XenonRecomp/recompiler.cpp | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index 801d80f..628fa6e 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -1928,7 +1928,12 @@ bool Recompiler::Recompile(
         break;
 
     case PPC_INST_VADDSWS:
-        println("\t_mm_store_si128((__m128i*){}.s32, _mm_adds_epi32(_mm_load_si128((__m128i*){}.s32), _mm_load_si128((__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        // TODO: vectorize
+        for (size_t i = 0; i < 4; i++)
+        {
+            println("\t{}.s64 = int64_t({}.s32[{}]) + int64_t({}.s32[{}]);", temp(), v(insn.operands[1]), i, v(insn.operands[2]), i);
+            println("\t{}.s32[{}] = {}.s64 > INT_MAX ? INT_MAX : {}.s64 < INT_MIN ? INT_MIN : {}.s64;", v(insn.operands[0]), i, temp(), temp(), temp());
+        }
         break;
 
     case PPC_INST_VADDUBM:

From 9ff80d8321063657ea0e44edcc7ae6a60532e2c1 Mon Sep 17 00:00:00 2001
From: DeaTh-G <hatvongeorge@gmail.com>
Date: Sun, 24 Nov 2024 10:50:58 +0100
Subject: [PATCH 8/8] make store instructions check for mmio

---
 XenonRecomp/recompiler.cpp | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index 628fa6e..3028b17 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -1606,7 +1606,7 @@ bool Recompiler::Recompile(
 
     case PPC_INST_STBU:
         println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U8({}, {}.u8);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u8);", mmioStore() ? "PPC_MM_STORE_U8(" : "PPC_STORE_U8(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[2]), ea());
         break;
 
@@ -1619,7 +1619,7 @@ bool Recompiler::Recompile(
 
     case PPC_INST_STBUX:
         println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U8({}, {}.u8);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u8);", mmioStore() ? "PPC_MM_STORE_U8(" : "PPC_STORE_U8(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[1]), ea());
         break;
 
@@ -1642,7 +1642,7 @@ bool Recompiler::Recompile(
 
     case PPC_INST_STDU:
         println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U64({}, {}.u64);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u64);", mmioStore() ? "PPC_MM_STORE_U64(" : "PPC_STORE_U64(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[2]), ea());
         break;
 
@@ -1655,7 +1655,7 @@ bool Recompiler::Recompile(
 
     case PPC_INST_STDUX:
         println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U64({}, {}.u64);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u64);", mmioStore() ? "PPC_MM_STORE_U64(" : "PPC_STORE_U64(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[1]), ea());
         break;
 
@@ -1670,7 +1670,7 @@ bool Recompiler::Recompile(
     case PPC_INST_STFDU:
         printSetFlushMode(false);
         println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U64({}, {}.u64);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u64);", mmioStore() ? "PPC_MM_STORE_U64(" : "PPC_STORE_U64(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[2]), ea());
         break;
 
@@ -1703,7 +1703,7 @@ bool Recompiler::Recompile(
         printSetFlushMode(false);
         println("\t{}.f32 = float({}.f64);", temp(), f(insn.operands[0]));
         println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U32({}, {}.u32);", ea(), temp());
+        println("\t{}{}, {}.u32);", mmioStore() ? "PPC_MM_STORE_U32(" : "PPC_STORE_U32(", ea(), temp());
         println("\t{}.u32 = {};", r(insn.operands[2]), ea());
         break;
 
@@ -1720,7 +1720,7 @@ bool Recompiler::Recompile(
         printSetFlushMode(false);
         println("\t{}.f32 = float({}.f64);", temp(), f(insn.operands[0]));
         println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U32({}, {}.u32);", ea(), temp());
+        println("\t{}{}, {}.u32);", mmioStore() ? "PPC_MM_STORE_U32(" : "PPC_STORE_U32(", ea(), temp());
         println("\t{}.u32 = {};", r(insn.operands[1]), ea());
         break;
 
@@ -1733,13 +1733,13 @@ bool Recompiler::Recompile(
 
     case PPC_INST_STHU:
         println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U16({}, {}.u16);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u16);", mmioStore() ? "PPC_MM_STORE_U16(" : "PPC_STORE_U16(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[2]), ea());
         break;
 
     case PPC_INST_STHUX:
         println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U16({}, {}.u16);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u16);", mmioStore() ? "PPC_MM_STORE_U16(" : "PPC_STORE_U16(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[1]), ea());
         break;
 
@@ -1838,13 +1838,13 @@ bool Recompiler::Recompile(
 
     case PPC_INST_STWU:
         println("\t{} = {} + {}.u32;", ea(), int32_t(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U32({}, {}.u32);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u32);", mmioStore() ? "PPC_MM_STORE_U32(" : "PPC_STORE_U32(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[2]), ea());
         break;
 
     case PPC_INST_STWUX:
         println("\t{} = {}.u32 + {}.u32;", ea(), r(insn.operands[1]), r(insn.operands[2]));
-        println("\tPPC_STORE_U32({}, {}.u32);", ea(), r(insn.operands[0]));
+        println("\t{}{}, {}.u32);", mmioStore() ? "PPC_MM_STORE_U32(" : "PPC_STORE_U32(", ea(), r(insn.operands[0]));
         println("\t{}.u32 = {};", r(insn.operands[1]), ea());
         break;