diff --git a/XenonRecomp/recompiler.cpp b/XenonRecomp/recompiler.cpp
index 39dc0e7..854fddc 100644
--- a/XenonRecomp/recompiler.cpp
+++ b/XenonRecomp/recompiler.cpp
@@ -1939,7 +1939,7 @@ bool Recompiler::Recompile(
         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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.s8, simde_mm_adds_epi8(simde_mm_load_si128((simde__m128i*){}.s8), simde_mm_load_si128((simde__m128i*){}.s8)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
     case PPC_INST_VADDSHS:
@@ -1998,7 +1998,7 @@ bool Recompiler::Recompile(
         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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_avg_epu16(simde_mm_load_si128((simde__m128i*){}.u16), simde_mm_load_si128((simde__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
     case PPC_INST_VCTSXS:
@@ -2014,11 +2014,11 @@ bool Recompiler::Recompile(
     case PPC_INST_VCTUXS:
     case PPC_INST_VCFPUXWS128:
         printSetFlushMode(true);
-        print("\t_mm_store_si128((__m128i*){}.u32, _mm_vctuxs(", v(insn.operands[0]));
+        print("\tsimde_mm_store_si128((simde__m128i*){}.u32, simde_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]);
+            println("simde_mm_mul_ps(simde_mm_load_ps({}.f32), simde_mm_set1_ps({}))));", v(insn.operands[1]), 1u << insn.operands[2]);
         else
-            println("_mm_load_ps({}.f32)));", v(insn.operands[1]));
+            println("simde_mm_load_ps({}.f32)));", v(insn.operands[1]));
         break;
 
     case PPC_INST_VCFSX:
@@ -2075,9 +2075,9 @@ bool Recompiler::Recompile(
         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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_cmpeq_epi16(simde_mm_load_si128((simde__m128i*){}.u16), simde_mm_load_si128((simde__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]));
+            println("\t{}.setFromMask(simde_mm_load_si128((simde__m128i*){}.u16), 0xFFFF);", cr(6), v(insn.operands[0]));
         break;
 
     case PPC_INST_VCMPEQUW:
@@ -2106,25 +2106,25 @@ bool Recompiler::Recompile(
     case PPC_INST_VCMPGTUB:
         println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_cmpgt_epu8(simde_mm_load_si128((simde__m128i*){}.u8), simde_mm_load_si128((simde__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]));
+            println("\t{}.setFromMask(simde_mm_load_si128((simde__m128i*){}.u8), 0xFFFF);", cr(6), v(insn.operands[0]));
         break;
 
     case PPC_INST_VCMPGTUH:
         println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_cmpgt_epu16(simde_mm_load_si128((simde__m128i*){}.u16), simde_mm_load_si128((simde__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]));
+            println("\t{}.setFromMask(simde_mm_load_si128((simde__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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.s8, simde_mm_cmpgt_epi16(simde_mm_load_si128((simde__m128i*){}.u16), simde_mm_load_si128((simde__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]));
+            println("\t{}.setFromMask(simde_mm_load_si128((simde__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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.s8, simde_mm_cmpgt_epi32(simde_mm_load_si128((simde__m128i*){}.u32), simde_mm_load_si128((simde__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]));
+            println("\t{}.setFromMask(simde_mm_load_si128((simde__m128i*){}.s32), 0xFFFF);", cr(6), v(insn.operands[0]));
         break;
 
     case PPC_INST_VEXPTEFP:
@@ -2157,7 +2157,7 @@ bool Recompiler::Recompile(
         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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u16, simde_mm_max_epi16(simde_mm_load_si128((simde__m128i*){}.u16), simde_mm_load_si128((simde__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
     case PPC_INST_VMAXSW:
@@ -2165,7 +2165,7 @@ bool Recompiler::Recompile(
         break;
 
     case PPC_INST_VMINSH:
-        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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u16, simde_mm_min_epi16(simde_mm_load_si128((simde__m128i*){}.u16), simde_mm_load_si128((simde__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
     case PPC_INST_VMINFP:
@@ -2223,12 +2223,12 @@ bool Recompiler::Recompile(
         break;
 
     case PPC_INST_VNOR:
-        print("\t_mm_store_si128((__m128i*){}.u8, ", v(insn.operands[0]));
+        print("\tsimde_mm_store_si128((simde__m128i*){}.u8, ", v(insn.operands[0]));
 
         if (insn.operands[1] != insn.operands[2])
-            println("_mm_andnot_si128(_mm_or_si128(_mm_load_si128((__m128i*){}.u8), _mm_load_si128((__m128i*){}.u8)), _mm_set1_epi8(0xFF)));", v(insn.operands[1]), v(insn.operands[2]));
+            println("simde_mm_andnot_si128(simde_mm_or_si128(simde_mm_load_si128((simde__m128i*){}.u8), simde_mm_load_si128((simde__m128i*){}.u8)), simde_mm_set1_epi8(0xFF)));", v(insn.operands[1]), v(insn.operands[2]));
         else
-            println("_mm_setzero_si128());");
+            println("simde_mm_setzero_si128());");
 
         break;
 
@@ -2288,12 +2288,12 @@ bool Recompiler::Recompile(
 
     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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_packs_epi16(simde_mm_load_si128((simde__m128i*){}.s16), simde_mm_load_si128((simde__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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_packs_epi32(simde_mm_load_si128((simde__m128i*){}.s32), simde_mm_load_si128((simde__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
         break;
 
     case PPC_INST_VPKSHUS:
@@ -2303,7 +2303,7 @@ bool Recompiler::Recompile(
 
     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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_packus_epi32(simde_mm_load_si128((simde__m128i*){}.s32), simde_mm_load_si128((simde__m128i*){}.s32)));", v(insn.operands[0]), v(insn.operands[2]), v(insn.operands[1]));
         break;
 
     case PPC_INST_VPKUHUS:
@@ -2415,7 +2415,7 @@ bool Recompiler::Recompile(
         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]);
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u16, simde_mm_set1_epi16(int(0x{:X})));", v(insn.operands[0]), insn.operands[1]);
         break;
 
     case PPC_INST_VSPLTISW:
@@ -2498,11 +2498,11 @@ bool Recompiler::Recompile(
         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]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_sub_epi8(simde_mm_load_si128((simde__m128i*){}.u8), simde_mm_load_si128((simde__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*){}.u16), _mm_load_si128((__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
+        println("\tsimde_mm_store_si128((simde__m128i*){}.u8, simde_mm_sub_epi16(simde_mm_load_si128((simde__m128i*){}.u16), simde_mm_load_si128((simde__m128i*){}.u16)));", v(insn.operands[0]), v(insn.operands[1]), v(insn.operands[2]));
         break;
 
     case PPC_INST_VUPKD3D128: