WIP

FEX-Emu · Mar 1, 2024 · 5067372 · 5067372
1 parent e87ca02
commit 5067372
Showing 1 changed file with 28 additions and 14 deletions.
diff --git a/FEXCore/Source/Interface/Core/JIT/Arm64/MemoryOps.cpp b/FEXCore/Source/Interface/Core/JIT/Arm64/MemoryOps.cpp
@@ -1718,7 +1718,7 @@ DEF_OP(MemSet) {
   ARMEmitter::SingleUseForwardLabel BackwardImpl{};
   ARMEmitter::SingleUseForwardLabel Done{};
 
-  mov(TMP1, Length.X());
+  mov(Size == 8 ? ARMEmitter::Size::i64Bit : ARMEmitter::Size::i32Bit, TMP1, Length);
   if (Op->Prefix.IsInvalid()) {
     mov(TMP2, MemReg.X());
   }
@@ -1810,27 +1810,34 @@ DEF_OP(MemSet) {
       and_(ARMEmitter::Size::i64Bit, TMP4, TMP2, 0x3);
       cbnz(ARMEmitter::Size::i64Bit, TMP4, &AgainInternal);
 
+      // Keep the counter one copy ahead, so that underflow can be used to detect when to fallback
+      // to the copy unit size copy loop for the last chunk.
+      // Do this in two parts, to fallback to the byte by byte loop if size < 32, and to the
+      // single copy loop if size < 64.
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
+      tbnz(TMP1, 63, &AgainInternal128Exit);
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
+      tbnz(TMP1, 63, &AgainInternal256Exit);
+
       // Fill VTMP2 with the set pattern
       dup(SubRegSize, VTMP2.Q(), Value);
 
       Bind(&AgainInternal256);
-      // Keep the counter one copy ahead, so that underflow can be used to detect when to fallback
-      // to the copy unit size copy loop for the last chunk.
-      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 64 / Size);
-      tbnz(TMP1, 63, &AgainInternal256Exit);
       stp<ARMEmitter::IndexType::POST>(VTMP2.Q(), VTMP2.Q(), TMP2, 32 * Direction);
       stp<ARMEmitter::IndexType::POST>(VTMP2.Q(), VTMP2.Q(), TMP2, 32 * Direction);
-      b(&AgainInternal256);
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 64 / Size);
+      tbz(TMP1, 63, &AgainInternal256);
 
       Bind(&AgainInternal256Exit);
       add(ARMEmitter::Size::i64Bit, TMP1, TMP1, 64 / Size);
       cbz(ARMEmitter::Size::i64Bit, TMP1, &DoneInternal);
 
-      Bind(&AgainInternal128);
       sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
       tbnz(TMP1, 63, &AgainInternal128Exit);
+      Bind(&AgainInternal128);
       stp<ARMEmitter::IndexType::POST>(VTMP2.Q(), VTMP2.Q(), TMP2, 32 * Direction);
-      b(&AgainInternal128);
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
+      tbz(TMP1, 63, &AgainInternal128);
 
       Bind(&AgainInternal128Exit);
       add(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
@@ -1941,7 +1948,7 @@ DEF_OP(MemCpy) {
   ARMEmitter::SingleUseForwardLabel BackwardImpl{};
   ARMEmitter::SingleUseForwardLabel Done{};
 
-  mov(TMP1, Length.X());
+  mov(Size == 8 ? ARMEmitter::Size::i64Bit : ARMEmitter::Size::i32Bit, TMP1, Length);
   if (Op->PrefixDest.IsInvalid()) {
     mov(TMP2, MemRegDest.X());
   }
@@ -2113,24 +2120,31 @@ DEF_OP(MemCpy) {
       and_(ARMEmitter::Size::i64Bit, TMP4, TMP4, 0x3);
       cbnz(ARMEmitter::Size::i64Bit, TMP4, &AgainInternal);
 
-      Bind(&AgainInternal256);
       // Keep the counter one copy ahead, so that underflow can be used to detect when to fallback
       // to the copy unit size copy loop for the last chunk.
-      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 64 / Size);
+      // Do this in two parts, to fallback to the byte by byte loop if size < 32, and to the
+      // single copy loop if size < 64.
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
+      tbnz(TMP1, 63, &AgainInternal128Exit);
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
       tbnz(TMP1, 63, &AgainInternal256Exit);
+
+      Bind(&AgainInternal256);
       MemCpy(32, 32 * Direction);
       MemCpy(32, 32 * Direction);
-      b(&AgainInternal256);
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 64 / Size);
+      tbz(TMP1, 63, &AgainInternal256);
 
       Bind(&AgainInternal256Exit);
       add(ARMEmitter::Size::i64Bit, TMP1, TMP1, 64 / Size);
       cbz(ARMEmitter::Size::i64Bit, TMP1, &DoneInternal);
 
-      Bind(&AgainInternal128);
       sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
       tbnz(TMP1, 63, &AgainInternal128Exit);
+      Bind(&AgainInternal128);
       MemCpy(32, 32 * Direction);
-      b(&AgainInternal128);
+      sub(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);
+      tbz(TMP1, 63, &AgainInternal128);
 
       Bind(&AgainInternal128Exit);
       add(ARMEmitter::Size::i64Bit, TMP1, TMP1, 32 / Size);