From eecd016ba8ce75baf33e0a7f9f225a8b0db769ea Mon Sep 17 00:00:00 2001
From: Lioncache <mai.iam2048@gmail.com>
Date: Mon, 27 Mar 2023 15:17:59 -0400
Subject: [PATCH] Arm64/VectorOps: Eliminate unnecessary EOR and MOV in
 VCMP{EQ,GT}

We can use the zeroing version of MOVPRFX to perform the same behavior.
---
 .../Interface/Core/JIT/Arm64/VectorOps.cpp       | 16 ++++------------
 1 file changed, 4 insertions(+), 12 deletions(-)

diff --git a/External/FEXCore/Source/Interface/Core/JIT/Arm64/VectorOps.cpp b/External/FEXCore/Source/Interface/Core/JIT/Arm64/VectorOps.cpp
index 2189b9012..ba53dc343 100644
--- a/External/FEXCore/Source/Interface/Core/JIT/Arm64/VectorOps.cpp
+++ b/External/FEXCore/Source/Interface/Core/JIT/Arm64/VectorOps.cpp
@@ -1531,17 +1531,13 @@ DEF_OP(VCMPEQ) {
     const auto Mask = PRED_TMP_32B.Zeroing();
     const auto ComparePred = ARMEmitter::PReg::p0;
 
-    // Ensure no junk is in the temp (important for ensuring
-    // non-equal entries remain as zero during the final bitwise OR).
-    eor(VTMP1.Z(), VTMP1.Z(), VTMP1.Z());
-
     // General idea is to compare for equality, not the equal vals
     // from one of the registers, then or both together to make the
     // relevant equal entries all 1s.
     cmpeq(SubRegSize.Vector, ComparePred, Mask, Vector1.Z(), Vector2.Z());
     not_(SubRegSize.Vector, VTMP1.Z(), ComparePred.Merging(), Vector1.Z());
-    orr(SubRegSize.Vector, VTMP1.Z(), ComparePred.Merging(), VTMP1.Z(), Vector1.Z());
-    mov(Dst.Z(), VTMP1.Z());
+    movprfx(SubRegSize.Vector, Dst.Z(), ComparePred.Zeroing(), Vector1.Z());
+    orr(SubRegSize.Vector, Dst.Z(), ComparePred.Merging(), Dst.Z(), VTMP1.Z());
   } else {
     if (IsScalar) {
       cmeq(SubRegSize.Scalar, Dst, Vector1, Vector2);
@@ -1615,17 +1611,13 @@ DEF_OP(VCMPGT) {
     const auto Mask = PRED_TMP_32B.Zeroing();
     const auto ComparePred = ARMEmitter::PReg::p0;
 
-    // Ensure no junk is in the temp (important for ensuring
-    // non greater-than values remain as zero).
-    eor(VTMP1.Z(), VTMP1.Z(), VTMP1.Z());
-
     // General idea is to compare for greater-than, bitwise NOT
     // the valid values, then ORR the NOTed values with the original
     // values to form entries that are all 1s.
     cmpgt(SubRegSize.Vector, ComparePred, Mask, Vector1.Z(), Vector2.Z());
     not_(SubRegSize.Vector, VTMP1.Z(), ComparePred.Merging(), Vector1.Z());
-    orr(SubRegSize.Vector, VTMP1.Z(), ComparePred.Merging(), VTMP1.Z(), Vector1.Z());
-    mov(Dst.Z(), VTMP1.Z());
+    movprfx(SubRegSize.Vector, Dst.Z(), ComparePred.Zeroing(), Vector1.Z());
+    orr(SubRegSize.Vector, Dst.Z(), ComparePred.Merging(), Dst.Z(), VTMP1.Z());
   } else {
     if (IsScalar) {
       cmgt(SubRegSize.Scalar, Dst, Vector1, Vector2);