2 files changed, 37 insertions, 7 deletions

diff --git a/src/arch/arm/isa.hh b/src/arch/arm/isa.hh
index 60c572833..a3e89b544 100644
--- a/src/arch/arm/isa.hh
+++ b/src/arch/arm/isa.hh
@@ -699,15 +699,28 @@ namespace ArmISA
 }
 
 template<>
-struct initRenameMode<ArmISA::ISA>
+struct RenameMode<ArmISA::ISA>
 {
-    static Enums::VecRegRenameMode mode(const ArmISA::ISA* isa)
+    static Enums::VecRegRenameMode
+    init(const ArmISA::ISA* isa)
     {
         return isa->vecRegRenameMode();
     }
-    static bool equals(const ArmISA::ISA* isa1, const ArmISA::ISA* isa2)
+
+    static Enums::VecRegRenameMode
+    mode(const ArmISA::PCState& pc)
+    {
+        if (pc.aarch64()) {
+            return Enums::Full;
+        } else {
+            return Enums::Elem;
+        }
+    }
+
+    static bool
+    equalsInit(const ArmISA::ISA* isa1, const ArmISA::ISA* isa2)
     {
-        return mode(isa1) == mode(isa2);
+        return init(isa1) == init(isa2);
     }
 };
 
diff --git a/src/arch/arm/utility.cc b/src/arch/arm/utility.cc
index 175fd7c00..11c3479c6 100644
--- a/src/arch/arm/utility.cc
+++ b/src/arch/arm/utility.cc
@@ -145,6 +145,24 @@ skipFunction(ThreadContext *tc)
     }
 }
 
+static void
+copyVecRegs(ThreadContext *src, ThreadContext *dest)
+{
+    auto src_mode = RenameMode<ArmISA::ISA>::mode(src->pcState());
+
+    // The way vector registers are copied (VecReg vs VecElem) is relevant
+    // in the O3 model only.
+    if (src_mode == Enums::Full) {
+        for (auto idx = 0; idx < NumVecRegs; idx++)
+            dest->setVecRegFlat(idx, src->readVecRegFlat(idx));
+    } else {
+        for (auto idx = 0; idx < NumVecRegs; idx++)
+            for (auto elem_idx = 0; elem_idx < NumVecElemPerVecReg; elem_idx++)
+                dest->setVecElemFlat(
+                    idx, elem_idx, src->readVecElemFlat(idx, elem_idx));
+    }
+}
+
 void
 copyRegs(ThreadContext *src, ThreadContext *dest)
 {
@@ -154,15 +172,14 @@ copyRegs(ThreadContext *src, ThreadContext *dest)
     for (int i = 0; i < NumFloatRegs; i++)
         dest->setFloatRegBitsFlat(i, src->readFloatRegBitsFlat(i));
 
-    for (int i = 0; i < NumVecRegs; i++)
-        dest->setVecRegFlat(i, src->readVecRegFlat(i));
-
     for (int i = 0; i < NumCCRegs; i++)
         dest->setCCReg(i, src->readCCReg(i));
 
     for (int i = 0; i < NumMiscRegs; i++)
         dest->setMiscRegNoEffect(i, src->readMiscRegNoEffect(i));
 
+    copyVecRegs(src, dest);
+
     // setMiscReg "with effect" will set the misc register mapping correctly.
     // e.g. updateRegMap(val)
     dest->setMiscReg(MISCREG_CPSR, src->readMiscRegNoEffect(MISCREG_CPSR));