From e9f2cc359521b7be6a6a7655c525bbd796135362 Mon Sep 17 00:00:00 2001
From: Andreas Kling <awesomekling@gmail.com>
Date: Wed, 27 Mar 2019 15:27:45 +0100
Subject: Kernel: Save/restore the SSE context on context switch.

---
 Kernel/Thread.cpp |  6 ++++--
 Kernel/Thread.h   |  4 ++--
 Kernel/i386.cpp   |  4 ++--
 Kernel/i386.h     | 11 ++---------
 4 files changed, 10 insertions(+), 15 deletions(-)

diff --git a/Kernel/Thread.cpp b/Kernel/Thread.cpp
index ef1a6ec5af..0e649106b3 100644
--- a/Kernel/Thread.cpp
+++ b/Kernel/Thread.cpp
@@ -15,7 +15,7 @@ Thread::Thread(Process& process)
 {
     dbgprintf("Thread: New thread TID=%u in %s(%u)\n", m_tid, process.name().characters(), process.pid());
     set_default_signal_dispositions();
-    memset(&m_fpu_state, 0, sizeof(FPUState));
+    m_fpu_state = (FPUState*)kmalloc_aligned(sizeof(FPUState), 16);
     memset(&m_tss, 0, sizeof(m_tss));
 
     // Only IF is set when a process boots.
@@ -68,6 +68,7 @@ Thread::Thread(Process& process)
 Thread::~Thread()
 {
     dbgprintf("~Thread{%p}\n", this);
+    kfree_aligned(m_fpu_state);
     {
         InterruptDisabler disabler;
         g_threads->remove(this);
@@ -493,7 +494,8 @@ Thread* Thread::clone(Process& process)
     auto* clone = new Thread(process);
     memcpy(clone->m_signal_action_data, m_signal_action_data, sizeof(m_signal_action_data));
     clone->m_signal_mask = m_signal_mask;
-    clone->m_fpu_state = m_fpu_state;
+    clone->m_fpu_state = (FPUState*)kmalloc_aligned(sizeof(FPUState), 16);
+    memcpy(clone->m_fpu_state, m_fpu_state, sizeof(FPUState));
     clone->m_has_used_fpu = m_has_used_fpu;
     return clone;
 }
diff --git a/Kernel/Thread.h b/Kernel/Thread.h
index abfb99251e..fec389c8b6 100644
--- a/Kernel/Thread.h
+++ b/Kernel/Thread.h
@@ -111,7 +111,7 @@ public:
     bool has_unmasked_pending_signals() const;
     void terminate_due_to_signal(byte signal);
 
-    FPUState& fpu_state() { return m_fpu_state; }
+    FPUState& fpu_state() { return *m_fpu_state; }
     bool has_used_fpu() const { return m_has_used_fpu; }
     void set_has_used_fpu(bool b) { m_has_used_fpu = b; }
 
@@ -159,7 +159,7 @@ private:
     Vector<int> m_select_write_fds;
     Vector<int> m_select_exceptional_fds;
     State m_state { Invalid };
-    FPUState m_fpu_state;
+    FPUState* m_fpu_state { nullptr };
     bool m_select_has_timeout { false };
     bool m_has_used_fpu { false };
     bool m_was_interrupted_while_blocked { false };
diff --git a/Kernel/i386.cpp b/Kernel/i386.cpp
index 1879ba9206..efdc3850d9 100644
--- a/Kernel/i386.cpp
+++ b/Kernel/i386.cpp
@@ -190,14 +190,14 @@ void exception_7_handler(RegisterDump& regs)
     if (g_last_fpu_thread == current)
         return;
     if (g_last_fpu_thread) {
-        asm volatile("fnsave %0":"=m"(g_last_fpu_thread->fpu_state()));
+        asm volatile("fxsave %0":"=m"(g_last_fpu_thread->fpu_state()));
     } else {
         asm volatile("fnclex");
     }
     g_last_fpu_thread = current;
 
     if (current->has_used_fpu()) {
-        asm volatile("frstor %0"::"m"(current->fpu_state()));
+        asm volatile("fxrstor %0"::"m"(current->fpu_state()));
     } else {
         asm volatile("fninit");
         current->set_has_used_fpu(true);
diff --git a/Kernel/i386.h b/Kernel/i386.h
index fcf70cf7cc..9d0ec625ad 100644
--- a/Kernel/i386.h
+++ b/Kernel/i386.h
@@ -234,15 +234,8 @@ struct [[gnu::packed]] RegisterDumpWithExceptionCode {
     word ss_if_crossRing;
 };
 
-struct FPUState {
-    dword cwd;
-    dword swd;
-    dword twd;
-    dword fip;
-    dword fcs;
-    dword foo;
-    dword fos;
-    dword st[20];
+struct [[gnu::aligned(16)]] FPUState {
+    byte buffer[512];
 };
 
 inline constexpr dword page_base_of(dword address)
-- 
cgit v1.2.3