[MODERN FEATURES] SSE

kernel/arch/sse.S

@@ -0,0 +1,45 @@
+#include <registers.h>
+
+.intel_syntax noprefix
+
+.global sse_init
+sse_init:
+  mov rax, cr0
+  and ax, 0xFFFF-CR0_EM
+  or ax, CR0_MP
+  mov cr0, rax
+
+  mov rax, cr4
+  or ax, CR4_OSFXSR
+  or ax, CR4_OSXMMEXCPT
+  mov cr4, rax
+
+  ret
+
+.global sse_save
+sse_save:
+  movabs rax, offset _sse_buffer
+  fxsave [rax]
+  movabs rsi, offset _sse_buffer
+  mov rdx, 512
+  call memcpy
+  ret
+
+.global sse_restore
+sse_restore:
+  mov rsi, rdi
+  movabs rdi, offset _sse_buffer
+  mov rdx, 512
+  call memcpy
+  movabs rax, offset _sse_buffer
+  fxrstor [rax]
+  ret
+
+
+.section .data
+.align 16
+_sse_buffer:
+ .rept 512
+ .byte 0
+ .endr
+

kernel/boot/kmain.c

@@ -9,6 +9,7 @@
 #include <cpuid.h>
 #include <acpi.h>
 #include <apic.h>
+#include <sse.h>
 
 void thread_function()
 {
@@ -36,6 +37,7 @@ int kmain(uint64_t multiboot_magic, void *multiboot_data)
   acpi_init();
   apic_init();
   ioapic_init();
+  sse_init();
 
 
   process_t *p1 = process_spawn(0);

kernel/include/registers.h

@@ -1,8 +1,16 @@
 #pragma once
 
+#ifndef __ASSEMBLER__
 void load_idt(void *);
 uint64_t read_cr0();
 uint64_t read_cr2();
 uint64_t read_cr3();
 void write_cr3(uint64_t);
 uint64_t read_cr4();
+#endif
+
+#define CR0_MP (1<<1)
+#define CR0_EM (1<<2)
+
+#define CR4_OSFXSR (1<<9)
+#define CR4_OSXMMEXCPT (1<<10)

kernel/include/sse.h

@@ -0,0 +1,6 @@
+#pragma once
+
+void sse_init();
+
+void sse_save(uint8_t *addr);
+void sse_restore(uint8_t *addr);

kernel/include/thread.h

@@ -11,6 +11,7 @@ typedef struct thread_st
   uint64_t stack_pointer; // Top of the kernel stack for thread
   uint64_t tid;
   uint64_t state;
+  uint8_t sse_registers[512];
   process_t *process;
   LIST(struct thread_st, process_threads);
   LIST(struct thread_st, ready_queue);

kernel/interrupts/ioapic.c

@@ -86,6 +86,4 @@ void ioapic_init()
   vmm_set_page(0, (uintptr_t)P2V(IOAPIC_BASE), IOAPIC_BASE, PAGE_PRESENT | PAGE_WRITE);
   for(int i = 0; i < 24; i++)
     ioapic_route(i, IOAPIC_DEST_LOGICAL | IOAPIC_DELIVERY_LOWEST, 0xFF, INT_IRQ0+i);
-
-  IRQ_UNMASK(IRQ_KEYBOARD);
 }

kernel/proc/scheduler.c

@@ -1,6 +1,7 @@
 #include <scheduler.h>
 #include <list.h>
 #include <debug.h>
+#include <sse.h>
 
 LIST(thread_t, ready_queue);
 
@@ -60,6 +61,7 @@ void scheduler()
     new->state = THREAD_STATE_RUNNING;
     set_last_thread(new);
     switch_process(new->process);
+    sse_restore(new->sse_registers);
     switch_thread(scheduler_th, new);
   }
 }
@@ -70,6 +72,11 @@ void schedule()
 
   thread_t *old = get_current_thread();
 
+  if(old)
+  {
+    sse_save(old->sse_registers);
+  }
+
   switch_process(0);
   switch_thread(old, scheduler_th);
 }