Espresso 0.0.1b

boot.s

@@ -74,6 +74,8 @@ enable_sse_asm:
     push %ebx
     push %ecx
     push %edx
+    push %esi
+    push %edi
 
     # Check CPUID support
     pushf
@@ -87,13 +89,16 @@ enable_sse_asm:
     xor %ecx, %eax
     jz .no_cpuid
 
-    # Check for SSE
+    # CPUID function 1
     mov $1, %eax
     cpuid
-    test $0x02000000, %edx
+    mov %edx, %ebx        # EDX = SSE1/SSE2 bits
+    mov %ecx, %esi        # ECX = SSE3/SSSE3/SSE4.1 bits
+
+    test $0x02000000, %ebx   # SSE
     jz .no_sse
 
-    # Enable SSE
+    # Enable SSE (required for SSE1/2/3/SSSE3/4.1)
     mov %cr0, %eax
     and $~0x4, %eax       # Clear EM (bit 2)
     or  $0x2, %eax        # Set MP (bit 1)
@@ -102,17 +107,48 @@ enable_sse_asm:
     mov %cr4, %eax
     or  $0x600, %eax      # Set OSFXSR | OSXMMEXCPT
     mov %eax, %cr4
-    
-    lea sse_initialized, %ebx
-    movl $1, (%ebx)
+
+    # Set version = 1 (SSE1)
+    mov $1, %eax
+
+    test $0x04000000, %ebx     # SSE2 (bit 26)
+    jz .check_sse3
+    mov $2, %eax
+
+.check_sse3:
+    test $0x00000001, %esi     # SSE3 (bit 0)
+    jz .check_ssse3
+    mov $3, %eax
+
+.check_ssse3:
+    test $0x00000200, %esi     # SSSE3 (bit 9)
+    jz .check_sse41
+    mov $4, %eax
+
+.check_sse41:
+    test $0x00080000, %esi     # SSE4.1 (bit 19)
+    jz .set_result
+    mov $5, %eax
+
+.set_result:
+    lea sse_initialized, %edi
+    mov %eax, (%edi)
+    jmp .done
 
 .no_sse:
 .no_cpuid:
+    lea sse_initialized, %edi
+    mov $0, (%edi)
+
+.done:
+    pop %edi
+    pop %esi
     pop %edx
     pop %ecx
     pop %ebx
     pop %eax
     ret
+
         
 _kernel_early:
         call _init
@@ -205,8 +241,9 @@ This is useful when debugging or when you implement call tracing.
 */
 .size _start, . - _start
 
+
 .section .data
 
 .global sse_initialized
 
-sse_initialized: .word 0
+sse_initialized: .int 0