Skip to contentSkip to conversation

x86: LLVMLinux: Reimplement current_stack_pointer without register usage.

PATCH In review [email protected] 3 participants 4 messages Feb 21, 2014 – Feb 26, 2014 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86: LLVMLinux: Reimplement current_stack_pointer without register usage.

From: Behan Webster <[email protected]>

Use asm to make the globally named register work again for gcc and clang. Much more efficient than copying the stack pointer to a variable and back again.

---
 arch/x86/include/asm/thread_info.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
arch/x86/include/asm/thread_info.h +4 -4
index e1940c0..e27ccc1 100644
@@ -163,10 +163,10 @@ struct thread_info {
  */
 #ifndef __ASSEMBLY__
 
-#define current_stack_pointer ({		\
-	unsigned long sp;			\
-	asm("mov %%esp,%0" : "=g" (sp));	\
-	sp;					\
+#define current_stack_pointer ({			\
+	register unsigned long sp asm("esp") __used;	\
+	asm("" : "=r" (sp));				\
+	sp;						\
 })
 
 /* how to get the thread information struct from C */
Signature · 1 line
1.8.3.2
#1 lore raw

This seems like really deep magic when looking at it... at the very least, this needs to be very carefully commented, including why it works on the various platforms.

How much does this actually affect the output? I only see three uses of current_stack_pointer:

/* how to get the thread information struct from C */ static inline struct thread_info *current_thread_info(void) {
return (struct thread_info *)
(current_stack_pointer & ~(THREAD_SIZE - 1));
}

... here we need the mov anyway, because we have to then AND it with a mask, which we obviously can't do inside the stack pointer.

kernel/irq_32.c: irqctx->tinfo.previous_esp = current_stack_pointer;

(two times)

Here we are moving it into a memory variable anyway, which the "=g" constraint should allow.

So I see no evidence this is more efficient in any way.

-hpa

Quoted 30 lines · On 02/20/2014 08:44 PM, [email protected] wrote:

From: Behan Webster <[email protected]>

Use asm to make the globally named register work again for gcc and clang. Much more efficient than copying the stack pointer to a variable and back again.

---
  arch/x86/include/asm/thread_info.h | 8 ++++----
  1 file changed, 4 insertions(+), 4 deletions(-)
arch/x86/include/asm/thread_info.h +4 -4
index e1940c0..e27ccc1 100644
@@ -163,10 +163,10 @@ struct thread_info {
   */
  #ifndef __ASSEMBLY__
 
-#define current_stack_pointer ({		\
-	unsigned long sp;			\
-	asm("mov %%esp,%0" : "=g" (sp));	\
-	sp;					\
+#define current_stack_pointer ({			\
+	register unsigned long sp asm("esp") __used;	\
+	asm("" : "=r" (sp));				\
+	sp;						\
  })
 
  /* how to get the thread information struct from C */
#2 lore raw
Quoted 14 lines · H. Peter Anvin

This seems like really deep magic when looking at it... at the very least, this needs to be very carefully commented, including why it works on the various platforms.

How much does this actually affect the output? I only see three uses of current_stack_pointer:

/* how to get the thread information struct from C */ static inline struct thread_info *current_thread_info(void) {
return (struct thread_info *)
(current_stack_pointer & ~(THREAD_SIZE - 1));
}

Quoted 2 lines

... here we need the mov anyway, because we have to then AND it with a mask, which we obviously can't do inside the stack pointer.

No clue what code is actually generated, but the new code could generate:

mov $MASK, %rax;
and %esp, %rax;

Admittedly, I can't see any reason why this would be an improvement.

--Andy

#3 lore raw
Quoted 14 lines · Andy Lutomirski
Quoted 11 lines

How much does this actually affect the output? I only see three uses of current_stack_pointer:

/* how to get the thread information struct from C */ static inline struct thread_info *current_thread_info(void) {
return (struct thread_info *)
(current_stack_pointer & ~(THREAD_SIZE - 1));
}

Quoted 2 lines

... here we need the mov anyway, because we have to then AND it with a mask, which we obviously can't do inside the stack pointer.

Quoted 7 lines

No clue what code is actually generated, but the new code could generate:

mov $MASK, %rax;
and %esp, %rax;

Admittedly, I can't see any reason why this would be an improvement.

You have to generate one of the code sequences:

mov $MASK, %eax
and %esp, %eax

... or ...

mov %esp, %eax
and $MASK, %eax

No real difference either way.

-hpa