arm: LLVMLinux: Add current_stack_pointer
Subsystems (2)
cover arm: LLVMLinux: Add current_stack_pointer
From: Behan Webster <[email protected]>
The LLVMLinux Project is working to be able to build the Linux kernel with clang/LLVM. With the release of LLVM 3.3 clang is now able to compile the Linux kernel with a number of small patches (available from the LLVMLinux git repo).
These patches add a macro to get the current stack pointer which allows for a single place in which to do so with ASM. Before this named registers (a gcc extension) was used to get the stack pointer. Using ASM is a more portable way of getting the stack pointer which works with both gcc and clang. This macro is of the same name used in the X86 arch.
Behan Webster (5):
arm: LLVMLinux: Add current_stack_pointer macro for ARM
arm: LLVMLinux: use current_stack_pointer for percpu
arm: LLVMLinux: Use current_stack_pointer for return_address
arm: LLVMLinux: Use current_stack_pointer in save_stack_trace_tsk
arm: LLVMLinux: Use current_stack_pointer in unwind_backtrace
arch/arm/include/asm/percpu.h | 4 ++-- arch/arm/include/asm/thread_info.h | 9 +++++++++ arch/arm/kernel/return_address.c | 3 +-- arch/arm/kernel/stacktrace.c | 4 +--- arch/arm/kernel/unwind.c | 3 +-- 5 files changed, 14 insertions(+), 9 deletions(-)
Signature · 1 line
1.8.1.2
PATCH 1/5 arm: LLVMLinux: Add current_stack_pointer macro for ARM
From: Behan Webster <[email protected]>
A macro to get the current stack pointer which allows for a single place in which to do so with ASM. Before this named registers (a gcc extension) was used to get the stack pointer. Using ASM is a more portable way of getting the stack pointer which works with both gcc and clang. This macro is of the same name used in the X86 arch.
Author: Behan Webster <[email protected]>
--- arch/arm/include/asm/thread_info.h | 9 +++++++++ 1 file changed, 9 insertions(+)
arch/arm/include/asm/thread_info.h +9 -0
@@ -100,6 +100,15 @@ struct thread_info {#define init_stack (init_thread_union.stack)/* +* how to get the current stack pointer from C +*/ +#define current_stack_pointer ({ \ + unsigned long current_sp; \ + asm ("mov %0, r13" : "=r" (current_sp)); \ + current_sp; \ +}) + +/* * how to get the thread information struct from C */ static inline struct thread_info *current_thread_info (void ) __attribute_const__;
Signature · 1 line
1.8.1.2
Re: [PATCH 1/5] arm: LLVMLinux: Add current_stack_pointer macro for ARM
arch/arm/include/asm/thread_info.h struct thread_info { · line 109
Quoted 5 lines · behanw at converseincode.com
arch/arm/include/asm/thread_info.h +5
-0+#define current_stack_pointer ({ \ + unsigned long current_sp; \ + asm ("mov %0, r13" : "=r" (current_sp)); \ + current_sp; \ +})
Why do you use 'r13' rather than the more common 'sp' alias?
Signature · 2 lines
M?ns Rullg?rd mans at mansr.com
PATCH 1/5 arm: LLVMLinux: Add current_stack_pointer macro for ARM
Quoted 8 lines · M?ns Rullg?rd
arch/arm/include/asm/thread_info.h struct thread_info { · line 109
Quoted 5 lines · behanw at converseincode.com
arch/arm/include/asm/thread_info.h +5
-0+#define current_stack_pointer ({ \ + unsigned long current_sp; \ + asm ("mov %0, r13" : "=r" (current_sp)); \ + current_sp; \ +})Why do you use 'r13' rather than the more common 'sp' alias?
Originally we were using LLVM's Integrated Assembler (IA), which didn't allow for that alias if I remember correctly. However, now we're using gas because IA only supports Unified Assembly Language grammar, and not the extensions that are common in the kernel code.
I can resubmit using that alias (after the rest of the discussion).
Behan
Signature · 2 lines
Behan Webster behanw at converseincode.com
Re: [PATCH 1/5] arm: LLVMLinux: Add current_stack_pointer macro for ARM
In reply to [email protected]Quoted 7 lines · behanw at converseincode.com
From: Behan Webster <[email protected]>
A macro to get the current stack pointer which allows for a single place in which to do so with ASM. Before this named registers (a gcc extension) was used to get the stack pointer. Using ASM is a more portable way of getting the stack pointer which works with both gcc and clang. This macro is of the same name used in the X86 arch.
This will result in less optimal code - rather than the compiler being able to mask directly with 'sp', it's going to have to use this bit of assembly to first move it into another register.
Why do we want this change?
PATCH 1/5 arm: LLVMLinux: Add current_stack_pointer macro for ARM
Quoted 11 lines · Russell King - ARM Linux
Quoted 7 lines · behanw at converseincode.com
From: Behan Webster <[email protected]>
A macro to get the current stack pointer which allows for a single place in which to do so with ASM. Before this named registers (a gcc extension) was used to get the stack pointer. Using ASM is a more portable way of getting the stack pointer which works with both gcc and clang. This macro is of the same name used in the X86 arch.
This will result in less optimal code - rather than the compiler being able to mask directly with 'sp', it's going to have to use this bit of assembly to first move it into another register.
I understand. The issue is that clang doesn't support naming registers like this. It's a gcc-ism.
I'm not entirely happy with this solution either, but it was what we could get to work for both compilers without the use of ifdefs.
Though also not ideal, how about an #ifdef for clang to do it this way, otherwise do it with named registers for gcc? Would that be acceptable?
Thanks,
Behan
Signature · 2 lines
Behan Webster behanw at converseincode.com
PATCH 2/5 arm: LLVMLinux: use current_stack_pointer for percpu
In reply to [email protected]From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new current_stack_pointer macro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.
--- arch/arm/include/asm/percpu.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
arch/arm/include/asm/percpu.h +2 -2
@@ -30,14 +30,14 @@ static inline void set_my_cpu_offset(unsigned long off)static inline unsigned long __my_cpu_offset (void ) {unsigned long off; - registerunsigned long *sp asm ("sp" ); +unsigned long sp = current_stack_pointer;/* * Read TPIDRPRW. * We want to allow caching the value, so avoid using volatile and * instead use a fake stack read to hazard against barrier(). */ - asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (*sp)); + asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (sp));return off; }
Signature · 1 line
1.8.1.2
Re: [PATCH 2/5] arm: LLVMLinux: use current_stack_pointer for percpu
Quoted 22 lines · behanw at converseincode.com
From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new
current_stack_pointermacro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.Signed-off-by Mark Charlebois <[email protected]>Signed-off-by Behan Webster <[email protected]>Reviewed-by Jan-Simon M?ller <[email protected]>--- arch/arm/include/asm/percpu.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/arm/include/asm/percpu.h +0
-0@@ -30,14 +30,14 @@ static inline void set_my_cpu_offset(unsigned long off)static inline unsigned long __my_cpu_offset (void ) {unsigned long off;
arch/arm/include/asm/percpu.h static inline void set_my_cpu_offset(unsigned long off) · line 40
Quoted 10 lines
arch/arm/include/asm/percpu.h +2
-2- registerunsigned long *sp asm ("sp" ); +unsigned long sp = current_stack_pointer;/* * Read TPIDRPRW. * We want to allow caching the value, so avoid using volatile and * instead use a fake stack read to hazard against barrier(). */ - asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (*sp)); + asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (sp));
This looks like it's breaking what's going on here. With the original code, we're passing the contents of the word at the stack pointer into the assembly via a "Q" constraint. After this change, we're passing the _value_ of the stack pointer.
Also, if you read the comment, it's certainly wrong.
PATCH 2/5 arm: LLVMLinux: use current_stack_pointer for percpu
Quoted 38 lines · Russell King - ARM Linux
Quoted 22 lines · behanw at converseincode.com
From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new
current_stack_pointermacro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.Signed-off-by Mark Charlebois <[email protected]>Signed-off-by Behan Webster <[email protected]>Reviewed-by Jan-Simon M?ller <[email protected]>--- arch/arm/include/asm/percpu.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/arm/include/asm/percpu.h +0
-0@@ -30,14 +30,14 @@ static inline void set_my_cpu_offset(unsigned long off)static inline unsigned long __my_cpu_offset (void ) {unsigned long off;arch/arm/include/asm/percpu.h static inline void set_my_cpu_offset(unsigned long off) · line 40
Quoted 10 lines
arch/arm/include/asm/percpu.h +2
-2- registerunsigned long *sp asm ("sp" ); +unsigned long sp = current_stack_pointer;/* * Read TPIDRPRW. * We want to allow caching the value, so avoid using volatile and * instead use a fake stack read to hazard against barrier(). */ - asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (*sp)); + asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (sp));This looks like it's breaking what's going on here. With the original code, we're passing the contents of the word at the stack pointer into the assembly via a "Q" constraint. After this change, we're passing the _value_ of the stack pointer.
Quoted 1 line
Also, if you read the comment, it's certainly wrong.
This code was rewritten a few times trying to remove the extra copy. I think this bug crept in.
Of course you're right. I will fix it.
Thanks,
Behan
Signature · 2 lines
Behan Webster behanw at converseincode.com
Re: [PATCH 2/5] arm: LLVMLinux: use current_stack_pointer for percpu
In reply to [email protected]Quoted 22 lines · behanw at converseincode.com
From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new
current_stack_pointermacro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.Signed-off-by Mark Charlebois <[email protected]>Signed-off-by Behan Webster <[email protected]>Reviewed-by Jan-Simon M?ller <[email protected]>--- arch/arm/include/asm/percpu.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/arm/include/asm/percpu.h +0
-0@@ -30,14 +30,14 @@ static inline void set_my_cpu_offset(unsigned long off)static inline unsigned long __my_cpu_offset (void ) {unsigned long off;
arch/arm/include/asm/percpu.h static inline void set_my_cpu_offset(unsigned long off) · line 40
Quoted 10 lines
arch/arm/include/asm/percpu.h +2
-2- registerunsigned long *sp asm ("sp" ); +unsigned long sp = current_stack_pointer;/* * Read TPIDRPRW. * We want to allow caching the value, so avoid using volatile and * instead use a fake stack read to hazard against barrier(). */ - asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (*sp)); + asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (sp));
This doesn't do quite the same thing. The existing code pretends to read something from the stack in order to create a barrier of some sort. Your new code stores the value of the stack pointer to a location on the stack for consumption by the "Q" memory constraint. This store is not necessary and should preferably be avoided.
Signature · 2 lines
M?ns Rullg?rd mans at mansr.com
PATCH 2/5 arm: LLVMLinux: use current_stack_pointer for percpu
Quoted 27 lines · M?ns Rullg?rd
Quoted 25 lines · behanw at converseincode.com
From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new
current_stack_pointermacro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.Signed-off-by Mark Charlebois <[email protected]>Signed-off-by Behan Webster <[email protected]>Reviewed-by Jan-Simon M?ller <[email protected]>--- arch/arm/include/asm/percpu.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/arm/include/asm/percpu.h +1
-1@@ -30,14 +30,14 @@ static inline void set_my_cpu_offset(unsigned long off)static inline unsigned long __my_cpu_offset (void ) {unsigned long off; - registerunsigned long *sp asm ("sp" ); +unsigned long sp = current_stack_pointer;
Quoted 11 lines
arch/arm/include/asm/percpu.h static inline void set_my_cpu_offset(unsigned long off) · line 40
Quoted 7 lines
arch/arm/include/asm/percpu.h +1
-1/* * Read TPIDRPRW. * We want to allow caching the value, so avoid using volatile and * instead use a fake stack read to hazard against barrier(). */ - asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (*sp)); + asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (sp));This doesn't do quite the same thing. The existing code pretends to read something from the stack in order to create a barrier of some sort. Your new code stores the value of the stack pointer to a location on the stack for consumption by the "Q" memory constraint.
Agreed. My bug. Will fix.
Quoted 1 line
This store is not necessary and should preferably be avoided.
I agree that the extra store should be avoided. I wasn't unable to remove it. Can you suggest how?
Thanks,
Behan
Signature · 2 lines
Behan Webster behanw at converseincode.com
Re: [PATCH 2/5] arm: LLVMLinux: use current_stack_pointer for percpu
In reply to [email protected]Quoted 22 lines · behanw at converseincode.com
From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new
current_stack_pointermacro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.Signed-off-by Mark Charlebois <[email protected]>Signed-off-by Behan Webster <[email protected]>Reviewed-by Jan-Simon M?ller <[email protected]>--- arch/arm/include/asm/percpu.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/arm/include/asm/percpu.h +0
-0@@ -30,14 +30,14 @@ static inline void set_my_cpu_offset(unsigned long off)static inline unsigned long __my_cpu_offset (void ) {unsigned long off;
arch/arm/include/asm/percpu.h static inline void set_my_cpu_offset(unsigned long off) · line 40
Quoted 10 lines
arch/arm/include/asm/percpu.h +2
-2- registerunsigned long *sp asm ("sp" ); +unsigned long sp = current_stack_pointer;/* * Read TPIDRPRW. * We want to allow caching the value, so avoid using volatile and * instead use a fake stack read to hazard against barrier(). */ - asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (*sp)); + asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (sp));
This change doesn't look to be equivalent. Previously the *sp implied a memory location which doesn't appear to be the case anymore.
this sp trickery was introduced in commit 509eb76ebf97 to solve bad code generation (the commit log has the details). It would be good if Will Deacon could confirm that his test case still works fine with your change.
Nicolas
Re: [PATCH 2/5] arm: LLVMLinux: use current_stack_pointer for percpu
Hi guys,
Quoted 26 lines · Nicolas Pitre
Quoted 25 lines · behanw at converseincode.com
From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new
current_stack_pointermacro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.Signed-off-by Mark Charlebois <[email protected]>Signed-off-by Behan Webster <[email protected]>Reviewed-by Jan-Simon M?ller <[email protected]>--- arch/arm/include/asm/percpu.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/arm/include/asm/percpu.h +1
-1@@ -30,14 +30,14 @@ static inline void set_my_cpu_offset(unsigned long off)static inline unsigned long __my_cpu_offset (void ) {unsigned long off; - registerunsigned long *sp asm ("sp" ); +unsigned long sp = current_stack_pointer;
Quoted 10 lines
arch/arm/include/asm/percpu.h static inline void set_my_cpu_offset(unsigned long off) · line 40
Quoted 7 lines
arch/arm/include/asm/percpu.h +1
-1/* * Read TPIDRPRW. * We want to allow caching the value, so avoid using volatile and * instead use a fake stack read to hazard against barrier(). */ - asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (*sp)); + asm("mrc p15, 0, %0, c13, c0, 4" :"=r" (off) :"Q" (sp));This change doesn't look to be equivalent. Previously the *sp implied a memory location which doesn't appear to be the case anymore.
Having looked at the other comments in this thread, I had a crack with the following diff:
arch/arm/include/asm/percpu.h +2 -1
@@ -30,7 +30,8 @@ static inline void set_my_cpu_offset(unsigned long off)static inline unsigned long __my_cpu_offset (void ) {unsigned long off; - registerunsigned long *sp asm ("sp" ); +// register unsigned long *sp asm ("sp"); +unsigned long *sp = (unsigned long *)current_stack_pointer;/* * Read TPIDRPRW.
arch/arm/include/asm/thread_info.h +9 -0
@@ -89,6 +89,15 @@ struct thread_info {#define init_stack (init_thread_union.stack)/* +* how to get the current stack pointer from C +*/ +#define current_stack_pointer ({ \ + unsigned long current_sp; \ + asm ("mov %0, r13" : "=r" (current_sp)); \ + current_sp; \ +}) + +/* * how to get the thread information struct from C */ static inline struct thread_info *current_thread_info (void ) __attribute_const__;
Quoted 4 lines
this sp trickery was introduced in commit
509eb76ebf97to solve bad code generation (the commit log has the details). It would be good if Will Deacon could confirm that his test case still works fine with your change.
I resurrected your original test case for the patch in question (see below) and the code is worse with the new sp accessor, since it insists on moving sp into r3, which forces us to push r4 onto the stack:
Before:
c001ce6c: ee1d3f90 mrc 15, 0, r3, cr13, cr0, {4}
c001ce70: e790c103 ldr ip, [r0, r3, lsl #2]
c001ce74: ee1d3f90 mrc 15, 0, r3, cr13, cr0, {4}
c001ce78: e7911103 ldr r1, [r1, r3, lsl #2]
c001ce7c: e7d22003 ldrb r2, [r2, r3]
c001ce80: ee1d3f90 mrc 15, 0, r3, cr13, cr0, {4}
c001ce84: e7903103 ldr r3, [r0, r3, lsl #2]
c001ce88: e08c0001 add r0, ip, r1
c001ce8c: e0800002 add r0, r0, r2
c001ce90: e0800003 add r0, r0, r3
c001ce94: e12fff1e bx lrAfter:
c001ce74: e52d4004 push {r4} ; (str r4, [sp, #-4]!)
c001ce78: e1a0300d mov r3, sp
c001ce7c: ee1dcf90 mrc 15, 0, ip, cr13, cr0, {4}
c001ce80: e790410c ldr r4, [r0, ip, lsl #2]
c001ce84: ee1dcf90 mrc 15, 0, ip, cr13, cr0, {4}
c001ce88: e791110c ldr r1, [r1, ip, lsl #2]
c001ce8c: e7d2200c ldrb r2, [r2, ip]
c001ce90: ee1d3f90 mrc 15, 0, r3, cr13, cr0, {4}
c001ce94: e7903103 ldr r3, [r0, r3, lsl #2]
c001ce98: e0840001 add r0, r4, r1
c001ce9c: e0800002 add r0, r0, r2
c001cea0: e0800003 add r0, r0, r3
c001cea4: e8bd0010 pop {r4}
c001cea8: e12fff1e bx lrWill
--->8
int foo(int *a, int *b, char *c)
{
int x, y, z;
x = a[__my_cpu_offset];
barrier();
y = b[__my_cpu_offset];
z = c[__my_cpu_offset];
barrier();
return x + y + z + a[__my_cpu_offset];
}PATCH 3/5 arm: LLVMLinux: Use current_stack_pointer for return_address
In reply to [email protected]From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new current_stack_pointer macro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.
--- arch/arm/kernel/return_address.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
arch/arm/kernel/return_address.c +1 -2
@@ -39,13 +39,12 @@ void *return_address(unsigned int level) {struct return_address_data data;struct stackframe frame; - registerunsigned long current_sp asm ("sp" ); data.level = level +2 ; data.addr =NULL ; frame.fp = (unsigned long )__builtin_frame_address (0 ); - frame.sp = current_sp; + frame.sp = current_stack_pointer; frame.lr = (unsigned long )__builtin_return_address (0 ); frame.pc = (unsigned long )return_address;
Signature · 1 line
1.8.1.2
PATCH 4/5 arm: LLVMLinux: Use current_stack_pointer in save_stack_trace_tsk
In reply to [email protected]From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new current_stack_pointer macro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.
--- arch/arm/kernel/stacktrace.c | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-)
arch/arm/kernel/stacktrace.c +1 -3
@@ -109,11 +109,9 @@ void save_stack_trace_tsk(struct task_struct *tsk, struct stack_trace *trace) frame.pc =thread_saved_pc (tsk);#endif }else { - registerunsigned long current_sp asm ("sp" ); - data.no_sched_functions =0 ; frame.fp = (unsigned long )__builtin_frame_address (0 ); - frame.sp = current_sp; + frame.sp = current_stack_pointer; frame.lr = (unsigned long )__builtin_return_address (0 ); frame.pc = (unsigned long )save_stack_trace_tsk; }
Signature · 1 line
1.8.1.2
PATCH 5/5 arm: LLVMLinux: Use current_stack_pointer in unwind_backtrace
In reply to [email protected]From: Behan Webster <[email protected]>
The existing code uses named registers to get the value of the stack pointer. The new current_stack_pointer macro is more readable and allows for a central portable implementation of how to get the stack pointer with ASM. This change supports being able to compile the kernel with both gcc and Clang.
--- arch/arm/kernel/unwind.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
arch/arm/kernel/unwind.c +1 -2
@@ -408,7 +408,6 @@ int unwind_frame(struct stackframe *frame)void unwind_backtrace (struct pt_regs *regs,struct task_struct *tsk) {struct stackframe frame; - registerunsigned long current_sp asm ("sp" );pr_debug ("%s(regs = %p tsk = %p)\n" , __func__, regs, tsk);
@@ -424,7 +423,7 @@ void unwind_backtrace(struct pt_regs *regs, struct task_struct *tsk) ? regs->ARM_pc : regs->ARM_lr; }else if (tsk == current) { frame.fp = (unsigned long )__builtin_frame_address (0 ); - frame.sp = current_sp; + frame.sp = current_stack_pointer; frame.lr = (unsigned long )__builtin_return_address (0 ); frame.pc = (unsigned long )unwind_backtrace; }else {
Signature · 1 line
1.8.1.2