diff options
| author |  Uros Bizjak <ubizjak@gmail.com>
| 2023-10-20 18:19:20 +0200 |
|---|---|---|
| committer |  Ingo Molnar <mingo@kernel.org>
| 2023-10-23 11:27:35 +0200 |
| commit | ed2f752e0e0a21d941ca0ee539ef3d4cd576bc5e (patch) | |
| tree | 69748caac82f958df3f9667e70347911f7ebc679 /arch/x86/kernel/vmlinux.lds.S | |
| parent | x86/percpu: Introduce %rip-relative addressing to PER_CPU_VAR() (diff) | |
| download | linux-ed2f752e0e0a21d941ca0ee539ef3d4cd576bc5e.tar.gz linux-ed2f752e0e0a21d941ca0ee539ef3d4cd576bc5e.tar.bz2 linux-ed2f752e0e0a21d941ca0ee539ef3d4cd576bc5e.zip | |
x86/percpu: Introduce const-qualified const_pcpu_hot to micro-optimize code generation
Some variables in pcpu_hot, currently current_task and top_of_stack
are actually per-thread variables implemented as per-CPU variables
and thus stable for the duration of the respective task. There is
already an attempt to eliminate redundant reads from these variables
using this_cpu_read_stable() asm macro, which hides the dependency
on the read memory address. However, the compiler has limited ability
to eliminate asm common subexpressions, so this approach results in a
limited success.
The solution is to allow more aggressive elimination by aliasing
pcpu_hot into a const-qualified const_pcpu_hot, and to read stable
per-CPU variables from this constant copy.
The current per-CPU infrastructure does not support reads from
const-qualified variables. However, when the compiler supports segment
qualifiers, it is possible to declare the const-aliased variable in
the relevant named address space. The compiler considers access to the
variable, declared in this way, as a read from a constant location,
and will optimize reads from the variable accordingly.
By implementing constant-qualified const_pcpu_hot, the compiler can
eliminate redundant reads from the constant variables, reducing the
number of loads from current_task from 3766 to 3217 on a test build,
a -14.6% reduction.
The reduction of loads translates to the following code savings:
text data bss dec hex filename
25,477,353 4389456 808452 30675261 1d4113d vmlinux-old.o
25,476,074 4389440 808452 30673966 1d40c2e vmlinux-new.o
representing a code size reduction of -1279 bytes.
[ mingo: Updated the changelog, EXPORT(const_pcpu_hot). ]
Co-developed-by: Nadav Amit <namit@vmware.com>
Signed-off-by: Nadav Amit <namit@vmware.com>
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20231020162004.135244-1-ubizjak@gmail.com
Diffstat (limited to 'arch/x86/kernel/vmlinux.lds.S')
| -rw-r--r-- | arch/x86/kernel/vmlinux.lds.S | 1 |
1 files changed, 1 insertions, 0 deletions
diff --git a/arch/x86/kernel/vmlinux.lds.S b/arch/x86/kernel/vmlinux.lds.S index 83d41c2601d7..e701f2dcea29 100644 --- a/arch/x86/kernel/vmlinux.lds.S +++ b/arch/x86/kernel/vmlinux.lds.S @@ -46,6 +46,7 @@ ENTRY(phys_startup_64) #endif jiffies = jiffies_64; +const_pcpu_hot = pcpu_hot; #if defined(CONFIG_X86_64) /* |