x86/percpu: Use C for arch_raw_cpu_ptr(), to improve code generation

Implement arch_raw_cpu_ptr() in C to allow the compiler to perform
better optimizations, such as setting an appropriate base to compute
the address. The compiler is free to choose either MOV or ADD from
this_cpu_off address to construct the optimal final address.

There are some other issues when memory access to the percpu area is
implemented with an asm. Compilers can not eliminate asm common
subexpressions over basic block boundaries, but are extremely good
at optimizing memory access. By implementing arch_raw_cpu_ptr() in C,
the compiler can eliminate additional redundant loads from this_cpu_off,
further reducing the number of percpu offset reads from 1646 to 1631
on a test build, a -0.9% reduction.

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>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Uros Bizjak <ubizjak@gmail.com>
Cc: Sean Christopherson <seanjc@google.com>
Link: https://lore.kernel.org/r/20231015202523.189168-2-ubizjak@gmail.com

diff --git a/arch/x86/include/asm/percpu.h b/arch/x86/include/asm/percpu.h
index 915675f3ad6055c75e63f02e205b5f2d45ce336b..54746903b8c3c9ab8ee75699a51b03cf85493e5b 100644 (file)
--- a/arch/x86/include/asm/percpu.h
+++ b/arch/x86/include/asm/percpu.h
@@ -49,6 +49,21 @@
 #define __force_percpu_prefix  "%%"__stringify(__percpu_seg)":"
 #define __my_cpu_offset                this_cpu_read(this_cpu_off)
 
+#ifdef CONFIG_USE_X86_SEG_SUPPORT
+/*
+ * Efficient implementation for cases in which the compiler supports
+ * named address spaces.  Allows the compiler to perform additional
+ * optimizations that can save more instructions.
+ */
+#define arch_raw_cpu_ptr(ptr)                                  \
+({                                                             \
+       unsigned long tcp_ptr__;                                \
+       tcp_ptr__ = __raw_cpu_read(, this_cpu_off);             \
+                                                               \
+       tcp_ptr__ += (unsigned long)(ptr);                      \
+       (typeof(*(ptr)) __kernel __force *)tcp_ptr__;           \
+})
+#else /* CONFIG_USE_X86_SEG_SUPPORT */
 /*
  * Compared to the generic __my_cpu_offset version, the following
  * saves one instruction and avoids clobbering a temp register.
@@ -63,6 +78,8 @@
        tcp_ptr__ += (unsigned long)(ptr);                      \
        (typeof(*(ptr)) __kernel __force *)tcp_ptr__;           \
 })
+#endif /* CONFIG_USE_X86_SEG_SUPPORT */
+
 #else /* CONFIG_SMP */
 #define __percpu_seg_override
 #define __percpu_prefix                ""