From: "Steven Rostedt (VMware)" <[email protected]> When running the likely/unlikely profiler, one of the results did not look accurate. It noted that the unlikely() in link_path_walk() was 100% incorrect. When I added a trace_printk() to see what was happening there, it became 80% correct! Looking deeper into what whas happening, I found that gcc split that if statement into two paths. One where the if statement became a constant, the other path a variable. The other path had the if statement always hit (making the unlikely there, always false), but since the #define unlikely() has:
#define unlikely() (__builtin_constant_p(x) ? !!(x) : __branch_check__(x, 0)) Where constants are ignored by the branch profiler, the "constant" path made by the compiler was ignored, even though it was hit 80% of the time. By just passing the constant value to the __branch_check__() function and tracing it out of line (as always correct, as likely/unlikely isn't a factor for constants), then we get back the accurate readings of branches that were optimized by gcc causing part of the execution to become constant. Signed-off-by: Steven Rostedt (VMware) <[email protected]> --- include/linux/compiler.h | 14 ++++++++------ kernel/trace/trace_branch.c | 6 +++++- 2 files changed, 13 insertions(+), 7 deletions(-) diff --git a/include/linux/compiler.h b/include/linux/compiler.h index cf0fa5d86059..bbbe1570de1c 100644 --- a/include/linux/compiler.h +++ b/include/linux/compiler.h @@ -107,12 +107,13 @@ struct ftrace_branch_data { */ #if defined(CONFIG_TRACE_BRANCH_PROFILING) \ && !defined(DISABLE_BRANCH_PROFILING) && !defined(__CHECKER__) -void ftrace_likely_update(struct ftrace_branch_data *f, int val, int expect); +void ftrace_likely_update(struct ftrace_branch_data *f, int val, + int expect, int is_constant); #define likely_notrace(x) __builtin_expect(!!(x), 1) #define unlikely_notrace(x) __builtin_expect(!!(x), 0) -#define __branch_check__(x, expect) ({ \ +#define __branch_check__(x, expect, is_constant) ({ \ int ______r; \ static struct ftrace_branch_data \ __attribute__((__aligned__(4))) \ @@ -122,8 +123,9 @@ void ftrace_likely_update(struct ftrace_branch_data *f, int val, int expect); .file = __FILE__, \ .line = __LINE__, \ }; \ - ______r = likely_notrace(x); \ - ftrace_likely_update(&______f, ______r, expect); \ + ______r = __builtin_expect(!!(x), expect); \ + ftrace_likely_update(&______f, ______r, \ + expect, is_constant); \ ______r; \ }) @@ -133,10 +135,10 @@ void ftrace_likely_update(struct ftrace_branch_data *f, int val, int expect); * written by Daniel Walker. */ # ifndef likely -# define likely(x) (__builtin_constant_p(x) ? !!(x) : __branch_check__(x, 1)) +# define likely(x) (__branch_check__(x, 1, __builtin_constant_p(x))) # endif # ifndef unlikely -# define unlikely(x) (__builtin_constant_p(x) ? !!(x) : __branch_check__(x, 0)) +# define unlikely(x) (__branch_check__(x, 0, __builtin_constant_p(x))) # endif #ifdef CONFIG_PROFILE_ALL_BRANCHES diff --git a/kernel/trace/trace_branch.c b/kernel/trace/trace_branch.c index 75489de546b6..7afe426ea528 100644 --- a/kernel/trace/trace_branch.c +++ b/kernel/trace/trace_branch.c @@ -200,8 +200,12 @@ void trace_likely_condition(struct ftrace_branch_data *f, int val, int expect) } #endif /* CONFIG_BRANCH_TRACER */ -void ftrace_likely_update(struct ftrace_branch_data *f, int val, int expect) +void ftrace_likely_update(struct ftrace_branch_data *f, int val, + int expect, int is_constant) { + /* A constant is always correct */ + if (is_constant) + val = expect; /* * I would love to have a trace point here instead, but the * trace point code is so inundated with unlikely and likely -- 2.10.2
