File jm_perf.h¶

just-makeit performance annotation macros. More...

#include "jm_simd.h"

Macros¶

Type	Name
define	JM_ALIGNED (n) `[\_JM\_ALIGNED\_](jm__perf_8h.md#define-_jm_aligned_) (n)`
define	JM_ASSUME_ALIGNED (ptr, n) `[\_JM\_ASSUME\_ALIGNED\_](jm__perf_8h.md#define-_jm_assume_aligned_) (ptr, n)` Inform the compiler that ptr is aligned to n bytes.
define	JM_FORCEINLINE `[\_JM\_FORCEINLINE\_](jm__perf_8h.md#define-_jm_forceinline_)`
define	JM_HOT `[\_JM\_HOT\_](jm__perf_8h.md#define-_jm_hot_)`
define	JM_LIKELY (x) `[\_JM\_LIKELY\_](jm__perf_8h.md#define-_jm_likely_) (x)`
define	JM_PREFETCH (ptr, rw, loc) `[\_JM\_PREFETCH\_](jm__perf_8h.md#define-_jm_prefetch_) (ptr, rw, loc)` Issue a non-blocking prefetch hint to the CPU.
define	JM_RESTRICT `[\_JM\_RESTRICT\_](jm__perf_8h.md#define-_jm_restrict_)`
define	JM_UNLIKELY (x) `[\_JM\_UNLIKELY\_](jm__perf_8h.md#define-_jm_unlikely_) (x)`
define	JM_UNROLL (n) `[\_JM\_UNROLL\_](jm__perf_8h.md#define-_jm_unroll_) (n)` Unroll the immediately following for-loop exactly n times.
define	_JM_ALIGNED_ (n)
define	_JM_ASSUME_ALIGNED_ (p, n) `(p)`
define	_JM_FORCEINLINE_ `inline`
define	_JM_HOT_
define	_JM_LIKELY_ (x) `(x)`
define	_JM_PREFETCH_ (p, rw, loc)
define	_JM_RESTRICT_ `restrict`
define	_JM_UNLIKELY_ (x) `(x)`
define	_JM_UNROLL_ (n)

Detailed Description¶

Portable hints for the compiler and CPU: hot functions, forced inlining, restrict aliasing, branch prediction, alignment, prefetch, and loop unrolling. All macros degrade gracefully to safe no-ops on unknown compilers.

Include order: jm_perf.h includes jm_simd.h automatically. Either header may be included standalone — jm_simd.h guards against redefining JM_RESTRICT if jm_perf.h was already included.

Usage:

JM_HOT static void process(const float * JM_RESTRICT in,
                            float * JM_RESTRICT out, size_t n)
{
    JM_PREFETCH(in + 16, 0, 1);
    JM_UNROLL(4)
    for (size_t i = 0; i < n; i++)
        out[i] = in[i] * 2.0f;
}

Macro Definition Documentation¶

define JM_ALIGNED¶

#define JM_ALIGNED (
    n
) `_JM_ALIGNED_ (n)`

Align a variable or struct member to n bytes.

define JM_ASSUME_ALIGNED¶

Inform the compiler that ptr is aligned to n bytes.

#define JM_ASSUME_ALIGNED (
    ptr,
    n
) `_JM_ASSUME_ALIGNED_ (ptr, n)`

Enables aligned SIMD loads/stores on ISAs that penalise unaligned access. Returns ptr so it can be used in expressions. Falls back to ptr unchanged on unknown compilers.

define JM_FORCEINLINE¶

#define JM_FORCEINLINE `_JM_FORCEINLINE_`

Force inlining regardless of the compiler's cost model.

define JM_HOT¶

#define JM_HOT `_JM_HOT_`

Mark a function as performance-critical (hot section).

define JM_LIKELY¶

#define JM_LIKELY (
    x
) `_JM_LIKELY_ (x)`

Hint that x is almost always true.

define JM_PREFETCH¶

Issue a non-blocking prefetch hint to the CPU.

#define JM_PREFETCH (
    ptr,
    rw,
    loc
) `_JM_PREFETCH_ (ptr, rw, loc)`

Parameters:

ptr Address to prefetch.
rw 0 = read (PLD), 1 = write (PSTL).
loc Cache level: 3 = L1 (hottest), 0 = NTA (streaming).

Prefetching 1–2 cache lines ahead of the load keeps the data pipeline full on high-latency memory. No-op on unknown compilers.

define JM_RESTRICT¶

#define JM_RESTRICT `_JM_RESTRICT_`

Assert a pointer does not alias any other; enables vectorisation.

define JM_UNLIKELY¶

#define JM_UNLIKELY (
    x
) `_JM_UNLIKELY_ (x)`

Hint that x is almost never true.

define JM_UNROLL¶

Unroll the immediately following for-loop exactly n times.

#define JM_UNROLL (
    n
) `_JM_UNROLL_ (n)`

Applied before a for loop, instructs GCC/Clang to unroll it unconditionally by the given factor, regardless of the compiler's own cost model. Use only on tight, well-measured inner loops with a known iteration count — a large n on a non-trivial body will bloat code size and hurt instruction-cache pressure.

No-op on compilers that do not support the pragma.

define _JM_ALIGNED_¶

#define _JM_ALIGNED_ (
    n
)

define _JM_ASSUME_ALIGNED_¶

#define _JM_ASSUME_ALIGNED_ (
    p,
    n
) `(p)`

define _JM_FORCEINLINE_¶

#define _JM_FORCEINLINE_ `inline`

define _JM_HOT_¶

#define _JM_HOT_

define _JM_LIKELY_¶

#define _JM_LIKELY_ (
    x
) `(x)`

define _JM_PREFETCH_¶

#define _JM_PREFETCH_ (
    p,
    rw,
    loc
)

define _JM_RESTRICT_¶

#define _JM_RESTRICT_ `restrict`

define _JM_UNLIKELY_¶

#define _JM_UNLIKELY_ (
    x
) `(x)`

define _JM_UNROLL_¶

#define _JM_UNROLL_ (
    n
)

The documentation for this class was generated from the following file native/inc/jm_perf.h