math: optimize lrint on 32bit targets

lrint in (LONG_MAX, 1/DBL_EPSILON) and in (-1/DBL_EPSILON, LONG_MIN)
is not trivial: rounding to int may be inexact, but the conversion to
int may overflow and then the inexact flag must not be raised. (the
overflow threshold is rounding mode dependent).

this matters on 32bit targets (without single instruction lrint or
rint), so the common case (when there is no overflow) is optimized by
inlining the lrint logic, otherwise the old code is kept as a fallback.

on my laptop an i486 lrint call is asm:10ns, old c:30ns, new c:21ns
on a smaller arm core: old c:71ns, new c:34ns
on a bigger arm core: old c:27ns, new c:19ns

1 files changed, 27 insertions, 1 deletions

diff --git a/src/math/lrint.c b/src/math/lrint.c
index bdca8b7c..ddee7a0d 100644
--- a/src/math/lrint.c
+++ b/src/math/lrint.c
@@ -1,5 +1,6 @@
 #include <limits.h>
 #include <fenv.h>
+#include <math.h>
 #include "libm.h"
 
 /*
@@ -26,7 +27,18 @@ as a double.
 */
 
 #if LONG_MAX < 1U<<53 && defined(FE_INEXACT)
-long lrint(double x)
+#include <float.h>
+#include <stdint.h>
+#if FLT_EVAL_METHOD==0 || FLT_EVAL_METHOD==1
+#define EPS DBL_EPSILON
+#elif FLT_EVAL_METHOD==2
+#define EPS LDBL_EPSILON
+#endif
+#ifdef __GNUC__
+/* avoid stack frame in lrint */
+__attribute__((noinline))
+#endif
+static long lrint_slow(double x)
 {
 	#pragma STDC FENV_ACCESS ON
 	int e;
@@ -38,6 +50,20 @@ long lrint(double x)
 	/* conversion */
 	return x;
 }
+
+long lrint(double x)
+{
+	uint32_t abstop = asuint64(x)>>32 & 0x7fffffff;
+	uint64_t sign = asuint64(x) & (1ULL << 63);
+
+	if (abstop < 0x41dfffff) {
+		/* |x| < 0x7ffffc00, no overflow */
+		double_t toint = asdouble(asuint64(1/EPS) | sign);
+		double_t y = x + toint - toint;
+		return (long)y;
+	}
+	return lrint_slow(x);
+}
 #else
 long lrint(double x)
 {