This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
[PATCH] PowerPC: fix hypot/hypof FP exceptions
- From: Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>
- To: "GNU C. Library" <libc-alpha at sourceware dot org>
- Date: Mon, 29 Apr 2013 14:28:15 -0300
- Subject: [PATCH] PowerPC: fix hypot/hypof FP exceptions
Hi,
This patch fixes spurious overflow FP generated in PowerPC hypot code triggered
by recent cacos/cacosh/casin/casinh tests with argument as [+/-]0x1.fp-129 for
float and [+/-] 0x1.fp-1025 for double.
For hypotf I removed the unneeded tests (since sqrt for double precision will be
used in the end).
For hypot I adjusted the test to check if x/y with x>y is higher than 2^60 by
checking if first y * 2^60 will overflow. If so, the test used is x > (y*2^60).
With this patch and the previous ULPs update , I'm not seeing any more issues
on float and double testcases for PowerPC. Tested on PPC32 and PPC64.
Any tips, comments, advices?
---
2013-04-26 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
* sysdeps/powerpc/fpu/e_hypot.c: Fix spurious FP exception generated by
internal tests.
* sysdeps/powerpc/fpu/e_hypotf.c: Likewise.
--
diff --git a/sysdeps/powerpc/fpu/e_hypot.c b/sysdeps/powerpc/fpu/e_hypot.c
index f236339..d8568d6 100644
--- a/sysdeps/powerpc/fpu/e_hypot.c
+++ b/sysdeps/powerpc/fpu/e_hypot.c
@@ -26,6 +26,7 @@ static const double two600 = 4.149515568880993e+180;
static const double two1022 = 4.49423283715579e+307;
static const double twoM500 = 3.054936363499605e-151;
static const double twoM600 = 2.4099198651028841e-181;
+static const double two60factor = 1.5592502418239997e+290;
static const double pdnum = 2.225073858507201e-308;
/* __ieee754_hypot(x,y)
@@ -87,9 +88,20 @@ __ieee754_hypot (double x, double y)
x = y;
y = t;
}
- if (y == 0.0 || (x / y) > two60)
+ if (y == 0.0)
+ return x;
+ /* if y is higher enough, y * 2^60 might overflow. This tests if
+ y >= 1.7976931348623157e+308/2^60 (two60factor) and uses the
+ appropriate check to avoid the overflow exception generation. */
+ if (y > two60factor)
{
- return x + y;
+ if ((x / y) > two60)
+ return x + y;
+ }
+ else
+ {
+ if (x > (y * two60))
+ return x + y;
}
if (x > two500)
{
diff --git a/sysdeps/powerpc/fpu/e_hypotf.c b/sysdeps/powerpc/fpu/e_hypotf.c
index e97f0c3..93055af 100644
--- a/sysdeps/powerpc/fpu/e_hypotf.c
+++ b/sysdeps/powerpc/fpu/e_hypotf.c
@@ -69,22 +69,8 @@ static const float two30 = 1.0737418e09;
float
__ieee754_hypotf (float x, float y)
{
- x = fabsf (x);
- y = fabsf (y);
-
TEST_INF_NAN (x, y);
- if (y > x)
- {
- float t = y;
- y = x;
- x = t;
- }
- if (y == 0.0 || (x / y) > two30)
- {
- return x + y;
- }
-
return __ieee754_sqrt ((double) x * x + (double) y * y);
}
strong_alias (__ieee754_hypotf, __hypotf_finite)