diff options
author | Joseph Myers <joseph@codesourcery.com> | 2012-10-31 13:01:17 +0000 |
---|---|---|
committer | Joseph Myers <joseph@codesourcery.com> | 2012-10-31 13:01:17 +0000 |
commit | ef82f4da79332e504e17729536d501add283719b (patch) | |
tree | 02fa059ea66f20f529767cedbcb25a8be3d2fd8b | |
parent | 954ef0d98d1fd8cc7a16a7289f286e885db36690 (diff) | |
download | glibc-ef82f4da79332e504e17729536d501add283719b.zip glibc-ef82f4da79332e504e17729536d501add283719b.tar.gz glibc-ef82f4da79332e504e17729536d501add283719b.tar.bz2 |
Fix fma underflow exceptions in after-rounding edge cases.
-rw-r--r-- | ChangeLog | 21 | ||||
-rw-r--r-- | math/libm-test.inc | 102 | ||||
-rw-r--r-- | sysdeps/ieee754/dbl-64/s_fma.c | 12 | ||||
-rw-r--r-- | sysdeps/ieee754/ldbl-128/s_fmal.c | 12 | ||||
-rw-r--r-- | sysdeps/ieee754/ldbl-96/s_fmal.c | 12 |
5 files changed, 159 insertions, 0 deletions
@@ -1,3 +1,24 @@ +2012-10-31 Joseph Myers <joseph@codesourcery.com> + + * sysdeps/ieee754/dbl-64/s_fma.c: Include <tininess.h>. + (__fma): Do not extract and scale down low bits on after-rounding + systems when result rounded to normal precision would have normal + exponent. + * sysdeps/ieee754/ldbl-128/s_fmal.c: Include <tininess.h>. + (__fmal): Do not extract and scale down low bits on after-rounding + systems when result rounded to normal precision would have normal + exponent. + * sysdeps/ieee754/ldbl-96/s_fmal.c: Include <tininess.h>. + (__fmal): Do not extract and scale down low bits on after-rounding + systems when result rounded to normal precision would have normal + exponent. + * math/libm-test.inc (UNDERFLOW_EXCEPTION_BEFORE_ROUNDING): New + macro. + (fma_test): Add more tests. + (fma_test_towardzero): Likewise. + (fma_test_downward): Likewise. + (fma_test_upward): Likewise. + 2012-10-30 H.J. Lu <hongjiu.lu@intel.com> * sysdeps/i386/tininess.h: Renamed to ... diff --git a/math/libm-test.inc b/math/libm-test.inc index 1e067fe..91019a1 100644 --- a/math/libm-test.inc +++ b/math/libm-test.inc @@ -132,6 +132,7 @@ #include <stdio.h> #include <string.h> #include <argp.h> +#include <tininess.h> /* Allow platforms without all rounding modes to test properly, assuming they provide an __FE_UNDEFINED in <bits/fenv.h> which @@ -187,6 +188,11 @@ #else # define UNDERFLOW_EXCEPTION_LDOUBLE_IBM 0 #endif +/* Values underflowing on architectures detecting tininess before + rounding, but not on those detecting tininess after rounding. */ +#define UNDERFLOW_EXCEPTION_BEFORE_ROUNDING (TININESS_AFTER_ROUNDING \ + ? 0 \ + : UNDERFLOW_EXCEPTION) /* Various constants (we must supply them precalculated for accuracy). */ #define M_PI_6l .52359877559829887307710723054658383L @@ -4618,6 +4624,12 @@ fma_test (void) TEST_fff_f (fma, 0x1.fffffep+127, 2.0, -0x1.fffffep+127, 0x1.fffffep+127); TEST_fff_f (fma, 0x1.4p-126, 0x1.000004p-1, 0x1p-128, 0x1.c00004p-127, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-126, 0x1.000004p-1, -0x1p-128, -0x1.c00004p-127, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffff8p-126, 0x1.000002p-1, 0x1p-149, 0x1p-126, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1.fffff8p-126, 0x1.000002p-1, -0x1p-149, -0x1p-126, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, 0x1p-149, 0x1p-1, 0x0.fffffep-126, 0x1p-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-149, 0x1p-1, -0x0.fffffep-126, -0x1p-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-149, 0x1.1p-1, 0x0.fffffep-126, 0x1p-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-149, 0x1.1p-1, -0x0.fffffep-126, -0x1p-126, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_DOUBLE) && DBL_MANT_DIG == 53 TEST_fff_f (fma, 0x1.7fp+13, 0x1.0000000000001p+0, 0x1.ffep-48, 0x1.7f00000000001p+13); @@ -4642,6 +4654,12 @@ fma_test (void) TEST_fff_f (fma, -0x1.fffffffffffffp-711, 0x1.fffffffffffffp-275, 0x1.fffffe00007ffp-983, 0x1.7ffffe00007ffp-983); TEST_fff_f (fma, 0x1.4p-1022, 0x1.0000000000002p-1, 0x1p-1024, 0x1.c000000000002p-1023, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-1022, 0x1.0000000000002p-1, -0x1p-1024, -0x1.c000000000002p-1023, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, 0x1p-1074, 0x1p-1022, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, -0x1p-1074, -0x1p-1022, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, 0x1p-1074, 0x1p-1, 0x0.fffffffffffffp-1022, 0x1p-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-1074, 0x1p-1, -0x0.fffffffffffffp-1022, -0x1p-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-1074, 0x1.1p-1, 0x0.fffffffffffffp-1022, 0x1p-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-1074, 0x1.1p-1, -0x0.fffffffffffffp-1022, -0x1p-1022, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 64 TEST_fff_f (fma, -0x8.03fcp+3696L, 0xf.fffffffffffffffp-6140L, 0x8.3ffffffffffffffp-2450L, -0x8.01ecp-2440L); @@ -4652,6 +4670,12 @@ fma_test (void) TEST_fff_f (fma, 0xb.ffffp-4777L, 0x8.000000fffffffffp-11612L, -0x0.3800fff8p-16385L, 0x5.c7fe80c7ffeffffp-16385L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, 0x1.4p-16382L, 0x1.0000000000000004p-1L, 0x1p-16384L, 0x1.c000000000000004p-16383L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-16382L, 0x1.0000000000000004p-1L, -0x1p-16384L, -0x1.c000000000000004p-16383L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffffffffffffff8p-16382L, 0x1.0000000000000002p-1L, 0x1p-16445L, 0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1.fffffffffffffff8p-16382L, 0x1.0000000000000002p-1L, -0x1p-16445L, -0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, 0x1p-16445L, 0x1p-1L, 0x0.fffffffffffffffep-16382L, 0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16445L, 0x1p-1L, -0x0.fffffffffffffffep-16382L, -0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16445L, 0x1.1p-1L, 0x0.fffffffffffffffep-16382L, 0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16445L, 0x1.1p-1L, -0x0.fffffffffffffffep-16382L, -0x1p-16382L, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 113 TEST_fff_f (fma, 0x1.bb2de33e02ccbbfa6e245a7c1f71p-2584L, -0x1.6b500daf0580d987f1bc0cadfcddp-13777L, 0x1.613cd91d9fed34b33820e5ab9d8dp-16378L, -0x1.3a79fb50eb9ce887cffa0f09bd9fp-16360L); @@ -4669,6 +4693,12 @@ fma_test (void) TEST_fff_f (fma, 0x1.801181509c03bdbef10d6165588cp-15131L, 0x1.ad86f8e57d3d40bfa8007780af63p-368L, -0x1.6e9df0dab1c9f1d7a6043c390741p-15507L, 0x1.417c9b2b15e2ad57dc9e0e920844p-15498L); TEST_fff_f (fma, 0x1.4p-16382L, 0x1.0000000000000000000000000002p-1L, 0x1p-16384L, 0x1.c000000000000000000000000002p-16383L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-16382L, 0x1.0000000000000000000000000002p-1L, -0x1p-16384L, -0x1.c000000000000000000000000002p-16383L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffffffffffffffffffffffffffcp-16382L, 0x1.0000000000000000000000000001p-1L, 0x1p-16494L, 0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1.fffffffffffffffffffffffffffcp-16382L, 0x1.0000000000000000000000000001p-1L, -0x1p-16494L, -0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, 0x1p-16494L, 0x1p-1L, 0x0.ffffffffffffffffffffffffffffp-16382L, 0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16494L, 0x1p-1L, -0x0.ffffffffffffffffffffffffffffp-16382L, -0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16494L, 0x1.1p-1L, 0x0.ffffffffffffffffffffffffffffp-16382L, 0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16494L, 0x1.1p-1L, -0x0.ffffffffffffffffffffffffffffp-16382L, -0x1p-16382L, UNDERFLOW_EXCEPTION); #endif END (fma); @@ -4727,18 +4757,42 @@ fma_test_towardzero (void) #if defined (TEST_FLOAT) && FLT_MANT_DIG == 24 TEST_fff_f (fma, 0x1.4p-126, 0x1.000004p-1, 0x1p-128, 0x1.c00004p-127, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-126, 0x1.000004p-1, -0x1p-128, -0x1.c00004p-127, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffff8p-126, 0x1.000002p-1, 0x1p-149, 0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1.fffff8p-126, 0x1.000002p-1, -0x1p-149, -0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-149, 0x1p-1, 0x0.fffffep-126, 0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-149, 0x1p-1, -0x0.fffffep-126, -0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-149, 0x1.1p-1, 0x0.fffffep-126, 0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-149, 0x1.1p-1, -0x0.fffffep-126, -0x0.fffffep-126, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_DOUBLE) && DBL_MANT_DIG == 53 TEST_fff_f (fma, 0x1.4p-1022, 0x1.0000000000002p-1, 0x1p-1024, 0x1.c000000000002p-1023, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-1022, 0x1.0000000000002p-1, -0x1p-1024, -0x1.c000000000002p-1023, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, 0x1p-1074, 0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, -0x1p-1074, -0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-1074, 0x1p-1, 0x0.fffffffffffffp-1022, 0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-1074, 0x1p-1, -0x0.fffffffffffffp-1022, -0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-1074, 0x1.1p-1, 0x0.fffffffffffffp-1022, 0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-1074, 0x1.1p-1, -0x0.fffffffffffffp-1022, -0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 64 TEST_fff_f (fma, 0x1.4p-16382L, 0x1.0000000000000004p-1L, 0x1p-16384L, 0x1.c000000000000004p-16383L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-16382L, 0x1.0000000000000004p-1L, -0x1p-16384L, -0x1.c000000000000004p-16383L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffffffffffffff8p-16382L, 0x1.0000000000000002p-1L, 0x1p-16445L, 0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1.fffffffffffffff8p-16382L, 0x1.0000000000000002p-1L, -0x1p-16445L, -0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16445L, 0x1p-1L, 0x0.fffffffffffffffep-16382L, 0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16445L, 0x1p-1L, -0x0.fffffffffffffffep-16382L, -0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16445L, 0x1.1p-1L, 0x0.fffffffffffffffep-16382L, 0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16445L, 0x1.1p-1L, -0x0.fffffffffffffffep-16382L, -0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 113 TEST_fff_f (fma, 0x1.4p-16382L, 0x1.0000000000000000000000000002p-1L, 0x1p-16384L, 0x1.c000000000000000000000000002p-16383L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-16382L, 0x1.0000000000000000000000000002p-1L, -0x1p-16384L, -0x1.c000000000000000000000000002p-16383L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffffffffffffffffffffffffffcp-16382L, 0x1.0000000000000000000000000001p-1L, 0x1p-16494L, 0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1.fffffffffffffffffffffffffffcp-16382L, 0x1.0000000000000000000000000001p-1L, -0x1p-16494L, -0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16494L, 0x1p-1L, 0x0.ffffffffffffffffffffffffffffp-16382L, 0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16494L, 0x1p-1L, -0x0.ffffffffffffffffffffffffffffp-16382L, -0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16494L, 0x1.1p-1L, 0x0.ffffffffffffffffffffffffffffp-16382L, 0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16494L, 0x1.1p-1L, -0x0.ffffffffffffffffffffffffffffp-16382L, -0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); #endif } @@ -4800,18 +4854,42 @@ fma_test_downward (void) #if defined (TEST_FLOAT) && FLT_MANT_DIG == 24 TEST_fff_f (fma, 0x1.4p-126, 0x1.000004p-1, 0x1p-128, 0x1.c00004p-127, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-126, 0x1.000004p-1, -0x1p-128, -0x1.c00008p-127, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffff8p-126, 0x1.000002p-1, 0x1p-149, 0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1.fffff8p-126, 0x1.000002p-1, -0x1p-149, -0x1p-126, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, 0x1p-149, 0x1p-1, 0x0.fffffep-126, 0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-149, 0x1p-1, -0x0.fffffep-126, -0x1p-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-149, 0x1.1p-1, 0x0.fffffep-126, 0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-149, 0x1.1p-1, -0x0.fffffep-126, -0x1p-126, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); #endif #if defined (TEST_DOUBLE) && DBL_MANT_DIG == 53 TEST_fff_f (fma, 0x1.4p-1022, 0x1.0000000000002p-1, 0x1p-1024, 0x1.c000000000002p-1023, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-1022, 0x1.0000000000002p-1, -0x1p-1024, -0x1.c000000000004p-1023, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, 0x1p-1074, 0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, -0x1p-1074, -0x1p-1022, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, 0x1p-1074, 0x1p-1, 0x0.fffffffffffffp-1022, 0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-1074, 0x1p-1, -0x0.fffffffffffffp-1022, -0x1p-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-1074, 0x1.1p-1, 0x0.fffffffffffffp-1022, 0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-1074, 0x1.1p-1, -0x0.fffffffffffffp-1022, -0x1p-1022, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); #endif #if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 64 TEST_fff_f (fma, 0x1.4p-16382L, 0x1.0000000000000004p-1L, 0x1p-16384L, 0x1.c000000000000004p-16383L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-16382L, 0x1.0000000000000004p-1L, -0x1p-16384L, -0x1.c000000000000008p-16383L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffffffffffffff8p-16382L, 0x1.0000000000000002p-1L, 0x1p-16445L, 0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1.fffffffffffffff8p-16382L, 0x1.0000000000000002p-1L, -0x1p-16445L, -0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, 0x1p-16445L, 0x1p-1L, 0x0.fffffffffffffffep-16382L, 0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16445L, 0x1p-1L, -0x0.fffffffffffffffep-16382L, -0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16445L, 0x1.1p-1L, 0x0.fffffffffffffffep-16382L, 0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16445L, 0x1.1p-1L, -0x0.fffffffffffffffep-16382L, -0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); #endif #if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 113 TEST_fff_f (fma, 0x1.4p-16382L, 0x1.0000000000000000000000000002p-1L, 0x1p-16384L, 0x1.c000000000000000000000000002p-16383L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-16382L, 0x1.0000000000000000000000000002p-1L, -0x1p-16384L, -0x1.c000000000000000000000000004p-16383L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffffffffffffffffffffffffffcp-16382L, 0x1.0000000000000000000000000001p-1L, 0x1p-16494L, 0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1.fffffffffffffffffffffffffffcp-16382L, 0x1.0000000000000000000000000001p-1L, -0x1p-16494L, -0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, 0x1p-16494L, 0x1p-1L, 0x0.ffffffffffffffffffffffffffffp-16382L, 0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16494L, 0x1p-1L, -0x0.ffffffffffffffffffffffffffffp-16382L, -0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16494L, 0x1.1p-1L, 0x0.ffffffffffffffffffffffffffffp-16382L, 0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16494L, 0x1.1p-1L, -0x0.ffffffffffffffffffffffffffffp-16382L, -0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); #endif } @@ -4873,18 +4951,42 @@ fma_test_upward (void) #if defined (TEST_FLOAT) && FLT_MANT_DIG == 24 TEST_fff_f (fma, 0x1.4p-126, 0x1.000004p-1, 0x1p-128, 0x1.c00008p-127, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-126, 0x1.000004p-1, -0x1p-128, -0x1.c00004p-127, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffff8p-126, 0x1.000002p-1, 0x1p-149, 0x1p-126, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1.fffff8p-126, 0x1.000002p-1, -0x1p-149, -0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-149, 0x1p-1, 0x0.fffffep-126, 0x1p-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-149, 0x1p-1, -0x0.fffffep-126, -0x0.fffffep-126, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-149, 0x1.1p-1, 0x0.fffffep-126, 0x1p-126, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1p-149, 0x1.1p-1, -0x0.fffffep-126, -0x0.fffffep-126, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_DOUBLE) && DBL_MANT_DIG == 53 TEST_fff_f (fma, 0x1.4p-1022, 0x1.0000000000002p-1, 0x1p-1024, 0x1.c000000000004p-1023, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-1022, 0x1.0000000000002p-1, -0x1p-1024, -0x1.c000000000002p-1023, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, 0x1p-1074, 0x1p-1022, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, -0x1p-1074, -0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-1074, 0x1p-1, 0x0.fffffffffffffp-1022, 0x1p-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-1074, 0x1p-1, -0x0.fffffffffffffp-1022, -0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-1074, 0x1.1p-1, 0x0.fffffffffffffp-1022, 0x1p-1022, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1p-1074, 0x1.1p-1, -0x0.fffffffffffffp-1022, -0x0.fffffffffffffp-1022, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 64 TEST_fff_f (fma, 0x1.4p-16382L, 0x1.0000000000000004p-1L, 0x1p-16384L, 0x1.c000000000000008p-16383L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-16382L, 0x1.0000000000000004p-1L, -0x1p-16384L, -0x1.c000000000000004p-16383L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffffffffffffff8p-16382L, 0x1.0000000000000002p-1L, 0x1p-16445L, 0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1.fffffffffffffff8p-16382L, 0x1.0000000000000002p-1L, -0x1p-16445L, -0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16445L, 0x1p-1L, 0x0.fffffffffffffffep-16382L, 0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16445L, 0x1p-1L, -0x0.fffffffffffffffep-16382L, -0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16445L, 0x1.1p-1L, 0x0.fffffffffffffffep-16382L, 0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1p-16445L, 0x1.1p-1L, -0x0.fffffffffffffffep-16382L, -0x0.fffffffffffffffep-16382L, UNDERFLOW_EXCEPTION); #endif #if defined (TEST_LDOUBLE) && LDBL_MANT_DIG == 113 TEST_fff_f (fma, 0x1.4p-16382L, 0x1.0000000000000000000000000002p-1L, 0x1p-16384L, 0x1.c000000000000000000000000004p-16383L, UNDERFLOW_EXCEPTION); TEST_fff_f (fma, -0x1.4p-16382L, 0x1.0000000000000000000000000002p-1L, -0x1p-16384L, -0x1.c000000000000000000000000002p-16383L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1.fffffffffffffffffffffffffffcp-16382L, 0x1.0000000000000000000000000001p-1L, 0x1p-16494L, 0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1.fffffffffffffffffffffffffffcp-16382L, 0x1.0000000000000000000000000001p-1L, -0x1p-16494L, -0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16494L, 0x1p-1L, 0x0.ffffffffffffffffffffffffffffp-16382L, 0x1p-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, -0x1p-16494L, 0x1p-1L, -0x0.ffffffffffffffffffffffffffffp-16382L, -0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); + TEST_fff_f (fma, 0x1p-16494L, 0x1.1p-1L, 0x0.ffffffffffffffffffffffffffffp-16382L, 0x1p-16382L, UNDERFLOW_EXCEPTION_BEFORE_ROUNDING); + TEST_fff_f (fma, -0x1p-16494L, 0x1.1p-1L, -0x0.ffffffffffffffffffffffffffffp-16382L, -0x0.ffffffffffffffffffffffffffffp-16382L, UNDERFLOW_EXCEPTION); #endif } diff --git a/sysdeps/ieee754/dbl-64/s_fma.c b/sysdeps/ieee754/dbl-64/s_fma.c index 108fd66..155a773 100644 --- a/sysdeps/ieee754/dbl-64/s_fma.c +++ b/sysdeps/ieee754/dbl-64/s_fma.c @@ -22,6 +22,7 @@ #include <fenv.h> #include <ieee754.h> #include <math_private.h> +#include <tininess.h> /* This implementation uses rounding to odd to avoid problems with double rounding. See a paper by Boldo and Melquiond: @@ -208,6 +209,17 @@ __fma (double x, double y, double z) for proper rounding. */ if (v.ieee.exponent == 106) { + /* If the exponent would be in the normal range when + rounding to normal precision with unbounded exponent + range, the exact result is known and spurious underflows + must be avoided on systems detecting tininess after + rounding. */ + if (TININESS_AFTER_ROUNDING) + { + w.d = a1 + u.d; + if (w.ieee.exponent == 107) + return w.d * 0x1p-106; + } /* v.ieee.mantissa1 & 2 is LSB bit of the result before rounding, v.ieee.mantissa1 & 1 is the round bit and j is our sticky bit. */ diff --git a/sysdeps/ieee754/ldbl-128/s_fmal.c b/sysdeps/ieee754/ldbl-128/s_fmal.c index 7fb9856..201bff9 100644 --- a/sysdeps/ieee754/ldbl-128/s_fmal.c +++ b/sysdeps/ieee754/ldbl-128/s_fmal.c @@ -22,6 +22,7 @@ #include <fenv.h> #include <ieee754.h> #include <math_private.h> +#include <tininess.h> /* This implementation uses rounding to odd to avoid problems with double rounding. See a paper by Boldo and Melquiond: @@ -209,6 +210,17 @@ __fmal (long double x, long double y, long double z) for proper rounding. */ if (v.ieee.exponent == 226) { + /* If the exponent would be in the normal range when + rounding to normal precision with unbounded exponent + range, the exact result is known and spurious underflows + must be avoided on systems detecting tininess after + rounding. */ + if (TININESS_AFTER_ROUNDING) + { + w.d = a1 + u.d; + if (w.ieee.exponent == 227) + return w.d * 0x1p-226L; + } /* v.ieee.mantissa3 & 2 is LSB bit of the result before rounding, v.ieee.mantissa3 & 1 is the round bit and j is our sticky bit. */ diff --git a/sysdeps/ieee754/ldbl-96/s_fmal.c b/sysdeps/ieee754/ldbl-96/s_fmal.c index 7310491..91333af 100644 --- a/sysdeps/ieee754/ldbl-96/s_fmal.c +++ b/sysdeps/ieee754/ldbl-96/s_fmal.c @@ -22,6 +22,7 @@ #include <fenv.h> #include <ieee754.h> #include <math_private.h> +#include <tininess.h> /* This implementation uses rounding to odd to avoid problems with double rounding. See a paper by Boldo and Melquiond: @@ -209,6 +210,17 @@ __fmal (long double x, long double y, long double z) for proper rounding. */ if (v.ieee.exponent == 128) { + /* If the exponent would be in the normal range when + rounding to normal precision with unbounded exponent + range, the exact result is known and spurious underflows + must be avoided on systems detecting tininess after + rounding. */ + if (TININESS_AFTER_ROUNDING) + { + w.d = a1 + u.d; + if (w.ieee.exponent == 129) + return w.d * 0x1p-128L; + } /* v.ieee.mantissa1 & 2 is LSB bit of the result before rounding, v.ieee.mantissa1 & 1 is the round bit and j is our sticky bit. */ |