Fix some regressions caused by the switch from gcc to clang. The fixes

are workarounds for various symptoms of the problem described in clang
bugs 3929, 8100, 8241, 10409, and 12958.

The regression tests did their job: they failed, someone brought it
up on the mailing lists, and then the issue got ignored for 6 months.
Oops. There may still be some regressions for functions we don't have
test coverage for yet.
This commit is contained in:
David Schultz 2013-05-27 08:50:10 +00:00
parent 6e7bc5b60d
commit 7dbbb6dde3
21 changed files with 78 additions and 59 deletions

View File

@ -121,12 +121,6 @@ COMMON_SRCS:= ${COMMON_SRCS:N${i:R}.c}
.endfor
.endif
# Some files need certain gcc built-in functions to be disabled, since gcc's
# model of the functions bogusly assumes -fno-trapping-math.
XRINT_CFLAGS= -fno-builtin-rint -fno-builtin-rintf -fno-builtin-rintl
CFLAGS+= ${XRINT_CFLAGS}
XRINT_CFLAGS:= ${.IMPSRC:M*/s_nearbyint.c:C/^.+$/${XRINT_CFLAGS}/:C/^$//}
SRCS= ${COMMON_SRCS} ${ARCH_SRCS}
INCS+= fenv.h math.h

View File

@ -39,14 +39,11 @@ __FBSDID("$FreeBSD$");
#define BIAS (LDBL_MAX_EXP - 1)
#define EXPMASK (BIAS + LDBL_MAX_EXP)
#if 0 /* XXX Prevent gcc from erroneously constant folding this. */
static const long double twom10000 = 0x1p-10000L;
#else
static volatile long double twom10000 = 0x1p-10000L;
#endif
static volatile long double
huge = 0x1p10000L,
twom10000 = 0x1p-10000L;
static const long double
huge = 0x1p10000L,
P1 = 0x1.62e42fefa39ef35793c7673007e6p-1L,
P2 = 0x1.ebfbdff82c58ea86f16b06ec9736p-3L,
P3 = 0x1.c6b08d704a0bf8b33a762bad3459p-5L,

View File

@ -43,12 +43,9 @@ __FBSDID("$FreeBSD$");
#define BIAS (LDBL_MAX_EXP - 1)
#define EXPMASK (BIAS + LDBL_MAX_EXP)
static const long double huge = 0x1p10000L;
#if 0 /* XXX Prevent gcc from erroneously constant folding this. */
static const long double twom10000 = 0x1p-10000L;
#else
static volatile long double twom10000 = 0x1p-10000L;
#endif
static volatile long double
huge = 0x1p10000L,
twom10000 = 0x1p-10000L;
static const double
redux = 0x1.8p63 / TBLSIZE,

View File

@ -84,7 +84,6 @@ __FBSDID("$FreeBSD$");
static const double
one = 1.0,
halF[2] = {0.5,-0.5,},
huge = 1.0e+300,
o_threshold= 7.09782712893383973096e+02, /* 0x40862E42, 0xFEFA39EF */
u_threshold= -7.45133219101941108420e+02, /* 0xc0874910, 0xD52D3051 */
ln2HI[2] ={ 6.93147180369123816490e-01, /* 0x3fe62e42, 0xfee00000 */
@ -99,6 +98,7 @@ P4 = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */
P5 = 4.13813679705723846039e-08; /* 0x3E663769, 0x72BEA4D0 */
static volatile double
huge = 1.0e+300,
twom1000= 9.33263618503218878990e-302; /* 2**-1000=0x01700000,0*/
double

View File

@ -24,7 +24,6 @@ __FBSDID("$FreeBSD$");
static const float
one = 1.0,
halF[2] = {0.5,-0.5,},
huge = 1.0e+30,
o_threshold= 8.8721679688e+01, /* 0x42b17180 */
u_threshold= -1.0397208405e+02, /* 0xc2cff1b5 */
ln2HI[2] ={ 6.9314575195e-01, /* 0x3f317200 */
@ -39,7 +38,9 @@ invln2 = 1.4426950216e+00, /* 0x3fb8aa3b */
P1 = 1.6666625440e-1, /* 0xaaaa8f.0p-26 */
P2 = -2.7667332906e-3; /* -0xb55215.0p-32 */
static volatile float twom100 = 7.8886090522e-31; /* 2**-100=0x0d800000 */
static volatile float
huge = 1.0e+30,
twom100 = 7.8886090522e-31; /* 2**-100=0x0d800000 */
float
__ieee754_expf(float x)

View File

@ -81,6 +81,7 @@ Lg6 = 1.531383769920937332e-01, /* 3FC39A09 D078C69F */
Lg7 = 1.479819860511658591e-01; /* 3FC2F112 DF3E5244 */
static const double zero = 0.0;
static volatile double vzero = 0.0;
double
__ieee754_log(double x)
@ -94,7 +95,7 @@ __ieee754_log(double x)
k=0;
if (hx < 0x00100000) { /* x < 2**-1022 */
if (((hx&0x7fffffff)|lx)==0)
return -two54/zero; /* log(+-0)=-inf */
return -two54/vzero; /* log(+-0)=-inf */
if (hx<0) return (x-x)/zero; /* log(-#) = NaN */
k -= 54; x *= two54; /* subnormal number, scale up x */
GET_HIGH_WORD(hx,x);

View File

@ -34,6 +34,7 @@ log10_2hi = 3.01029995663611771306e-01, /* 0x3FD34413, 0x509F6000 */
log10_2lo = 3.69423907715893078616e-13; /* 0x3D59FEF3, 0x11F12B36 */
static const double zero = 0.0;
static volatile double vzero = 0.0;
double
__ieee754_log10(double x)
@ -47,7 +48,7 @@ __ieee754_log10(double x)
k=0;
if (hx < 0x00100000) { /* x < 2**-1022 */
if (((hx&0x7fffffff)|lx)==0)
return -two54/zero; /* log(+-0)=-inf */
return -two54/vzero; /* log(+-0)=-inf */
if (hx<0) return (x-x)/zero; /* log(-#) = NaN */
k -= 54; x *= two54; /* subnormal number, scale up x */
GET_HIGH_WORD(hx,x);

View File

@ -28,6 +28,7 @@ log10_2hi = 3.0102920532e-01, /* 0x3e9a2080 */
log10_2lo = 7.9034151668e-07; /* 0x355427db */
static const float zero = 0.0;
static volatile float vzero = 0.0;
float
__ieee754_log10f(float x)
@ -40,7 +41,7 @@ __ieee754_log10f(float x)
k=0;
if (hx < 0x00800000) { /* x < 2**-126 */
if ((hx&0x7fffffff)==0)
return -two25/zero; /* log(+-0)=-inf */
return -two25/vzero; /* log(+-0)=-inf */
if (hx<0) return (x-x)/zero; /* log(-#) = NaN */
k -= 25; x *= two25; /* subnormal number, scale up x */
GET_FLOAT_WORD(hx,x);

View File

@ -34,6 +34,7 @@ ivln2hi = 1.44269504072144627571e+00, /* 0x3ff71547, 0x65200000 */
ivln2lo = 1.67517131648865118353e-10; /* 0x3de705fc, 0x2eefa200 */
static const double zero = 0.0;
static volatile double vzero = 0.0;
double
__ieee754_log2(double x)
@ -47,7 +48,7 @@ __ieee754_log2(double x)
k=0;
if (hx < 0x00100000) { /* x < 2**-1022 */
if (((hx&0x7fffffff)|lx)==0)
return -two54/zero; /* log(+-0)=-inf */
return -two54/vzero; /* log(+-0)=-inf */
if (hx<0) return (x-x)/zero; /* log(-#) = NaN */
k -= 54; x *= two54; /* subnormal number, scale up x */
GET_HIGH_WORD(hx,x);

View File

@ -26,6 +26,7 @@ ivln2hi = 1.4428710938e+00, /* 0x3fb8b000 */
ivln2lo = -1.7605285393e-04; /* 0xb9389ad4 */
static const float zero = 0.0;
static volatile float vzero = 0.0;
float
__ieee754_log2f(float x)
@ -38,7 +39,7 @@ __ieee754_log2f(float x)
k=0;
if (hx < 0x00800000) { /* x < 2**-126 */
if ((hx&0x7fffffff)==0)
return -two25/zero; /* log(+-0)=-inf */
return -two25/vzero; /* log(+-0)=-inf */
if (hx<0) return (x-x)/zero; /* log(-#) = NaN */
k -= 25; x *= two25; /* subnormal number, scale up x */
GET_FLOAT_WORD(hx,x);

View File

@ -30,6 +30,7 @@ Lg3 = 0x91e9ee.0p-25, /* 0.28498786688 */
Lg4 = 0xf89e26.0p-26; /* 0.24279078841 */
static const float zero = 0.0;
static volatile float vzero = 0.0;
float
__ieee754_logf(float x)
@ -42,7 +43,7 @@ __ieee754_logf(float x)
k=0;
if (ix < 0x00800000) { /* x < 2**-126 */
if ((ix&0x7fffffff)==0)
return -two25/zero; /* log(+-0)=-inf */
return -two25/vzero; /* log(+-0)=-inf */
if (ix<0) return (x-x)/zero; /* log(-#) = NaN */
k -= 25; x *= two25; /* subnormal number, scale up x */
GET_FLOAT_WORD(ix,x);

View File

@ -36,7 +36,6 @@ __FBSDID("$FreeBSD$");
#define TBLSIZE (1 << TBLBITS)
static const double
huge = 0x1p1000,
redux = 0x1.8p52 / TBLSIZE,
P1 = 0x1.62e42fefa39efp-1,
P2 = 0x1.ebfbdff82c575p-3,
@ -44,7 +43,9 @@ static const double
P4 = 0x1.3b2ab88f70400p-7,
P5 = 0x1.5d88003875c74p-10;
static volatile double twom1000 = 0x1p-1000;
static volatile double
huge = 0x1p1000,
twom1000 = 0x1p-1000;
static const double tbl[TBLSIZE * 2] = {
/* exp2(z + eps) eps */

View File

@ -36,14 +36,15 @@ __FBSDID("$FreeBSD$");
#define TBLSIZE (1 << TBLBITS)
static const float
huge = 0x1p100f,
redux = 0x1.8p23f / TBLSIZE,
P1 = 0x1.62e430p-1f,
P2 = 0x1.ebfbe0p-3f,
P3 = 0x1.c6b348p-5f,
P4 = 0x1.3b2c9cp-7f;
static volatile float twom100 = 0x1p-100f;
static volatile float
huge = 0x1p100f,
twom100 = 0x1p-100f;
static const double exp2ft[TBLSIZE] = {
0x1.6a09e667f3bcdp-1,

View File

@ -115,7 +115,6 @@ __FBSDID("$FreeBSD$");
static const double
one = 1.0,
huge = 1.0e+300,
tiny = 1.0e-300,
o_threshold = 7.09782712893383973096e+02,/* 0x40862E42, 0xFEFA39EF */
ln2_hi = 6.93147180369123816490e-01,/* 0x3fe62e42, 0xfee00000 */
@ -128,6 +127,8 @@ Q3 = -7.93650757867487942473e-05, /* BF14CE19 9EAADBB7 */
Q4 = 4.00821782732936239552e-06, /* 3ED0CFCA 86E65239 */
Q5 = -2.01099218183624371326e-07; /* BE8AFDB7 6E09C32D */
static volatile double huge = 1.0e+300;
double
expm1(double x)
{

View File

@ -23,7 +23,6 @@ __FBSDID("$FreeBSD$");
static const float
one = 1.0,
huge = 1.0e+30,
tiny = 1.0e-30,
o_threshold = 8.8721679688e+01,/* 0x42b17180 */
ln2_hi = 6.9313812256e-01,/* 0x3f317180 */
@ -37,6 +36,8 @@ invln2 = 1.4426950216e+00,/* 0x3fb8aa3b */
Q1 = -3.3333212137e-2, /* -0x888868.0p-28 */
Q2 = 1.5807170421e-3; /* 0xcf3010.0p-33 */
static volatile float huge = 1.0e+30;
float
expm1f(float x)
{

View File

@ -238,6 +238,8 @@ fma(double x, double y, double z)
zs = copysign(DBL_MIN, zs);
fesetround(FE_TONEAREST);
/* work around clang bug 8100 */
volatile double vxs = xs;
/*
* Basic approach for round-to-nearest:
@ -247,7 +249,7 @@ fma(double x, double y, double z)
* adj = xy.lo + r.lo (inexact; low bit is sticky)
* result = r.hi + adj (correctly rounded)
*/
xy = dd_mul(xs, ys);
xy = dd_mul(vxs, ys);
r = dd_add(xy.hi, zs);
spread = ex + ey;
@ -268,7 +270,9 @@ fma(double x, double y, double z)
* rounding modes.
*/
fesetround(oround);
adj = r.lo + xy.lo;
/* work around clang bug 8100 */
volatile double vrlo = r.lo;
adj = vrlo + xy.lo;
return (ldexp(r.hi + adj, spread));
}

View File

@ -226,6 +226,8 @@ fmal(long double x, long double y, long double z)
zs = copysignl(LDBL_MIN, zs);
fesetround(FE_TONEAREST);
/* work around clang bug 8100 */
volatile long double vxs = xs;
/*
* Basic approach for round-to-nearest:
@ -235,7 +237,7 @@ fmal(long double x, long double y, long double z)
* adj = xy.lo + r.lo (inexact; low bit is sticky)
* result = r.hi + adj (correctly rounded)
*/
xy = dd_mul(xs, ys);
xy = dd_mul(vxs, ys);
r = dd_add(xy.hi, zs);
spread = ex + ey;
@ -256,7 +258,9 @@ fmal(long double x, long double y, long double z)
* rounding modes.
*/
fesetround(oround);
adj = r.lo + xy.lo;
/* work around clang bug 8100 */
volatile long double vrlo = r.lo;
adj = vrlo + xy.lo;
return (ldexpl(r.hi + adj, spread));
}

View File

@ -96,6 +96,7 @@ Lp6 = 1.531383769920937332e-01, /* 3FC39A09 D078C69F */
Lp7 = 1.479819860511658591e-01; /* 3FC2F112 DF3E5244 */
static const double zero = 0.0;
static volatile double vzero = 0.0;
double
log1p(double x)
@ -109,7 +110,7 @@ log1p(double x)
k = 1;
if (hx < 0x3FDA827A) { /* 1+x < sqrt(2)+ */
if(ax>=0x3ff00000) { /* x <= -1.0 */
if(x==-1.0) return -two54/zero; /* log1p(-1)=+inf */
if(x==-1.0) return -two54/vzero; /* log1p(-1)=+inf */
else return (x-x)/(x-x); /* log1p(x<-1)=NaN */
}
if(ax<0x3e200000) { /* |x| < 2**-29 */

View File

@ -34,6 +34,7 @@ Lp6 = 1.5313838422e-01, /* 3E1CD04F */
Lp7 = 1.4798198640e-01; /* 3E178897 */
static const float zero = 0.0;
static volatile float vzero = 0.0;
float
log1pf(float x)
@ -47,7 +48,7 @@ log1pf(float x)
k = 1;
if (hx < 0x3ed413d0) { /* 1+x < sqrt(2)+ */
if(ax>=0x3f800000) { /* x <= -1.0 */
if(x==(float)-1.0) return -two25/zero; /* log1p(-1)=+inf */
if(x==(float)-1.0) return -two25/vzero; /* log1p(-1)=+inf */
else return (x-x)/(x-x); /* log1p(x<-1)=NaN */
}
if(ax<0x38000000) { /* |x| < 2**-15 */

View File

@ -36,12 +36,16 @@ __FBSDID("$FreeBSD$");
* instead of feclearexcept()/feupdateenv() to restore the environment
* because the only exception defined for rint() is overflow, and
* rounding can't overflow as long as emax >= p.
*
* The volatile keyword is needed below because clang incorrectly assumes
* that rint won't raise any floating-point exceptions. Declaring ret volatile
* is sufficient to trick the compiler into doing the right thing.
*/
#define DECL(type, fn, rint) \
type \
fn(type x) \
{ \
type ret; \
volatile type ret; \
fenv_t env; \
\
fegetenv(&env); \

View File

@ -76,6 +76,12 @@ __FBSDID("$FreeBSD$");
test((func), (x), (y), (z), (rz), (exceptmask), (excepts)); \
} while (0)
/*
* This is needed because clang constant-folds fma in ways that are incorrect
* in rounding modes other than FE_TONEAREST.
*/
volatile double one = 1.0;
/*
* Determine whether x and y are equal, with two special rules:
* +0.0 != -0.0
@ -108,9 +114,9 @@ test_zeroes(void)
testall(-0.0, 0.0, -0.0, -0.0, ALL_STD_EXCEPT, 0);
testall(0.0, -0.0, -0.0, -0.0, ALL_STD_EXCEPT, 0);
testall(-1.0, 1.0, 1.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
testall(1.0, -1.0, 1.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
testall(-1.0, -1.0, -1.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
testall(-one, one, one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
testall(one, -one, one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
testall(-one, -one, -one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
switch (fegetround()) {
case FE_TONEAREST:
@ -190,53 +196,53 @@ test_small_z(void)
/* x*y positive, z positive */
if (fegetround() == FE_UPWARD) {
test(fmaf, 1.0, 1.0, 0x1.0p-100, 1.0 + FLT_EPSILON,
test(fmaf, one, one, 0x1.0p-100, 1.0 + FLT_EPSILON,
ALL_STD_EXCEPT, FE_INEXACT);
test(fma, 1.0, 1.0, 0x1.0p-200, 1.0 + DBL_EPSILON,
test(fma, one, one, 0x1.0p-200, 1.0 + DBL_EPSILON,
ALL_STD_EXCEPT, FE_INEXACT);
test(fmal, 1.0, 1.0, 0x1.0p-200, 1.0 + LDBL_EPSILON,
test(fmal, one, one, 0x1.0p-200, 1.0 + LDBL_EPSILON,
ALL_STD_EXCEPT, FE_INEXACT);
} else {
testall(0x1.0p100, 1.0, 0x1.0p-100, 0x1.0p100,
testall(0x1.0p100, one, 0x1.0p-100, 0x1.0p100,
ALL_STD_EXCEPT, FE_INEXACT);
}
/* x*y negative, z negative */
if (fegetround() == FE_DOWNWARD) {
test(fmaf, -1.0, 1.0, -0x1.0p-100, -(1.0 + FLT_EPSILON),
test(fmaf, -one, one, -0x1.0p-100, -(1.0 + FLT_EPSILON),
ALL_STD_EXCEPT, FE_INEXACT);
test(fma, -1.0, 1.0, -0x1.0p-200, -(1.0 + DBL_EPSILON),
test(fma, -one, one, -0x1.0p-200, -(1.0 + DBL_EPSILON),
ALL_STD_EXCEPT, FE_INEXACT);
test(fmal, -1.0, 1.0, -0x1.0p-200, -(1.0 + LDBL_EPSILON),
test(fmal, -one, one, -0x1.0p-200, -(1.0 + LDBL_EPSILON),
ALL_STD_EXCEPT, FE_INEXACT);
} else {
testall(0x1.0p100, -1.0, -0x1.0p-100, -0x1.0p100,
testall(0x1.0p100, -one, -0x1.0p-100, -0x1.0p100,
ALL_STD_EXCEPT, FE_INEXACT);
}
/* x*y positive, z negative */
if (fegetround() == FE_DOWNWARD || fegetround() == FE_TOWARDZERO) {
test(fmaf, 1.0, 1.0, -0x1.0p-100, 1.0 - FLT_EPSILON / 2,
test(fmaf, one, one, -0x1.0p-100, 1.0 - FLT_EPSILON / 2,
ALL_STD_EXCEPT, FE_INEXACT);
test(fma, 1.0, 1.0, -0x1.0p-200, 1.0 - DBL_EPSILON / 2,
test(fma, one, one, -0x1.0p-200, 1.0 - DBL_EPSILON / 2,
ALL_STD_EXCEPT, FE_INEXACT);
test(fmal, 1.0, 1.0, -0x1.0p-200, 1.0 - LDBL_EPSILON / 2,
test(fmal, one, one, -0x1.0p-200, 1.0 - LDBL_EPSILON / 2,
ALL_STD_EXCEPT, FE_INEXACT);
} else {
testall(0x1.0p100, 1.0, -0x1.0p-100, 0x1.0p100,
testall(0x1.0p100, one, -0x1.0p-100, 0x1.0p100,
ALL_STD_EXCEPT, FE_INEXACT);
}
/* x*y negative, z positive */
if (fegetround() == FE_UPWARD || fegetround() == FE_TOWARDZERO) {
test(fmaf, -1.0, 1.0, 0x1.0p-100, -1.0 + FLT_EPSILON / 2,
test(fmaf, -one, one, 0x1.0p-100, -1.0 + FLT_EPSILON / 2,
ALL_STD_EXCEPT, FE_INEXACT);
test(fma, -1.0, 1.0, 0x1.0p-200, -1.0 + DBL_EPSILON / 2,
test(fma, -one, one, 0x1.0p-200, -1.0 + DBL_EPSILON / 2,
ALL_STD_EXCEPT, FE_INEXACT);
test(fmal, -1.0, 1.0, 0x1.0p-200, -1.0 + LDBL_EPSILON / 2,
test(fmal, -one, one, 0x1.0p-200, -1.0 + LDBL_EPSILON / 2,
ALL_STD_EXCEPT, FE_INEXACT);
} else {
testall(-0x1.0p100, 1.0, 0x1.0p-100, -0x1.0p100,
testall(-0x1.0p100, one, 0x1.0p-100, -0x1.0p100,
ALL_STD_EXCEPT, FE_INEXACT);
}
}