cc Aldy Hernandez <aldyh@redhat.com> and Andrew MacLeod <amacleod@redhat.com>,
they are author of new VRP analysis for GCC, just to make sure I didn't
mis-understanding or mis-interpreting anything on GCC site.
GCC 11 have better value range analysis, that give GCC more confidence
to perform more aggressive optimization, but it cause scalbn/scalbnf get
wrong result.
Using scalbn to demostrate what happened on GCC 11, see comments with VRP
prefix:
```c
double scalbn (double x, int n)
{
/* VRP RESULT: n = [-INF, +INF] */
__int32_t k,hx,lx;
...
k = (hx&0x7ff00000)>>20;
/* VRP RESULT: k = [0, 2047] */
if (k==0) {
/* VRP RESULT: k = 0 */
...
k = ((hx&0x7ff00000)>>20) - 54;
if (n< -50000) return tiny*x; /*underflow*/
/* VRP RESULT: k = -54 */
}
/* VRP RESULT: k = [-54, 2047] */
if (k==0x7ff) return x+x; /* NaN or Inf */
/* VRP RESULT: k = [-54, 2046] */
k = k+n;
if (k > 0x7fe) return huge*copysign(huge,x); /* overflow */
/* VRP RESULT: k = [-INF, 2046] */
/* VRP RESULT: n = [-INF, 2100],
because k + n <= 0x7fe is false, so:
1. -INF < [-54, 2046] + n <= 0x7fe(2046) < INF
2. -INF < [-54, 2046] + n <= 2046 < INF
3. -INF < n <= 2046 - [-54, 2046] < INF
4. -INF < n <= [0, 2100] < INF
5. n = [-INF, 2100] */
if (k > 0) /* normal result */
{SET_HIGH_WORD(x,(hx&0x800fffff)|(k<<20)); return x;}
if (k <= -54) {
/* VRP OPT: Evaluate n > 50000 as true...*/
if (n > 50000) /* in case integer overflow in n+k */
return huge*copysign(huge,x); /*overflow*/
else return tiny*copysign(tiny,x); /*underflow*/
}
k += 54; /* subnormal result */
SET_HIGH_WORD(x,(hx&0x800fffff)|(k<<20));
return x*twom54;
}
```
However give the input n = INT32_MAX, k = k+n will overflow, and then we
expect got `huge*copysign(huge,x)`, but new VRP optimization think
`n > 50000` is never be true, so optimize that into `tiny*copysign(tiny,x)`.
so the solution here is to moving the overflow handle logic before `k = k + n`.
- GCC will set __FLT_EVAL_METHOD__ to 16 if __fp16 supported, e.g.
cortex-a55/aarch64.
- $ aarch64-unknown-elf-gcc -v 2>&1 |grep version
gcc version 9.2.0 (GCC)
- $ aarch64-unknown-elf-gcc -E -dM -mcpu=cortex-a55 - < /dev/null |grep FLT_EVAL_METHOD
#define __FLT_EVAL_METHOD__ 16
#define __FLT_EVAL_METHOD_TS_18661_3__ 16
#define __FLT_EVAL_METHOD_C99__ 16
- The behavior of __FLT_EVAL_METHOD__ == 16 is same as
__FLT_EVAL_METHOD__ == 0 except for float16_t, but newlib didn't
support float16_t.
ISO/IEC TS 18661-3:
http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2405.pdf
V2 Changes:
- List Howland, Craig D as co-author since he provide the draft of comment
in math.h.
Co-authored-by: "Howland, Craig D" <howland@LGSInnovations.com>
The C standard says that errno may acquire the value ERANGE if the
result from strtod underflows. According to IEEE 754, underflow occurs
whenever the value cannot be represented in normalized form.
Newlib is inconsistent in this, setting errno to ERANGE only if the
value underflows to zero, but not for denorm values, and never for hex
format floats.
This patch attempts to consistently set errno to ERANGE for all
'underflow' conditions, which is to say all values which are not
exactly zero and which cannot be represented in normalized form.
This matches glibc behavior, as well as the Linux, Mac OS X, OpenBSD,
FreeBSD and SunOS strtod man pages.
Signed-off-by: Keith Packard <keithp@keithp.com>
The scanf code was skipping the '0' after the 'x' causing the
resulting buffer to contain an invalid number when passed to strtod.
Signed-off-by: Keith Packard <keithp@keithp.com>
- compiler is sometimes optimizing out the rounding check in
e_sqrt.c and ef_sqrt.c which uses two constants to create
an inexact operation
- there is a similar constant operation in s_tanh.c/sf_tanh.c
- make the one and tiny constants volatile to stop this
Newlib for aarch64 uses libgloss for the backend. One common libgloss
implementation is the 'rdimon' implementation, which uses the Arm
Semihosting protocol. In order to support a remote host that runs on
Windows we need to know whether a file is to be opened in binary or
text mode. That means that we need to preserve this information via
O_BINARY until we know what the libgloss binding will be.
This patch simply copies the arm implementation from sys/arm/sys and
puts it in machine/aarch64/sys, because we don't have a 'sys' subtree
on aarch64.
The only reason why it is tough for us to use nano malloc
is because of the small shortcoming where nano_malloc()
splits a bigger chunk from the free list into two pieces
while handing back the second one (the tail) to the user.
This is error prone and especially bad for smaller heaps,
where nano malloc is supposed to be superior. The normal
malloc doesn't have this issue and we need to use it even
though it costs us ~2k bytes compared to nano-malloc.
The problem arise especially after giving back _every_
malloced memory to the heap and then starting to exercise
the heap again by allocating something small. This small
item might split the whole heap in two equally big parts
depending on how the heap has been exercised before.
I have uploaded the smallest possible application
(only tested on ST and Nordic devices) to show the issue
while the real customer applications are far more complicated:
https://drive.google.com/file/d/1kfSC2KOm3Os3mI7EBd-U0j63qVs8xMbt/view?usp=sharing
The application works like the following pseudo code,
where we assume a heap of 100 bytes
(I haven't taken padding and other nitty and gritty
details into account. Everything to simplify understanding):
void *ptr = malloc(52); // We get 52 bytes and we have
// 48 bytes to use.
free(ptr); // Hand back the 52 bytes to nano_malloc
// This is the magic line that shows the issue of
// nano_malloc
ptr = malloc(1); // Nano malloc will split the 52 bytes
// in the free list and hand you a pointer
// somewhere in the
// middle of the heap.
ptr2 = malloc(52); // Out of memory...
I have done a fix which hands back the first part of the
splitted chunk. Once this is fixed we obviously
have the 1 byte placed in position 0 of the heap instead
of somewhere in the middle.
However, this won't let us malloc 52 new bytes even though
we potentially have 99 bytes left to use in the heap. The
reason is that when we try to do the allocation,
nano-malloc looks into the free list and sees a 51 byte
chunk to be used.
This is not big enough so nano-malloc decides to call
sbrk for _another_ 52 bytes which is not possible since
there is only 48 bytes left to ask for.
The solution for this problem is to check if the last
item in the free list is adjacent to sbrk(0). If it is,
as it is in this case, we can just ask sbrk for the
remainder of what is needed. In this case 1 byte.
NB! I have only tested the solution on our ST device.
Use the more official fesetenv(FE_DFL_ENV) from _dll_crt0, thus
allowing to drop the _feinitialise declaration from fenv.h.
Provide a no-op _feinitialise in Cygwin as exportable symbol for really
old applications when _feinitialise was called from mainCRTStartup in
crt0.o.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Drop the Cygwin-specific fenv.cc and fenv.h file and use the equivalent
newlib functionality now, so we have at least one example of a user for
this new mechanism.
fenv.c: allow _feinitialise to be called from Cygwin startup code
fenv.h: add declarations for fegetprec and fesetprec for Cygwin only.
Fix a comment.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Add a directory libc/machine/shared_x86 to share header files
between ix86 and x86_64 architectures.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
So far the build mechanism in newlib only allowed to either define
machine-specific headers, or headers shared between all machines.
In some cases, architectures are sufficiently alike to share header
files between them, but not with other architectures. A good example
is ix86 vs. x86_64, which share certain traits with each other, but
not with other architectures.
Introduce a new configure variable called "shared_machine_dir". This
dir can then be used for headers shared between architectures.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Without this, for a bare-iron/simulator target such as cris-elf,
you'll see, at newlib build time:
/x/gccobj/./gcc/xgcc -B/x/gccobj/./gcc/ <many options elided> -c -o lib_a-aligned_alloc.o \
`test -f 'aligned_alloc.c' || echo '/y/newlib/libc/stdlib/'`aligned_alloc.c
/y/newlib/libc/stdlib/aligned_alloc.c: In function 'aligned_alloc':
/y/newlib/libc/stdlib/aligned_alloc.c:35:10: warning: implicit declaration of function \
'_memalign_r' [-Wimplicit-function-declaration]
35 | return _memalign_r (_REENT, align, size);
| ^~~~~~~~~~~
The revert-part of the revert-and-fix commit, b99887c428 a.k.a.
"Revert previous change to sys/stat.h and fix cris libgloss",
apparently intending to revert f75aa67851 a.k.a. "Fix regression in
cris-elf caused by sys/stat.h change" and fix it in another way,
wasn't complete. Although the fix-part added the prerequisite "#undef
st_atime" (et al) to gensyscalls, the revert-part didn't revert the
"&& !defined(__cris__)" in sys/stat.h, stopping st_atime (et al) from
being defined.
The effect of the unreverted change is that accessing the struct stat
compatibility member names "st_atime" (et al) as in "struct stat
mystat; mystat.st_atime;" yields errors, observable for example when
building libgfortran in gcc:
/x/gcc/libgfortran/intrinsics/stat.c:114:42: error: 'struct stat' has \
no member named 'st_atime'; did you mean 'st_atim'?
114 | sarray->base_addr[8 * stride] = sb.st_atime;
| ^~~~~~~~
| st_atim
(etc.)
Trivially fixed by completing the reversion, removing the "&&
!defined(__cris__)" in sys/stat.h.
Beware: the net effect of the earlier related change to struct stat in
sys/stat.h, leading up to the fix, *does* change its definition as a
type. Thankfully, replacing members like "time_t st_atime; long
st_spare1;" by "struct timespec st_atim;", ditto st_mtim and st_ctim,
is layout-compatible. To wit, that change is "binary compatible".
Incidentally, related to the simulator / Linux ABI, there's a
transitional stage (see gensyscalls), reloading between "struct stat"
(sys/stat.h) and "struct new_stat" (kernel/simulator) as necessary.
Tested by a cris-elf gcc build (including libgfortran).
This Patch removes Soft Float code from MIPS.
Instead It adds the soft float code from RISCV
The code came from FreeBSD and assumes the FreeBSD softfp
implementation not the one with GCC. That was an overlooked and
fixed in the other fenv code already.
Signed-off-by: Eshan Dhawan <eshandhawan51@gmail.com>
Change the prototypes to be in line with POSIX/glibc. This may fix
issues with new warnings produced by GCC 11.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
Add the POSIX header file <poll.h> which is used by the GCC 11 Ada
runtime support.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
The Cortex-R52 processor is an Armv8-R processor with a NEON unit. This
fix prevents conflicting architecture profiles A/R errors issued by the
linker.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
This patch fixes the error found by Paul Zimmermann (see
https://homepages.loria.fr/PZimmermann/papers/#accuracy) regarding x
close to 1 and rather large y (specifically he found the case
powf(0x1.ffffeep-1,-0x1.000002p+27) which returns +Inf instead of the
correct value). We found 2 more values for x which show the same faulty
behaviour, and all 3 are fixed with this patch. We have tested all
combinations for x in [+1.fffdfp-1, +1.00020p+0] and y in
[-1.000007p+27, -1.000002p+27] and [1.000002p+27,1.000007p+27].
The overflow check in mEMALIGn erroneously checks for INT_MAX,
albeit the input parameter is size_t. Fix this to check for
__SIZE_MAX__ instead. Also, it misses to check the req against
adding the alignment before calling mALLOc.
While at it, add out-of-bounds checks to pvALLOc, nano_memalign,
nano_valloc, and Cygwin's (unused) dlpvalloc.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Kito Cheng
Keith Packard
Thomas Wolff
Jonathan Wakely
Joel Sherrill
Dimitar Dimitrov
Jeff Johnston
Richard Earnshaw
Ola Olsson
Mike Frysinger
Ties Stuij
Corinna Vinschen
David Macek
Marcus Comstedt
Hans-Peter Nilsson
Eshan dhawan
Sebastian Huber
Paul Zimmermann
Fabian Schriever
Corinna Vinschen
Ivan Grokhotov