[3/3] x86: Do not raises floating-point exception traps on fesetexceptflag (BZ 30990)
Checks
Context |
Check |
Description |
redhat-pt-bot/TryBot-apply_patch |
success
|
Patch applied to master at the time it was sent
|
linaro-tcwg-bot/tcwg_glibc_check--master-arm |
success
|
Testing passed
|
redhat-pt-bot/TryBot-32bit |
success
|
Build for i686
|
linaro-tcwg-bot/tcwg_glibc_build--master-aarch64 |
success
|
Testing passed
|
linaro-tcwg-bot/tcwg_glibc_check--master-aarch64 |
success
|
Testing passed
|
linaro-tcwg-bot/tcwg_glibc_build--master-arm |
success
|
Testing passed
|
Commit Message
From: Bruno Haible <bruno@clisp.org>
According to ISO C23 (7.6.4.5), fesetexceptflag is supposed to set
floating-point exception flags without raising a trap.
The flags can be set in the 387 unit or in the SSE unit. When we need
to clear a flag, we need to do so in both units, due to the way
fetestexcept is implemented.
When we need to set a flag, it is sufficient to do it in the SSE unit,
because that is guaranteed to not trap. However, on i386 CPUs that have
only a 387 unit, set the flags in the 387, as long as this cannot trap.
Co-authored-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
---
math/test-fexcept-traps.c | 11 ++++++
sysdeps/i386/fpu/fsetexcptflg.c | 58 ++++++++++++++++++++-----------
sysdeps/x86_64/fpu/fsetexcptflg.c | 24 +++++++------
3 files changed, 62 insertions(+), 31 deletions(-)
Comments
Adhemerval Zanella wrote:
> diff --git a/math/test-fexcept-traps.c b/math/test-fexcept-traps.c
> index 9b8f583ae6..a486d17951 100644
> --- a/math/test-fexcept-traps.c
> +++ b/math/test-fexcept-traps.c
> @@ -67,6 +68,16 @@ do_test (void)
> where setting the exception might result in traps the function should
> return a nonzero value. */
> ret = fesetexceptflag (&saved, FE_ALL_EXCEPT);
> +
> + /* Execute some floating-point operations, since on some CPUs exceptions
> + triggers a trap only at the next floating-point instruction. */
> + double a = 1.0;
> + double b = a + a;
> + math_force_eval (b);
> + long double al = 1.0L;
> + long double bl = al + al;
Like in [2/3], I would mark the variables a, b, al, bl as 'volatile',
otherwise it's too easy for GCC to do constant-folding and thus eliminate
the floating-point operations.
Bruno
On 23/10/23 21:19, Bruno Haible wrote:
> Adhemerval Zanella wrote:
>> diff --git a/math/test-fexcept-traps.c b/math/test-fexcept-traps.c
>> index 9b8f583ae6..a486d17951 100644
>> --- a/math/test-fexcept-traps.c
>> +++ b/math/test-fexcept-traps.c
>> @@ -67,6 +68,16 @@ do_test (void)
>> where setting the exception might result in traps the function should
>> return a nonzero value. */
>> ret = fesetexceptflag (&saved, FE_ALL_EXCEPT);
>> +
>> + /* Execute some floating-point operations, since on some CPUs exceptions
>> + triggers a trap only at the next floating-point instruction. */
>> + double a = 1.0;
>> + double b = a + a;
>> + math_force_eval (b);
>> + long double al = 1.0L;
>> + long double bl = al + al;
>
> Like in [2/3], I would mark the variables a, b, al, bl as 'volatile',
> otherwise it's too easy for GCC to do constant-folding and thus eliminate
> the floating-point operations.
Ack.
@@ -19,6 +19,7 @@
#include <fenv.h>
#include <stdio.h>
#include <math-tests.h>
+#include <math-barriers.h>
static int
do_test (void)
@@ -67,6 +68,16 @@ do_test (void)
where setting the exception might result in traps the function should
return a nonzero value. */
ret = fesetexceptflag (&saved, FE_ALL_EXCEPT);
+
+ /* Execute some floating-point operations, since on some CPUs exceptions
+ triggers a trap only at the next floating-point instruction. */
+ double a = 1.0;
+ double b = a + a;
+ math_force_eval (b);
+ long double al = 1.0L;
+ long double bl = al + al;
+ math_force_eval (bl);
+
if (ret != 0 && !EXCEPTION_SET_FORCES_TRAP)
{
puts ("fesetexceptflag failed");
@@ -17,42 +17,58 @@
<https://www.gnu.org/licenses/>. */
#include <fenv.h>
-#include <math.h>
-#include <unistd.h>
#include <ldsodefs.h>
-#include <dl-procinfo.h>
int
__fesetexceptflag (const fexcept_t *flagp, int excepts)
{
- fenv_t temp;
+ /* The flags can be set in the 387 unit or in the SSE unit. When we need to
+ clear a flag, we need to do so in both units, due to the way fetestexcept
+ is implemented.
+ When we need to set a flag, it is sufficient to do it in the SSE unit,
+ because that is guaranteed to not trap. However, on i386 CPUs that have
+ only a 387 unit, set the flags in the 387, as long as this cannot trap. */
- /* Get the current environment. We have to do this since we cannot
- separately set the status word. */
- __asm__ ("fnstenv %0" : "=m" (*&temp));
+ fenv_t temp;
- temp.__status_word &= ~(excepts & FE_ALL_EXCEPT);
- temp.__status_word |= *flagp & excepts & FE_ALL_EXCEPT;
+ excepts &= FE_ALL_EXCEPT;
- /* Store the new status word (along with the rest of the environment.
- Possibly new exceptions are set but they won't get executed unless
- the next floating-point instruction. */
- __asm__ ("fldenv %0" : : "m" (*&temp));
+ /* Get the current x87 FPU environment. We have to do this since we
+ cannot separately set the status word. */
+ __asm__ ("fnstenv %0" : "=m" (*&temp));
- /* If the CPU supports SSE, we set the MXCSR as well. */
if (CPU_FEATURE_USABLE (SSE))
{
- unsigned int xnew_exc;
+ unsigned int mxcsr;
+
+ /* Clear relevant flags. */
+ temp.__status_word &= ~(excepts & ~ *flagp);
- /* Get the current MXCSR. */
- __asm__ ("stmxcsr %0" : "=m" (*&xnew_exc));
+ /* Store the new status word (along with the rest of the environment). */
+ __asm__ ("fldenv %0" : : "m" (*&temp));
- /* Set the relevant bits. */
- xnew_exc &= ~(excepts & FE_ALL_EXCEPT);
- xnew_exc |= *flagp & excepts & FE_ALL_EXCEPT;
+ /* And now similarly for SSE. */
+ __asm__ ("stmxcsr %0" : "=m" (*&mxcsr));
+
+ /* Clear or set relevant flags. */
+ mxcsr ^= (mxcsr ^ *flagp) & excepts;
/* Put the new data in effect. */
- __asm__ ("ldmxcsr %0" : : "m" (*&xnew_exc));
+ __asm__ ("ldmxcsr %0" : : "m" (*&mxcsr));
+ }
+ else
+ {
+ /* Clear or set relevant flags. */
+ temp.__status_word ^= (temp.__status_word ^ *flagp) & excepts;
+
+ if ((~temp.__control_word) & temp.__status_word & excepts)
+ /* Setting the exception flags may trigger a trap (at the next
+ floating-point instruction, but that does not matter).
+ ISO C 23 § 7.6.4.5 does not allow it. */
+ return -1;
+
+ /* Store the new status word (along with the rest of the environment). */
+ __asm__ ("fldenv %0" : : "m" (*&temp));
}
/* Success. */
@@ -22,30 +22,34 @@
int
fesetexceptflag (const fexcept_t *flagp, int excepts)
{
+ /* The flags can be set in the 387 unit or in the SSE unit.
+ When we need to clear a flag, we need to do so in both units,
+ due to the way fetestexcept() is implemented.
+ When we need to set a flag, it is sufficient to do it in the SSE unit,
+ because that is guaranteed to not trap. */
+
fenv_t temp;
unsigned int mxcsr;
- /* XXX: Do we really need to set both the exception in both units?
- Shouldn't it be enough to set only the SSE unit? */
+ excepts &= FE_ALL_EXCEPT;
/* Get the current x87 FPU environment. We have to do this since we
cannot separately set the status word. */
__asm__ ("fnstenv %0" : "=m" (*&temp));
- temp.__status_word &= ~(excepts & FE_ALL_EXCEPT);
- temp.__status_word |= *flagp & excepts & FE_ALL_EXCEPT;
+ /* Clear relevant flags. */
+ temp.__status_word &= ~(excepts & ~ *flagp);
- /* Store the new status word (along with the rest of the environment.
- Possibly new exceptions are set but they won't get executed unless
- the next floating-point instruction. */
+ /* Store the new status word (along with the rest of the environment). */
__asm__ ("fldenv %0" : : "m" (*&temp));
- /* And now the same for SSE. */
+ /* And now similarly for SSE. */
__asm__ ("stmxcsr %0" : "=m" (*&mxcsr));
- mxcsr &= ~(excepts & FE_ALL_EXCEPT);
- mxcsr |= *flagp & excepts & FE_ALL_EXCEPT;
+ /* Clear or set relevant flags. */
+ mxcsr ^= (mxcsr ^ *flagp) & excepts;
+ /* Put the new data in effect. */
__asm__ ("ldmxcsr %0" : : "m" (*&mxcsr));
/* Success. */