Re: [PATCH v2 3/3] x86: Do not raises floating-point exception traps on fesetexceptflag (BZ 30990)

[Bruno Haible <bruno@clisp.org>]

[-- Attachment #1: Type: text/plain, Size: 731 bytes --]

Hi Adhemerval,

Like in the PATCH v2 2/3, in the error case, the new code mistakenly masks
all floating-point exceptions (i.e. as if someone had called
fedisableexcept (FE_ALL_EXCEPT)).
This is because the FNSTENV instruction is documented as
   "Saves the current FPU operating environment at the memory
    location specified with the destination operand, and then
    masks all floating-point exceptions."

The mistake came from my initial proposed fix. Sorry about that.

Here's a proposed fix, on top of your patch. I've verified that the sequence
of instructions
    __asm__ ("fnstenv %0" : "=m" (*&temp));
    __asm__ volatile ("fldcw %0" : : "m" (*&temp.__control_word));
does restore the exceptions trapping bits.

Bruno


[-- Attachment #2: BZ30990-fnstenv-fix.diff --]
[-- Type: text/x-patch, Size: 1478 bytes --]

diff --git a/sysdeps/i386/fpu/fsetexcptflg.c b/sysdeps/i386/fpu/fsetexcptflg.c
index ccbcf35e8e..8bff9026a1 100644
--- a/sysdeps/i386/fpu/fsetexcptflg.c
+++ b/sysdeps/i386/fpu/fsetexcptflg.c
@@ -36,6 +36,8 @@ __fesetexceptflag (const fexcept_t *flagp, int excepts)
   /* Get the current x87 FPU environment.  We have to do this since we
      cannot separately set the status word.  */
   __asm__ ("fnstenv %0" : "=m" (*&temp));
+  /* Note: fnstenv masks all floating-point exceptions until the fldenv
+     or fldcw below.  */
 
   if (CPU_FEATURE_USABLE (SSE))
     {
@@ -62,10 +64,13 @@ __fesetexceptflag (const fexcept_t *flagp, int excepts)
       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;
+        {
+          /* 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.  */
+          __asm__ volatile ("fldcw %0" : : "m" (*&temp.__control_word));
+          return -1;
+        }
 
       /* Store the new status word (along with the rest of the environment).  */
       __asm__ ("fldenv %0" : : "m" (*&temp));