================
@@ -2707,27 +2707,31 @@ static bool checkFloatingPointResult(EvalInfo &Info,
const Expr *E,
if (Info.InConstantContext)
return true;
+ // The output result is exact and no exceptions are raised, so it is safe to
+ // perform compile-time evaluation.
+ if (St == APFloat::opOK)
+ return true;
+
FPOptions FPO = E->getFPFeaturesInEffect(Info.getLangOpts());
- if ((St & APFloat::opInexact) &&
- FPO.getRoundingMode() == llvm::RoundingMode::Dynamic) {
- // Inexact result means that it depends on rounding mode. If the requested
- // mode is dynamic, the evaluation cannot be made in compile time.
+
+ if (FPO.getRoundingMode() == llvm::RoundingMode::Dynamic) {
----------------
hubert-reinterpretcast wrote:
This logic breaks once `<math.h>` functions are considered.
For `<math.h>` functions, a constant rounding mode (which we ignore for folding
of run-time evaluations) should not allow folding that ignores the dynamic
rounding mode without regard to `-frounding-math`. Such folding occurs with
this PR in its current state:
```console
$ clang++ -w -frounding-math -S -o - -emit-llvm -Xclang -disable-llvm-passes
-xc -<<<'double exp1() { _Pragma("STDC FENV_ROUND FE_DOWNWARD") return
__builtin_exp(1.); }' | grep ret
ret double f0x4005BF0A8B145769
```
**Note 1:**
https://clang.llvm.org/docs/UsersManual.html#accessing-the-floating-point-environment
makes a statement about the C standard in direct contradiction to C23
subclause 7.6.2 paragraph 3. According to the C standard, `#pragma STDC
FENV_ROUND FE_DYNAMIC` affords no permission for the program to change the
dynamic rounding mode when `FENV_ACCESS` is "off".
**Note 2:** From the same section of the Clang documentation and with reference
to the same paragraph of the C standard,
> The command line option `-frounding-math` behaves as if the translation unit
> began with `#pragma STDC FENV_ROUND FE_DYNAMIC`.
is vacuous at best and incorrect at worst. Having the translation unit begin
that way has no effect according to the C standard, and that is true of the
Clang implementation; however, the option _does_ have an effect:
https://godbolt.org/z/d8bcfe4jx.
https://github.com/llvm/llvm-project/pull/199808
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
