[flang] Build fix

[vdonaldson]

Reinstate a preprocessor define for a nonstandard exception.

124978

[llvmbot]

@llvm/pr-subscribers-flang-runtime

Author: None (vdonaldson)

Changes

Reinstate a preprocessor define for a nonstandard exception.

124978

---

Full diff: https://github.com/llvm/llvm-project/pull/125087.diff

1 Files Affected:

• (modified) flang/runtime/exceptions.cpp (+12-7)

diff --git a/flang/runtime/exceptions.cpp b/flang/runtime/exceptions.cpp
index b41c7cf438f569..257c71b51edb3d 100644
--- a/flang/runtime/exceptions.cpp
+++ b/flang/runtime/exceptions.cpp
@@ -13,11 +13,14 @@
 #include <cfenv>
 #if defined(__aarch64__) && !defined(_WIN32)
 #include <fpu_control.h>
-#elif defined(__x86_64__)
+#elif defined(__x86_64__) && !defined(_WIN32)
 #include <xmmintrin.h>
 #endif
 
-// fenv.h may not define exception macros.
+// File fenv.h usually, but not always, defines standard exceptions as both
+// enumerator values and preprocessor #defines. Some x86 environments also
+// define a nonstandard __FE_DENORM enumerator, but without a corresponding
+// #define, which makes it more difficult to determine if it is present or not.
 #ifndef FE_INVALID
 #define FE_INVALID 0
 #endif
@@ -33,6 +36,12 @@
 #ifndef FE_INEXACT
 #define FE_INEXACT 0
 #endif
+#if FE_INVALID == 1 && FE_DIVBYZERO == 4 && FE_OVERFLOW == 8 && \
+    FE_UNDERFLOW == 16 && FE_INEXACT == 32
+#define __FE_DENORM 2
+#else
+#define __FE_DENORM 0
+#endif
 
 namespace Fortran::runtime {
 
@@ -44,11 +53,7 @@ uint32_t RTNAME(MapException)(uint32_t excepts) {
   Terminator terminator{__FILE__, __LINE__};
 
   static constexpr uint32_t v{FE_INVALID};
-#if __x86_64__
-  static constexpr uint32_t s{__FE_DENORM}; // nonstandard, not a #define
-#else
-  static constexpr uint32_t s{0};
-#endif
+  static constexpr uint32_t s{__FE_DENORM};
   static constexpr uint32_t z{FE_DIVBYZERO};
   static constexpr uint32_t o{FE_OVERFLOW};
   static constexpr uint32_t u{FE_UNDERFLOW};

[brooksdavis]

Seems plausible to me and likely to fix FreeBSD builds, but far from my area of expertise.

Thanks,
Brooks

[rorth]

Solaris/amd64 was broken the same. This patch restored the build.

[jeanPerier]

This "by elimination logic" makes some sense and is valid with my configs. I think if you wanted something more robust, you could use cmake to try compiling a a file using __FE_DENORM and define a macro according to the compilation success/failure. Then that macro could be used here.

[h-vetinari]

Can someone add this to the 20.x milestone please? Without that the following command will fail (and we need this for 20.1 release).

/cherry-pick 7fa1257

[h-vetinari]

Ah, nevermind, I just saw this was already requested elsewhere, leading to the PR: #125213

[llvmbot]

Can someone add this to the 20.x milestone please? Without that the following command will fail (and we need this for 20.1 release).

/cherry-pick 7fa1257

Error: Command failed due to missing milestone.