[Clang][CodeGen][AIX] Map __builtin_frexpl, __builtin_ldexpl, and __builtin_modfl to 'double' version lib calls in 64-bit 'long double' mode

[xingxue]

Summary

AIX library functions frexpl(), ldexpl(), and modfl() are for 128-bit IBM long double, i.e. __ibm128. Other *l() functions, e.g., acosl(), are for 64-bit long double. The AIX Clang compiler currently maps builtin functions __builtin_frexpl(), __builtin_ldexpl(), and __builtin_modfl() to frexpl(), ldexpl(), and modfl() in 64-bit long double mode which results in seg-faults or incorrect return values. This patch changes to map __builtin_frexpl(), __builtin_ldexpl(), and __builtin_modfl() to double version lib functions frexp(), ldexp() and modf() in 64-bit long double mode. The following is from AIX <math.h>.

/*
 * frexpl(), ldexpl(), and modfl() have preexisting shared versions which are
 * 128-bit only.  64-bit versions must be made available for C99 for the
 * default 64-bit long double.  These cannot simply be macros because the
 * actual routines will be the incorrect form in 64-bit mode if the user
 * forces the actual routines to be used through undef or macro suppression
 * per the standard.
 */
#if defined(_ISOC99_SOURCE) && !defined(__LONGDOUBLE128)
static long double _NOTHROW(frexpl, (long double __x, int *__i))
{
     return (long double) frexp((double) __x, __i);
}

static long double _NOTHROW(ldexpl, (long double __x, int __i))
{
     return (long double) ldexp((double) __x, __i);
}

#ifndef __MODFL
static long double _NOTHROW(modfl, (long double __x, long double *__y))
{
     return (long double) modf((double) __x, (double *) __y);
}
#endif
#endif

[hubert.reinterpretcast]

clang/lib/CodeGen/CGBuiltin.cpp

112

Please rename to "AIXLongDouble64Builtins".

clang/test/CodeGen/aix-builtin-mapping.c

4

Fix typo.

6–7

This should be an IR test. Checking assembly means actually needing the PowerPC target enabled in the LLVM configuration.

[daltenty]

clang/lib/CodeGen/CGBuiltin.cpp

110

I feel like we should be clear about which 128-bit double format we are talking about, since it may not be immediately clear for folks who aren't intimately familiar with AIX. Maybe we can reference __ibm128 or ibm 128-bit, so it's clear what we mean.

clang/test/CodeGen/aix-builtin-mapping.c

3

[xingxue]

Addressed comments:

• rename AIXLongDoubleBuiltins to AIXLongDouble64Builtins
• test IR output instead of assembly output
• fix a typo in a comment
• mention frexpl(), ldexpl(), and modfl() are for 128-bit IBM long double, i.e. __ibm128

[xingxue]

clang/lib/CodeGen/CGBuiltin.cpp

110

Modified as suggested, thanks!

112

Renamed as suggested, thanks!

clang/test/CodeGen/aix-builtin-mapping.c

3

Modified as suggested, thanks!

4

Fixed, thanks for catching it!

6–7

Changed to use IR, thanks!

[hubert.reinterpretcast]

LGTM; thanks!

[daltenty]

LGTM, thanks!

clang/lib/CodeGen/CGBuiltin.cpp

static SmallDenseMap<unsigned, StringRef, 8> F128Builtins{
       {Builtin::BI__builtin_vsnprintf, "__vsnprintfieee128"},
       {Builtin::BI__builtin_vsprintf, "__vsprintfieee128"},
       {Builtin::BI__builtin_sprintf, "__sprintfieee128"},
       {Builtin::BI__builtin_snprintf, "__snprintfieee128"},
       {Builtin::BI__builtin_fprintf, "__fprintfieee128"},
       {Builtin::BI__builtin_nexttowardf128, "__nexttowardieee128"},
   };
 
+  // The AIX library functions frexpl, ldexpl, and modfl are for 128-bit
+  // IBM 'long double' (i.e. __ibm128). Map to the 'double' versions
+  // if it is 64-bit 'long double' mode.
+  static SmallDenseMap<unsigned, StringRef, 4> AIXLongDouble64Builtins{
+      {Builtin::BI__builtin_frexpl, "frexp"},
+      {Builtin::BI__builtin_ldexpl, "ldexp"},
+      {Builtin::BI__builtin_modfl, "modf"},
+  };
+
   // If the builtin has been declared explicitly with an assembler label,
   // use the mangled name. This differs from the plain label on platforms
   // that prefix labels.
   if (FD->hasAttr<AsmLabelAttr>())
     Name = getMangledName(D);
   else {
     // TODO: This mutation should also be applied to other targets other than
     // PPC, after backend supports IEEE 128-bit style libcalls.
     if (getTriple().isPPC64() &&
         &getTarget().getLongDoubleFormat() == &llvm::APFloat::IEEEquad() &&
         F128Builtins.find(BuiltinID) != F128Builtins.end())
       Name = F128Builtins[BuiltinID];
+    else if (getTriple().isOSAIX() &&
+             &getTarget().getLongDoubleFormat() ==
+                 &llvm::APFloat::IEEEdouble() &&
+             AIXLongDouble64Builtins.find(BuiltinID) !=
+                 AIXLongDouble64Builtins.end())
+      Name = AIXLongDouble64Builtins[BuiltinID];
     else
       Name = Context.BuiltinInfo.getName(BuiltinID) + 10;
   }
 
   llvm::FunctionType *Ty =
     cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
 
   return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);

clang/test/CodeGen/aix-builtin-mapping.c

+// AIX library functions frexpl, ldexpl, and modfl are for 128-bit IBM
+// 'long double' (i.e. __ibm128). Check that the compiler generates
+// calls to the 'double' versions for corresponding builtin functions in
+// 64-bit 'long double' mode.
+
+// RUN: %clang_cc1 -triple powerpc-ibm-aix -mlong-double-64 -emit-llvm -o - %s | FileCheck -check-prefix=CHECK %s
+// RUN: %clang_cc1 -triple powerpc64-ibm-aix -mlong-double-64 -emit-llvm -o - %s | FileCheck -check-prefix=CHECK %s
+
+int main()
+{
+  int DummyInt;
+  long double DummyLongDouble;
+  long double returnValue;
+
+  returnValue = __builtin_modfl(1.0L, &DummyLongDouble);
+  returnValue = __builtin_frexpl(0.0L, &DummyInt);
+  returnValue = __builtin_ldexpl(1.0L, 1);
+}
+
+// CHECK: %call = call double @modf(double noundef 1.000000e+00, ptr noundef %DummyLongDouble) #3
+// CHECK: %call1 = call double @frexp(double noundef 0.000000e+00, ptr noundef %DummyInt) #3
+// CHECK: %call2 = call double @ldexp(double noundef 1.000000e+00, i32 noundef {{(signext )?}}1) #4