CLANG specific issues.

[munroesj52]

• Can not tell what is up with Float128 support. FLOAT128 is not defined.
  • ./pveclib/vec_f128_ppc.h:209:17: error: cannot combine with previous declaration specifier
  • typedef vf128_t __float128;
  • Clang claims to support __float128 but may require explicit -mfloat128 for use.
    • Clang seems to require both -mcpu=power9 and -mfloat128 for clang-8
    • This defines FLOAT128 and enables the __float128 type
    • Otherwise you are not allowed to use or redefine (typedef) __float128
  • But Clang (attempts) to use the GCC/GLIBC installed math.h, stdlib.h ... headers which will get compile errors if -mcpu=power9 -mfloat128 is enabled.
    • Basically clang chokes on code in bits/floatn.h what uses attribute ((mode (KC)))
• Seems to be still using long double == _IBM128 but the type constrains do not seem to be working.
• No Decimal FP support. Had to disable BCD.
  • Can't even use in-line asm because clang does no recognize the constraints for FPR pairs (FPRp) required for _Decimal128.
• No constant folding for __int128 for example:
  • (__int128 ) 10ll * (__int128 ) 1000000000000000000ll, /* 10**19 */
  • works on GCC for both scalar and vector initializers.
• No support for vector bool _int128. GCC supports.
• Does not support #pagma GCC target
• vec_ld () does not handle vector __int128 (as specified in the Power Vector Intrinsic Programming Reference). CLang only supports the original VMX types. GCC support vector __int128 and long long.
• Clang does not support __builtin_unpack_longdouble, should be a direct register transfer.
  • Again can't use in-line asm because clang does not support the FPR pair constaints
• Code gen bug in clang8 for vec_sra vector long long.
• Clang8 for -mcpu=power9 does not support inline asm for vmsumud which is a valid instruction and mnemonic. This cripples the implementation of quadword and multi-quadword precision multiply.
• Before clang 9 clang miss assembled the DCMX of the Test Data Class instructions.

[munroesj52]

Clang can also be more picky about pointer types in ASM "m" and "Z" constraints. GCC does not care as long as it can see the EA calculation. For example:

"=Z" (*((char *)rb+rt))

Is OK for GCC, but will generate a char storage copy for Clang. Clang seems to need pointer type like:

"=Z" ((double)((char *)rb+rt))

[ecnelises]

Clang 14 (next major release, estimated to be available early next year) will support:

• __ibm128 type for traditional double-double format on Power
• mode(IF)/mode(IC) for forced double-double format
• -mfloat128 will be enabled by default for VSX targets (Power7 and later)
• long double may switch to __float128 as default semantics, and there seems to be similar changes in GCC
  • In clang 9 and later, -mabi=ieeelongdouble/ibmlongdouble is available to switch long double semantics, so we cannot expect long double is __ibm128
  • Use __LONG_DOUBLE_IEEE128__ and __LONG_DOUBLE_IBM128__ to tell current long double semantics (available since clang 12)

[munroesj52]

Thanks!

…

On Mon, Oct 11, 2021 at 1:05 AM Qiu Chaofan ***@***.***> wrote: Clang 14 (next major release, estimated to be available early next year) will support: - __ibm128 type for traditional double-double format on Power - mode(IF)/mode(IC) for forced double-double format - -mfloat128 will be enabled by default for VSX targets (Power7 and later) - long double may switch to __float128 as default semantics, and there seems to be similar changes in GCC - In clang 9 and later, -mabi=ieeelongdouble/ibmlongdouble is available to switch long double semantics, so we cannot expect long double is __ibm128 - Use __LONG_DOUBLE_IEEE128__ and __LONG_DOUBLE_IBM128__ to tell current long double semantics (available since clang 12) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#94 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIEI2IJPIPEC6ZAV2K3MAWLUGJ5CFANCNFSM4M6HH7FA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[munroesj52]

So it looks like Clang 12 with -mfloat128 will define __FLOAT128__ but will not let you define or use __float128 or __ieee128. Does _Float128 work as a type in C? It would be helpful if there was a type I could use for unions and Q const values. As is I have to use vector unsigned __int128 to typedef __binary128 as a work around and all constants are in hex.

…

On Mon, Oct 11, 2021 at 1:05 AM Qiu Chaofan ***@***.***> wrote: Clang 14 (next major release, estimated to be available early next year) will support: - __ibm128 type for traditional double-double format on Power - mode(IF)/mode(IC) for forced double-double format - -mfloat128 will be enabled by default for VSX targets (Power7 and later) - long double may switch to __float128 as default semantics, and there seems to be similar changes in GCC - In clang 9 and later, -mabi=ieeelongdouble/ibmlongdouble is available to switch long double semantics, so we cannot expect long double is __ibm128 - Use __LONG_DOUBLE_IEEE128__ and __LONG_DOUBLE_IBM128__ to tell current long double semantics (available since clang 12) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#94 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIEI2IJPIPEC6ZAV2K3MAWLUGJ5CFANCNFSM4M6HH7FA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[ecnelises]

So it looks like Clang 12 with -mfloat128 will define FLOAT128 but will
not let you define or use __float128 or __ieee128.

Clang doesn't have __ieee128, but __float128 (and 'q/Q' as constant suffix) is available under -mfloat128 -mcpu=power9.

Does _Float128 work as a type in C?

Currently not. There is patching adding this (https://reviews.llvm.org/D111382) while some reviewers are against it (making difference between C and C++).