testcae:
```
#pragma GCC target("sse4.1")
#include <immintrin.h>
int main(){return 0;}
```

Works on x86_64-linux-gnu with `-m32 -march=i686  -mno-sse`

_Float16 is supportted with target sse2, that's why it report an error.
It looks like you want to use just gpr intrinsics, for that case, x86gprintrin.h can be used.

From pr108883, maybe we shouldnot restrict _Float16 under TARGET_SSE2.

Jakub Jelinek 2023-02-22 12:21:24 UTC
Created attachment 54506 [details]
gcc13-pr108883.patch

Untested fix on the compiler side of emit_support_tinfos.

That said, these fundamental types whose presence/absence depends on ISA flags
are quite problematic IMHO, as they are incompatible with the target attribute/pragmas.  Whether they are available or not available depends on whether in this case SSE2 is enabled during compiler initialization (aka after
parsing command line options) and then they are available or unavailable to
everything else based on that.

Yeah.  Enable all the time and have say the
targetm.invalid_conversion, targetm.invalid_unary_op, targetm.invalid_binary_op
and something in argument/return value passing reject _Float16/__bf16 in functions without SSE2.
That will not be enough though, we'll need to arrange e.g. for the spot where we #undef/#define target  macros based on currently active ISA in pragmas to also
do that for __STDCPP_FLOAT16_T__ and __STDCPP_BFLOAT16_T__ for C++, and change libstdc++
such that for x86 it adds similarly to x86 intrin headers something like
#ifndef __SSE2__
#pragma GCC push_options
#pragma GCC target("sse2")
#define __DISABLE_SSE2__
#endif /* __SSE2__ */
...
#ifdef __DISABLE_SSE2__
#undef __DISABLE_SSE2__
#pragma GCC pop_options
#endif /* __DISABLE_SSE2__ */
around std::float16_t/std::bfloat16_t stuff.
Definitely not stage4 material I'm afraid.

Note, until then we could perhaps do something like:
--- gcc/config/i386/avx512fp16intrin.h.jj	2023-01-16 11:52:15.946736083 +0100
+++ gcc/config/i386/avx512fp16intrin.h	2023-04-14 10:13:30.054297659 +0200
@@ -28,6 +28,8 @@
 #ifndef __AVX512FP16INTRIN_H_INCLUDED
 #define __AVX512FP16INTRIN_H_INCLUDED
 
+#ifdef __FLT16_DIG__
+
 #ifndef __AVX512FP16__
 #pragma GCC push_options
 #pragma GCC target("avx512fp16")
@@ -7216,4 +7218,6 @@ _mm512_set1_pch (_Float16 _Complex __A)
 #pragma GCC pop_options
 #endif /* __DISABLE_AVX512FP16__ */
 
+#endif
+
 #endif /* __AVX512FP16INTRIN_H_INCLUDED */
--- gcc/config/i386/avx512fp16vlintrin.h.jj	2023-01-16 11:52:15.946736083 +0100
+++ gcc/config/i386/avx512fp16vlintrin.h	2023-04-14 10:13:54.152946427 +0200
@@ -28,6 +28,8 @@
 #ifndef __AVX512FP16VLINTRIN_H_INCLUDED
 #define __AVX512FP16VLINTRIN_H_INCLUDED
 
+#ifdef __FLT16_DIG__
+
 #if !defined(__AVX512VL__) || !defined(__AVX512FP16__)
 #pragma GCC push_options
 #pragma GCC target("avx512fp16,avx512vl")
@@ -3359,4 +3361,6 @@ _mm_set1_pch (_Float16 _Complex __A)
 #pragma GCC pop_options
 #endif /* __DISABLE_AVX512FP16VL__ */
 
+#endif
+
 #endif /* __AVX512FP16VLINTRIN_H_INCLUDED */
--- gcc/config/i386/avxneconvertintrin.h.jj	2023-01-16 11:52:15.949736039 +0100
+++ gcc/config/i386/avxneconvertintrin.h	2023-04-14 10:14:51.280113813 +0200
@@ -48,6 +48,7 @@ _mm256_bcstnebf16_ps (const void *__P)
   return (__m256) __builtin_ia32_vbcstnebf162ps256 ((const __bf16 *) __P);
 }
 
+#ifdef __FLT16_DIG__
 extern __inline __m128
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm_bcstnesh_ps (const void *__P)
@@ -61,6 +62,7 @@ _mm256_bcstnesh_ps (const void *__P)
 {
   return (__m256) __builtin_ia32_vbcstnesh2ps256 ((const _Float16 *) __P);
 }
+#endif
 
 extern __inline __m128
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))

i.e. if _Float16 type isn't usable, just hide all the _Float16 related stuff from the headers.
But don't we need the same thing for __bf16 too?  There I'm afraid aren't macros for that.

(In reply to Jakub Jelinek from comment #4)
> Yeah.  Enable all the time and have say the
> targetm.invalid_conversion, targetm.invalid_unary_op,
> targetm.invalid_binary_op
> and something in argument/return value passing reject _Float16/__bf16 in
> functions without SSE2.
> That will not be enough though, we'll need to arrange e.g. for the spot
> where we #undef/#define target  macros based on currently active ISA in
> pragmas to also
> do that for __STDCPP_FLOAT16_T__ and __STDCPP_BFLOAT16_T__ for C++, and
> change libstdc++
> such that for x86 it adds similarly to x86 intrin headers something like
Can we just cpp_undef _STDCPP_FLOAT16_T__ and __STDCPP_BFLOAT16_T__ for C++ in ix86_target_macros when !TARGET_SSE2 so that no need to change libstdc++ part.

(In reply to Hongtao.liu from comment #6)
> (In reply to Jakub Jelinek from comment #4)
> > Yeah.  Enable all the time and have say the
> > targetm.invalid_conversion, targetm.invalid_unary_op,
> > targetm.invalid_binary_op
> > and something in argument/return value passing reject _Float16/__bf16 in
> > functions without SSE2.
> > That will not be enough though, we'll need to arrange e.g. for the spot
> > where we #undef/#define target  macros based on currently active ISA in
> > pragmas to also
> > do that for __STDCPP_FLOAT16_T__ and __STDCPP_BFLOAT16_T__ for C++, and
> > change libstdc++
> > such that for x86 it adds similarly to x86 intrin headers something like
> Can we just cpp_undef _STDCPP_FLOAT16_T__ and __STDCPP_BFLOAT16_T__ for C++
> in ix86_target_macros when !TARGET_SSE2 so that no need to change libstdc++
> part.

Also need to undef
__LIBGCC_HAS_%d_MODE__, __LIBGCC_%d_FUNC_EXT__,__LIBGCC_%d_MANT_DIG__, __LIBGCC_%d_EXCESS_PRECISION__, __LIBGCC_%d_EPSILON__, __LIBGCC_%d_MAX__, __LIBGCC_%d_MIN__

Which are used for building libgcc(found in libbid)

And then ix86_emit_support_tinfos is not needed any more since type is always supported?

Or just adjust to

-      gcc_checking_assert (!float16_type_node && !bfloat16_type_node);
-      float16_type_node = ix86_float16_type_node;
-      bfloat16_type_node = ix86_bf16_type_node;
+      float16_type_node
+       = float16_type_node ? float16_type_node : ix86_float16_type_node;
+      bfloat16_type_node
+       = bfloat16_type_node ? bfloat16_type_node : ix86_bf16_type_node;

GCC 12.3 is being released, retargeting bugs to GCC 12.4.

The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:9a19fa8b616f83474c35cc5b34a3865073ced829

commit r14-2628-g9a19fa8b616f83474c35cc5b34a3865073ced829
Author: liuhongt <hongtao.liu@intel.com>
Date:   Tue Apr 18 14:53:04 2023 +0800

    Support type _Float16/__bf16 independent of SSE2.
    
    Enable _Float16 and __bf16 all the time but issue errors when the
    types are used in conversion, unary operation, binary operation,
    parameter passing or value return when TARGET_SSE2 is not available.
    
    Also undef macros which are used by libgcc/libstdc++ to check the
    backend support of the _Float16/__bf16 types when TARGET_SSE2 is not
    available.
    
    gcc/ChangeLog:
    
            PR target/109504
            * config/i386/i386-builtins.cc
            (ix86_register_float16_builtin_type): Remove TARGET_SSE2.
            (ix86_register_bf16_builtin_type): Ditto.
            * config/i386/i386-c.cc (ix86_target_macros): When TARGET_SSE2
            isn't available, undef the macros which are used to check the
            backend support of the _Float16/__bf16 types when building
            libstdc++ and libgcc.
            * config/i386/i386.cc (construct_container): Issue errors for
            HFmode/BFmode when TARGET_SSE2 is not available.
            (function_value_32): Ditto.
            (ix86_scalar_mode_supported_p): Remove TARGET_SSE2 for HFmode/BFmode.
            (ix86_libgcc_floating_mode_supported_p): Ditto.
            (ix86_emit_support_tinfos): Adjust codes.
            (ix86_invalid_conversion): Return diagnostic message string
            when there's conversion from/to BF/HFmode w/o TARGET_SSE2.
            (ix86_invalid_unary_op): New function.
            (ix86_invalid_binary_op): Ditto.
            (TARGET_INVALID_UNARY_OP): Define.
            (TARGET_INVALID_BINARY_OP): Define.
            * config/i386/immintrin.h [__SSE2__]: Remove for fp16/bf16
            related instrinsics header files.
            * config/i386/i386.h (VALID_SSE2_TYPE_MODE): New macro.
    
    gcc/testsuite/ChangeLog:
    
            * gcc.target/i386/pr109504.c: New test.
            * gcc.target/i386/sse2-bfloat16-1.c: Adjust error info.
            * gcc.target/i386/sse2-float16-1.c: Ditto.
            * gcc.target/i386/sse2-float16-4.c: New test.
            * gcc.target/i386/sse2-float16-5.c: New test.
            * g++.target/i386/float16-1.C: Adjust error info.
    
    libgcc/ChangeLog:
    
            * config/i386/t-softfp: Add -msse2 to libbid HFtype related
            files.

Fixed in GCC14.

.