Re: [PATCH] i386: Handle target of __builtin_ia32_cmp[p|s][s|d] from avx into sse/sse2/avx
On Thu, May 30, 2024 at 1:52 PM Hu, Lin1 wrote:
>
> Hi, all
>
> This patch aims to extend __builtin_ia32_cmp[p|s][s|d] from avx to
> sse/sse2/avx, where its immediate is in range of [0, 7].
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu, OK for trunk?
Ok.
>
> BRs,
> Lin
>
> gcc/ChangeLog:
>
> * config/i386/avxintrin.h: Move cmp[p|s][s|d] to [e|x]mmintrin.h,
> and move macros to xmmintrin.h
> * config/i386/emmintrin.h: Add cmp[p|s]s intrins.
> * config/i386/i386-builtin.def: Modify __builtin_ia32_cmp[p|s][s|d].
> * config/i386/i386-expand.cc
> (ix86_expand_args_builtin): Raise error when imm is in range of
> [8, 32] without avx.
> * config/i386/sse.md (avx_cmp3): Modefy define_insn.
> (avx_vmcmp3): Ditto.
> * config/i386/xmmintrin.h (_CMP_EQ_OQ): New macro for sse/sse2.
> (_CMP_LT_OS): Ditto
> (_CMP_LE_OS): Ditto
> (_CMP_UNORD_Q): Ditto
> (_CMP_NEQ_UQ): Ditto
> (_CMP_NLT_US): Ditto
> (_CMP_NLE_US): Ditto
> (_CMP_ORD_Q): Ditto
> (_mm_cmp_ps): Move intrin from avxintrin.h to xmmintrin.h
> (_mm_cmp_ss): Ditto.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/sse-cmp-1.c: New test.
> * gcc.target/i386/sse-cmp-2.c: Ditto.
> * gcc.target/i386/sse-cmp-error-1.c: Ditto.
> ---
> gcc/config/i386/avxintrin.h | 56 ---
> gcc/config/i386/emmintrin.h | 22 +
> gcc/config/i386/i386-builtin.def | 10 +-
> gcc/config/i386/i386-expand.cc| 6 ++
> gcc/config/i386/predicates.md | 5 +
> gcc/config/i386/sse.md| 42
> gcc/config/i386/xmmintrin.h | 41
> gcc/testsuite/gcc.target/i386/sse-cmp-1.c | 20
> gcc/testsuite/gcc.target/i386/sse-cmp-2.c | 96 +++
> gcc/testsuite/gcc.target/i386/sse-cmp-error.c | 16
> 10 files changed, 236 insertions(+), 78 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/i386/sse-cmp-1.c
> create mode 100644 gcc/testsuite/gcc.target/i386/sse-cmp-2.c
> create mode 100644 gcc/testsuite/gcc.target/i386/sse-cmp-error.c
>
> diff --git a/gcc/config/i386/avxintrin.h b/gcc/config/i386/avxintrin.h
> index 80214540888..ec9b9905b5f 100644
> --- a/gcc/config/i386/avxintrin.h
> +++ b/gcc/config/i386/avxintrin.h
> @@ -72,22 +72,6 @@ typedef double __m256d_u __attribute__ ((__vector_size__
> (32),
>
> /* Compare predicates for scalar and packed compare intrinsics. */
>
> -/* Equal (ordered, non-signaling) */
> -#define _CMP_EQ_OQ 0x00
> -/* Less-than (ordered, signaling) */
> -#define _CMP_LT_OS 0x01
> -/* Less-than-or-equal (ordered, signaling) */
> -#define _CMP_LE_OS 0x02
> -/* Unordered (non-signaling) */
> -#define _CMP_UNORD_Q 0x03
> -/* Not-equal (unordered, non-signaling) */
> -#define _CMP_NEQ_UQ0x04
> -/* Not-less-than (unordered, signaling) */
> -#define _CMP_NLT_US0x05
> -/* Not-less-than-or-equal (unordered, signaling) */
> -#define _CMP_NLE_US0x06
> -/* Ordered (nonsignaling) */
> -#define _CMP_ORD_Q 0x07
> /* Equal (unordered, non-signaling) */
> #define _CMP_EQ_UQ 0x08
> /* Not-greater-than-or-equal (unordered, signaling) */
> @@ -381,18 +365,6 @@ _mm256_xor_ps (__m256 __A, __m256 __B)
> }
>
> #ifdef __OPTIMIZE__
> -extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__,
> __artificial__))
> -_mm_cmp_pd (__m128d __X, __m128d __Y, const int __P)
> -{
> - return (__m128d) __builtin_ia32_cmppd ((__v2df)__X, (__v2df)__Y, __P);
> -}
> -
> -extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__,
> __artificial__))
> -_mm_cmp_ps (__m128 __X, __m128 __Y, const int __P)
> -{
> - return (__m128) __builtin_ia32_cmpps ((__v4sf)__X, (__v4sf)__Y, __P);
> -}
> -
> extern __inline __m256d __attribute__((__gnu_inline__, __always_inline__,
> __artificial__))
> _mm256_cmp_pd (__m256d __X, __m256d __Y, const int __P)
> {
> @@ -406,27 +378,7 @@ _mm256_cmp_ps (__m256 __X, __m256 __Y, const int __P)
>return (__m256) __builtin_ia32_cmpps256 ((__v8sf)__X, (__v8sf)__Y,
>__P);
> }
> -
> -extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__,
> __artificial__))
> -_mm_cmp_sd (__m128d __X, __m128d __Y, const int __P)
> -{
> - return (__m128d) __builtin_ia32_cmpsd ((__v2df)__X, (__v2df)__Y, __P);
> -}
> -
> -extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__,
> __artificial__))
> -_mm_cmp_ss (__m128 __X, __m128 __Y, const int __P)
> -{
> - return (__m128) __builtin_ia32_cmpss ((__v4sf)__X, (__v4sf)__Y, __P);
> -}
> #else
> -#define _mm_cmp_pd(X, Y, P)\
> - ((__m128d) __builtin_ia32_cmppd ((__v2df)(__m128d)(X), \
> - (__v2df)(__m128d)(Y), (int)(P)))
> -
> -#define
[PATCH] i386: Handle target of __builtin_ia32_cmp[p|s][s|d] from avx into sse/sse2/avx
Hi, all
This patch aims to extend __builtin_ia32_cmp[p|s][s|d] from avx to
sse/sse2/avx, where its immediate is in range of [0, 7].
Bootstrapped and regtested on x86_64-pc-linux-gnu, OK for trunk?
BRs,
Lin
gcc/ChangeLog:
* config/i386/avxintrin.h: Move cmp[p|s][s|d] to [e|x]mmintrin.h,
and move macros to xmmintrin.h
* config/i386/emmintrin.h: Add cmp[p|s]s intrins.
* config/i386/i386-builtin.def: Modify __builtin_ia32_cmp[p|s][s|d].
* config/i386/i386-expand.cc
(ix86_expand_args_builtin): Raise error when imm is in range of
[8, 32] without avx.
* config/i386/sse.md (avx_cmp3): Modefy define_insn.
(avx_vmcmp3): Ditto.
* config/i386/xmmintrin.h (_CMP_EQ_OQ): New macro for sse/sse2.
(_CMP_LT_OS): Ditto
(_CMP_LE_OS): Ditto
(_CMP_UNORD_Q): Ditto
(_CMP_NEQ_UQ): Ditto
(_CMP_NLT_US): Ditto
(_CMP_NLE_US): Ditto
(_CMP_ORD_Q): Ditto
(_mm_cmp_ps): Move intrin from avxintrin.h to xmmintrin.h
(_mm_cmp_ss): Ditto.
gcc/testsuite/ChangeLog:
* gcc.target/i386/sse-cmp-1.c: New test.
* gcc.target/i386/sse-cmp-2.c: Ditto.
* gcc.target/i386/sse-cmp-error-1.c: Ditto.
---
gcc/config/i386/avxintrin.h | 56 ---
gcc/config/i386/emmintrin.h | 22 +
gcc/config/i386/i386-builtin.def | 10 +-
gcc/config/i386/i386-expand.cc| 6 ++
gcc/config/i386/predicates.md | 5 +
gcc/config/i386/sse.md| 42
gcc/config/i386/xmmintrin.h | 41
gcc/testsuite/gcc.target/i386/sse-cmp-1.c | 20
gcc/testsuite/gcc.target/i386/sse-cmp-2.c | 96 +++
gcc/testsuite/gcc.target/i386/sse-cmp-error.c | 16
10 files changed, 236 insertions(+), 78 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/i386/sse-cmp-1.c
create mode 100644 gcc/testsuite/gcc.target/i386/sse-cmp-2.c
create mode 100644 gcc/testsuite/gcc.target/i386/sse-cmp-error.c
diff --git a/gcc/config/i386/avxintrin.h b/gcc/config/i386/avxintrin.h
index 80214540888..ec9b9905b5f 100644
--- a/gcc/config/i386/avxintrin.h
+++ b/gcc/config/i386/avxintrin.h
@@ -72,22 +72,6 @@ typedef double __m256d_u __attribute__ ((__vector_size__
(32),
/* Compare predicates for scalar and packed compare intrinsics. */
-/* Equal (ordered, non-signaling) */
-#define _CMP_EQ_OQ 0x00
-/* Less-than (ordered, signaling) */
-#define _CMP_LT_OS 0x01
-/* Less-than-or-equal (ordered, signaling) */
-#define _CMP_LE_OS 0x02
-/* Unordered (non-signaling) */
-#define _CMP_UNORD_Q 0x03
-/* Not-equal (unordered, non-signaling) */
-#define _CMP_NEQ_UQ0x04
-/* Not-less-than (unordered, signaling) */
-#define _CMP_NLT_US0x05
-/* Not-less-than-or-equal (unordered, signaling) */
-#define _CMP_NLE_US0x06
-/* Ordered (nonsignaling) */
-#define _CMP_ORD_Q 0x07
/* Equal (unordered, non-signaling) */
#define _CMP_EQ_UQ 0x08
/* Not-greater-than-or-equal (unordered, signaling) */
@@ -381,18 +365,6 @@ _mm256_xor_ps (__m256 __A, __m256 __B)
}
#ifdef __OPTIMIZE__
-extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
-_mm_cmp_pd (__m128d __X, __m128d __Y, const int __P)
-{
- return (__m128d) __builtin_ia32_cmppd ((__v2df)__X, (__v2df)__Y, __P);
-}
-
-extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
-_mm_cmp_ps (__m128 __X, __m128 __Y, const int __P)
-{
- return (__m128) __builtin_ia32_cmpps ((__v4sf)__X, (__v4sf)__Y, __P);
-}
-
extern __inline __m256d __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm256_cmp_pd (__m256d __X, __m256d __Y, const int __P)
{
@@ -406,27 +378,7 @@ _mm256_cmp_ps (__m256 __X, __m256 __Y, const int __P)
return (__m256) __builtin_ia32_cmpps256 ((__v8sf)__X, (__v8sf)__Y,
__P);
}
-
-extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
-_mm_cmp_sd (__m128d __X, __m128d __Y, const int __P)
-{
- return (__m128d) __builtin_ia32_cmpsd ((__v2df)__X, (__v2df)__Y, __P);
-}
-
-extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
-_mm_cmp_ss (__m128 __X, __m128 __Y, const int __P)
-{
- return (__m128) __builtin_ia32_cmpss ((__v4sf)__X, (__v4sf)__Y, __P);
-}
#else
-#define _mm_cmp_pd(X, Y, P)\
- ((__m128d) __builtin_ia32_cmppd ((__v2df)(__m128d)(X), \
- (__v2df)(__m128d)(Y), (int)(P)))
-
-#define _mm_cmp_ps(X, Y, P)\
- ((__m128) __builtin_ia32_cmpps ((__v4sf)(__m128)(X), \
- (__v4sf)(__m128)(Y), (int)(P)))
-
#define _mm256_cmp_pd(X, Y, P) \
((__m256d)
