[Clang][AVX512][BuiltIn] Adding avx512 ( vcvt ) builtin to clang

[delena]

Please add tests where the rounding mode is not "current"

test/CodeGen/avx512f-builtins.c

2786

Please use _CURRENT_instead of 4.

include/clang/Basic/BuiltinsX86.def

 TARGET_BUILTIN(__builtin_ia32_loaddqudi128_mask, "V2LLiV2LLi*V2LLiUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loaddqudi256_mask, "V4LLiV4LLi*V4LLiUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loaddqusi128_mask, "V4iV4i*V4iUc","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_loaddqusi256_mask, "V8iV8i*V8iUc","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_loadupd128_mask, "V2dV2d*V2dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loadupd256_mask, "V4dV4d*V4dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loadups128_mask, "V4fV4f*V4fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_loadups256_mask, "V8fV8f*V8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2si64, "LLiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2si32, "iV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi32, "UiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi64, "ULLiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2si32, "iV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2si64, "LLiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2usi32, "UiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvtss2usi64, "ULLiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2si32, "iV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2si64, "LLiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi32, "UiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi64, "ULLiV2dIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2si32, "iV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2si64, "LLiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2usi32, "UiV4fIi","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_vcvttss2usi64, "ULLiV4fIi","","avx512f")
 
 #undef BUILTIN
 #undef TARGET_BUILTIN

lib/Headers/avx512fintrin.h

 #define _mm_comi_round_sd(__A, __B, __P, __R) __extension__ ({\
 __builtin_ia32_vcomisd ((__v2df) (__A), (__v2df) (__B), ( __P), (  __R));\
 })
 
 #define _mm_comi_round_ss( __A,  __B, __P, __R) __extension__ ({\
 __builtin_ia32_vcomiss ((__v4sf) (__A), (__v4sf) (__B), ( __P), ( __R));\
 })
 
+#define _mm_cvt_roundsd_si64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2si64 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvt_roundsd_i64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2si64 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvt_roundsd_si32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2si32 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvt_roundsd_i32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2si32 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvt_roundsd_u32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2usi32 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ unsigned __DEFAULT_FN_ATTRS
+_mm_cvtsd_u32 (__m128d __A)
+{
+  return (unsigned) __builtin_ia32_vcvtsd2usi32 ((__v2df) __A,
+             _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvt_roundsd_u64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtsd2usi64 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ unsigned long long __DEFAULT_FN_ATTRS
+_mm_cvtsd_u64 (__m128d __A)
+{
+  return (unsigned long long) __builtin_ia32_vcvtsd2usi64 ((__v2df)
+                 __A,
+                 _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvt_roundss_si32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2si32 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvt_roundss_i32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2si32 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvt_roundss_si64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2si64 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvt_roundss_i64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2si64 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvt_roundss_u32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2usi32 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ unsigned __DEFAULT_FN_ATTRS
+_mm_cvtss_u32 (__m128 __A)
+{
+  return (unsigned) __builtin_ia32_vcvtss2usi32 ((__v4sf) __A,
+             _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvt_roundss_u64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvtss2usi64 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ unsigned long long __DEFAULT_FN_ATTRS
+_mm_cvtss_u64 (__m128 __A)
+{
+  return (unsigned long long) __builtin_ia32_vcvtss2usi64 ((__v4sf)
+                 __A,
+                 _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundsd_i32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2si32 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvtt_roundsd_si32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2si32 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ int __DEFAULT_FN_ATTRS
+_mm_cvttsd_i32 (__m128d __A)
+{
+  return (int) __builtin_ia32_vcvttsd2si32 ((__v2df) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundsd_si64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2si64 ((__v2df)( __A),( __R));\
+})
+
+#define _mm_cvtt_roundsd_i64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2si64 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ long long __DEFAULT_FN_ATTRS
+_mm_cvttsd_i64 (__m128d __A)
+{
+  return (long long) __builtin_ia32_vcvttsd2si64 ((__v2df) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundsd_u32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2usi32 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ unsigned __DEFAULT_FN_ATTRS
+_mm_cvttsd_u32 (__m128d __A)
+{
+  return (unsigned) __builtin_ia32_vcvttsd2usi32 ((__v2df) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundsd_u64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttsd2usi64 ((__v2df)( __A),( __R));\
+})
+
+static __inline__ unsigned long long __DEFAULT_FN_ATTRS
+_mm_cvttsd_u64 (__m128d __A)
+{
+  return (unsigned long long) __builtin_ia32_vcvttsd2usi64 ((__v2df)
+                  __A,
+                  _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundss_i32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2si32 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvtt_roundss_si32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2si32 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ int __DEFAULT_FN_ATTRS
+_mm_cvttss_i32 (__m128 __A)
+{
+  return (int) __builtin_ia32_vcvttss2si32 ((__v4sf) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundss_i64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2si64 ((__v4sf)( __A),( __R));\
+})
+
+#define _mm_cvtt_roundss_si64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2si64 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ long long __DEFAULT_FN_ATTRS
+_mm_cvttss_i64 (__m128 __A)
+{
+  return (long long) __builtin_ia32_vcvttss2si64 ((__v4sf) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundss_u32( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2usi32 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ unsigned __DEFAULT_FN_ATTRS
+_mm_cvttss_u32 (__m128 __A)
+{
+  return (unsigned) __builtin_ia32_vcvttss2usi32 ((__v4sf) __A,
+              _MM_FROUND_CUR_DIRECTION);
+}
+
+#define _mm_cvtt_roundss_u64( __A, __R) __extension__ ({ \
+__builtin_ia32_vcvttss2usi64 ((__v4sf)( __A),( __R));\
+})
+
+static __inline__ unsigned long long __DEFAULT_FN_ATTRS
+_mm_cvttss_u64 (__m128 __A)
+{
+  return (unsigned long long) __builtin_ia32_vcvttss2usi64 ((__v4sf)
+                  __A,
+                  _MM_FROUND_CUR_DIRECTION);
+}
 
 #undef __DEFAULT_FN_ATTRS
 
 #endif // __AVX512FINTRIN_H

test/CodeGen/avx512f-builtins.c

 }
 
 __mmask16 test_mm512_kmov(__mmask16 __A) {
   // CHECK-LABEL: @test_mm512_kmov
   // CHECK: load i16, i16* %__A.addr.i, align 2
   return _mm512_kmov(__A); 
 }
 
+unsigned long long test_mm_cvt_roundsd_si64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_si64
+  // CHECK: @llvm.x86.avx512.vcvtsd2si64
+  return _mm_cvt_roundsd_si64(__A, 4); 
+}
+
+long long test_mm_cvt_roundsd_i64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_i64
+  // CHECK: @llvm.x86.avx512.vcvtsd2si64
+  return _mm_cvt_roundsd_i64(__A, 4); 
+}
+
+int test_mm_cvt_roundsd_si32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_si32
+  // CHECK: @llvm.x86.avx512.vcvtsd2si32
+  return _mm_cvt_roundsd_si32(__A, 4); 
+}
+
+int test_mm_cvt_roundsd_i32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_i32
+  // CHECK: @llvm.x86.avx512.vcvtsd2si32
+  return _mm_cvt_roundsd_i32(__A, 4); 
+}
+
+unsigned test_mm_cvt_roundsd_u32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_u32
+  // CHECK: @llvm.x86.avx512.vcvtsd2usi32
+  return _mm_cvt_roundsd_u32(__A, 4); 
+}
 
+unsigned test_mm_cvtsd_u32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtsd_u32
+  // CHECK: @llvm.x86.avx512.vcvtsd2usi32
+  return _mm_cvtsd_u32(__A); 
+}
+
+unsigned long long test_mm_cvt_roundsd_u64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundsd_u64
+  // CHECK: @llvm.x86.avx512.vcvtsd2usi64
+  return _mm_cvt_roundsd_u64(__A, 4); 
+}
+
+unsigned long long test_mm_cvtsd_u64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtsd_u64
+  // CHECK: @llvm.x86.avx512.vcvtsd2usi64
+  return _mm_cvtsd_u64(__A); 
+}
+
+int test_mm_cvt_roundss_si32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_si32
+  // CHECK: @llvm.x86.avx512.vcvtss2si32
+  return _mm_cvt_roundss_si32(__A, 4); 
+}
+
+int test_mm_cvt_roundss_i32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_i32
+  // CHECK: @llvm.x86.avx512.vcvtss2si32
+  return _mm_cvt_roundss_i32(__A, 4); 
+}
 
+int test_mm_cvt_roundss_si64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_si64
+  // CHECK: @llvm.x86.avx512.vcvtss2si64
+  return _mm_cvt_roundss_si64(__A, 4); 
+}
+
+long long test_mm_cvt_roundss_i64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_i64
+  // CHECK: @llvm.x86.avx512.vcvtss2si64
+  return _mm_cvt_roundss_i64(__A, 4); 
+}
+
+unsigned test_mm_cvt_roundss_u32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_u32
+  // CHECK: @llvm.x86.avx512.vcvtss2usi32
+  return _mm_cvt_roundss_u32(__A, 4); 
+}
+
+unsigned test_mm_cvtss_u32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtss_u32
+  // CHECK: @llvm.x86.avx512.vcvtss2usi32
+  return _mm_cvtss_u32(__A); 
+}
+
+unsigned long long test_mm_cvt_roundss_u64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvt_roundss_u64
+  // CHECK: @llvm.x86.avx512.vcvtss2usi64
+  return _mm_cvt_roundss_u64(__A, 4); 
+}
+
+unsigned long long test_mm_cvtss_u64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtss_u64
+  // CHECK: @llvm.x86.avx512.vcvtss2usi64
+  return _mm_cvtss_u64(__A); 
+}
+
+int test_mm_cvtt_roundsd_i32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_i32
+  // CHECK: @llvm.x86.avx512.cvttsd2si
+  return _mm_cvtt_roundsd_i32(__A, 4); 
+}
+
+int test_mm_cvtt_roundsd_si32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_si32
+  // CHECK: @llvm.x86.avx512.cvttsd2si
+  return _mm_cvtt_roundsd_si32(__A, 4); 
+}
+
+int test_mm_cvttsd_i32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvttsd_i32
+  // CHECK: @llvm.x86.avx512.cvttsd2si
+  return _mm_cvttsd_i32(__A); 
+}
+
+unsigned long long test_mm_cvtt_roundsd_si64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_si64
+  // CHECK: @llvm.x86.avx512.cvttsd2si64
+  return _mm_cvtt_roundsd_si64(__A, 4); 
+}
+
+long long test_mm_cvtt_roundsd_i64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_i64
+  // CHECK: @llvm.x86.avx512.cvttsd2si64
+  return _mm_cvtt_roundsd_i64(__A, 4); 
+}
+
+long long test_mm_cvttsd_i64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvttsd_i64
+  // CHECK: @llvm.x86.avx512.cvttsd2si64
+  return _mm_cvttsd_i64(__A); 
+}
+
+unsigned test_mm_cvtt_roundsd_u32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_u32
+  // CHECK: @llvm.x86.avx512.cvttsd2usi
+  return _mm_cvtt_roundsd_u32(__A, 4); 
+}
+
+unsigned test_mm_cvttsd_u32(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvttsd_u32
+  // CHECK: @llvm.x86.avx512.cvttsd2usi
+  return _mm_cvttsd_u32(__A); 
+}
+
+unsigned long long test_mm_cvtt_roundsd_u64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundsd_u64
+  // CHECK: @llvm.x86.avx512.cvttsd2usi64
+  return _mm_cvtt_roundsd_u64(__A, 4); 
+}
+
+unsigned long long test_mm_cvttsd_u64(__m128d __A) {
+  // CHECK-LABEL: @test_mm_cvttsd_u64
+  // CHECK: @llvm.x86.avx512.cvttsd2usi64
+  return _mm_cvttsd_u64(__A); 
+}
+
+int test_mm_cvtt_roundss_i32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_i32
+  // CHECK: @llvm.x86.avx512.cvttss2si
+  return _mm_cvtt_roundss_i32(__A, 4); 
+}
+
+int test_mm_cvtt_roundss_si32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_si32
+  // CHECK: @llvm.x86.avx512.cvttss2si
+  return _mm_cvtt_roundss_si32(__A, 4); 
+}
+
+int test_mm_cvttss_i32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvttss_i32
+  // CHECK: @llvm.x86.avx512.cvttss2si
+  return _mm_cvttss_i32(__A); 
+}
+
+float test_mm_cvtt_roundss_i64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_i64
+  // CHECK: @llvm.x86.avx512.cvttss2si64
+  return _mm_cvtt_roundss_i64(__A, 4); 
+}
+
+long long test_mm_cvtt_roundss_si64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_si64
+  // CHECK: @llvm.x86.avx512.cvttss2si64
+  return _mm_cvtt_roundss_si64(__A, 4); 
+}
+
+long long test_mm_cvttss_i64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvttss_i64
+  // CHECK: @llvm.x86.avx512.cvttss2si64
+  return _mm_cvttss_i64(__A); 
+}
+
+unsigned test_mm_cvtt_roundss_u32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_u32
+  // CHECK: @llvm.x86.avx512.cvttss2usi
+  return _mm_cvtt_roundss_u32(__A, 4); 
+}
+
+unsigned test_mm_cvttss_u32(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvttss_u32
+  // CHECK: @llvm.x86.avx512.cvttss2usi
+  return _mm_cvttss_u32(__A); 
+}
+
+unsigned long long test_mm_cvtt_roundss_u64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvtt_roundss_u64
+  // CHECK: @llvm.x86.avx512.cvttss2usi64
+  return _mm_cvtt_roundss_u64(__A, 4); 
+}
+
+unsigned long long test_mm_cvttss_u64(__m128 __A) {
+  // CHECK-LABEL: @test_mm_cvttss_u64
+  // CHECK: @llvm.x86.avx512.cvttss2usi64
+  return _mm_cvttss_u64(__A); 
+}