[X86] Lowering integer truncation intrinsics to native IR

This patch lowers the _mm[256|512]_cvtepi{64|32|16}_epi{32|16|8} intrinsics to
native IR in cases where the result's length is less than 128 bits.

The resulting IR for 256-bit inputs is folded into VPMOV instructions, while for
128-bit inputs the vpshufb (or, in the 64-to-32-bit case, vinsertps)
instructions are generated instead

Differential Revision: https://reviews.llvm.org/D48712

llvm-svn: 336643

4 files changed, 50 insertions, 37 deletions

diff --git a/clang/lib/Headers/avx512vlbwintrin.h b/clang/lib/Headers/avx512vlbwintrin.h
index 72e07d05419..1b038dd04df 100644
--- a/clang/lib/Headers/avx512vlbwintrin.h
+++ b/clang/lib/Headers/avx512vlbwintrin.h
@@ -1496,10 +1496,10 @@ _mm256_maskz_cvtusepi16_epi8 (__mmask16 __M, __m256i __A) {
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_cvtepi16_epi8 (__m128i __A) {
-
-  return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A,
-               (__v16qi) _mm_setzero_si128(),
-               (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v8hi)__A, __v8qi),
+      (__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
+      12, 13, 14, 15);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
diff --git a/clang/lib/Headers/avx512vlintrin.h b/clang/lib/Headers/avx512vlintrin.h
index de898845ce5..0ee1d00ef4d 100644
--- a/clang/lib/Headers/avx512vlintrin.h
+++ b/clang/lib/Headers/avx512vlintrin.h
@@ -31,6 +31,10 @@
 #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl"), __min_vector_width__(128)))
 #define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl"), __min_vector_width__(256)))
 
+typedef short __v2hi __attribute__((__vector_size__(4)));
+typedef char __v4qi __attribute__((__vector_size__(4)));
+typedef char __v2qi __attribute__((__vector_size__(2)));
+
 /* Integer compare */
 
 #define _mm_cmpeq_epi32_mask(A, B) \
@@ -7341,9 +7345,9 @@ _mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_cvtepi32_epi8 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
-              (__v16qi)_mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v4si)__A, __v4qi), (__v4qi){0, 0, 0, 0}, 0, 1,
+      2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
@@ -7371,9 +7375,10 @@ _mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_cvtepi32_epi8 (__m256i __A)
 {
-  return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
-              (__v16qi)_mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v8si)__A, __v8qi),
+      (__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
+      12, 13, 14, 15);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
@@ -7400,9 +7405,9 @@ _mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_cvtepi32_epi16 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
-              (__v8hi) _mm_setzero_si128 (),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v4si)__A, __v4hi), (__v4hi){0, 0, 0, 0}, 0, 1,
+      2, 3, 4, 5, 6, 7);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
@@ -7456,9 +7461,9 @@ _mm256_mask_cvtepi32_storeu_epi16 (void *  __P, __mmask8 __M, __m256i __A)
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_cvtepi64_epi8 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
-              (__v16qi) _mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v2di)__A, __v2qi), (__v2qi){0, 0}, 0, 1, 2, 3,
+      3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
@@ -7485,9 +7490,9 @@ _mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_cvtepi64_epi8 (__m256i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
-              (__v16qi) _mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v4di)__A, __v4qi), (__v4qi){0, 0, 0, 0}, 0, 1,
+      2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
@@ -7514,9 +7519,8 @@ _mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_cvtepi64_epi32 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
-              (__v4si)_mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v2di)__A, __v2si), (__v2si){0, 0}, 0, 1, 2, 3);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
@@ -7571,9 +7575,9 @@ _mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_cvtepi64_epi16 (__m128i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
-              (__v8hi) _mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v2di)__A, __v2hi), (__v2hi){0, 0}, 0, 1, 2, 3,
+      3, 3, 3, 3);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
@@ -7601,9 +7605,9 @@ _mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_cvtepi64_epi16 (__m256i __A)
 {
-  return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
-              (__v8hi)_mm_undefined_si128(),
-              (__mmask8) -1);
+  return (__m128i)__builtin_shufflevector(
+      __builtin_convertvector((__v4di)__A, __v4hi), (__v4hi){0, 0, 0, 0}, 0, 1,
+      2, 3, 4, 5, 6, 7);
 }
 
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
diff --git a/clang/test/CodeGen/avx512vl-builtins.c b/clang/test/CodeGen/avx512vl-builtins.c
index 540ea223d93..7e4e64381c1 100644
--- a/clang/test/CodeGen/avx512vl-builtins.c
+++ b/clang/test/CodeGen/avx512vl-builtins.c
@@ -8503,7 +8503,8 @@ void test_mm256_mask_cvtusepi64_storeu_epi16(void * __P, __mmask8 __M, __m256i _
 
 __m128i test_mm_cvtepi32_epi8(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi32_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.db.128
+  // CHECK: trunc <4 x i32> %{{.*}} to <4 x i8>
+  // CHECK: shufflevector <4 x i8> %{{.*}}, <4 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>
   return _mm_cvtepi32_epi8(__A); 
 }
 
@@ -8527,7 +8528,8 @@ void test_mm_mask_cvtepi32_storeu_epi8(void * __P, __mmask8 __M, __m128i __A) {
 
 __m128i test_mm256_cvtepi32_epi8(__m256i __A) {
   // CHECK-LABEL: @test_mm256_cvtepi32_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.db.256
+  // CHECK: trunc <8 x i32> %{{.*}} to <8 x i8>
+  // CHECK: shufflevector <8 x i8> %{{.*}}, <8 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
   return _mm256_cvtepi32_epi8(__A); 
 }
 
@@ -8551,7 +8553,8 @@ void test_mm256_mask_cvtepi32_storeu_epi8(void * __P, __mmask8 __M, __m256i __A)
 
 __m128i test_mm_cvtepi32_epi16(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi32_epi16
-  // CHECK: @llvm.x86.avx512.mask.pmov.dw.128
+  // CHECK: trunc <4 x i32> %{{.*}} to <4 x i16>
+  // CHECK: shufflevector <4 x i16> %{{.*}}, <4 x i16> %{{.*}}, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   return _mm_cvtepi32_epi16(__A); 
 }
 
@@ -8599,7 +8602,8 @@ void test_mm256_mask_cvtepi32_storeu_epi16(void *  __P, __mmask8 __M, __m256i __
 
 __m128i test_mm_cvtepi64_epi8(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi64_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.qb.128
+  // CHECK: trunc <2 x i64> %{{.*}} to <2 x i8>
+  // CHECK: shufflevector <2 x i8> %{{.*}}, <2 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>
   return _mm_cvtepi64_epi8(__A); 
 }
 
@@ -8623,7 +8627,8 @@ void test_mm_mask_cvtepi64_storeu_epi8(void * __P, __mmask8 __M, __m128i __A) {
 
 __m128i test_mm256_cvtepi64_epi8(__m256i __A) {
   // CHECK-LABEL: @test_mm256_cvtepi64_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.qb.256
+  // CHECK: trunc <4 x i64> %{{.*}} to <4 x i8>
+  // CHECK: shufflevector <4 x i8> %{{.*}}, <4 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>
   return _mm256_cvtepi64_epi8(__A); 
 }
 
@@ -8647,7 +8652,8 @@ void test_mm256_mask_cvtepi64_storeu_epi8(void * __P, __mmask8 __M, __m256i __A)
 
 __m128i test_mm_cvtepi64_epi32(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi64_epi32
-  // CHECK: @llvm.x86.avx512.mask.pmov.qd.128
+  // CHECK: trunc <2 x i64> %{{.*}} to <2 x i32>
+  // CHECK: shufflevector <2 x i32> %{{.*}}, <2 x i32> %{{.*}}, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   return _mm_cvtepi64_epi32(__A); 
 }
 
@@ -8697,7 +8703,8 @@ void test_mm256_mask_cvtepi64_storeu_epi32(void * __P, __mmask8 __M, __m256i __A
 
 __m128i test_mm_cvtepi64_epi16(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi64_epi16
-  // CHECK: @llvm.x86.avx512.mask.pmov.qw.128
+  // CHECK: trunc <2 x i64> %{{.*}} to <2 x i16>
+  // CHECK: shufflevector <2 x i16> %{{.*}}, <2 x i16> %{{.*}}, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 3, i32 3, i32 3, i32 3>
   return _mm_cvtepi64_epi16(__A); 
 }
 
@@ -8721,7 +8728,8 @@ void test_mm_mask_cvtepi64_storeu_epi16(void * __P, __mmask8 __M, __m128i __A) {
 
 __m128i test_mm256_cvtepi64_epi16(__m256i __A) {
   // CHECK-LABEL: @test_mm256_cvtepi64_epi16
-  // CHECK: @llvm.x86.avx512.mask.pmov.qw.256
+  // CHECK: trunc <4 x i64> %{{.*}} to <4 x i16>
+  // CHECK: shufflevector <4 x i16> %{{.*}}, <4 x i16> %{{.*}}, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   return _mm256_cvtepi64_epi16(__A); 
 }
 
diff --git a/clang/test/CodeGen/avx512vlbw-builtins.c b/clang/test/CodeGen/avx512vlbw-builtins.c
index 607b0117944..c28c3fbd7f8 100644
--- a/clang/test/CodeGen/avx512vlbw-builtins.c
+++ b/clang/test/CodeGen/avx512vlbw-builtins.c
@@ -1792,7 +1792,8 @@ __m128i test_mm256_maskz_cvtusepi16_epi8(__mmask16 __M, __m256i __A) {
 
 __m128i test_mm_cvtepi16_epi8(__m128i __A) {
   // CHECK-LABEL: @test_mm_cvtepi16_epi8
-  // CHECK: @llvm.x86.avx512.mask.pmov.wb.128
+  // CHECK: trunc <8 x i16> %{{.*}} to <8 x i8>
+  // CHECK: shufflevector <8 x i8> %{{.*}}, <8 x i8> %{{.*}}, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
   return _mm_cvtepi16_epi8(__A); 
 }