[X86] Support some missing intrinsics

[LiuChen3]

Summary

Support for _mm512_i32logather_pd, _mm512_mask_i32logather_pd,
_mm512_i32logather_epi64, _mm512_mask_i32logather_epi64, _mm512_i32loscatter_pd,
_mm512_mask_i32loscatter_pd.

[RKSimon]

Add _mm512_i32loscatter_epi64 and _mm512_mask_i32loscatter_epi64 for completeness?

[LiuChen3]

In D100368#2685189, @RKSimon wrote:

Add _mm512_i32loscatter_epi64 and _mm512_mask_i32loscatter_epi64 for completeness?

It seems we doesn't support KNCNI in LLVM. Am I right?

[craig.topper]

In D100368#2685559, @LiuChen3 wrote:

In D100368#2685189, @RKSimon wrote:

Add _mm512_i32loscatter_epi64 and _mm512_mask_i32loscatter_epi64 for completeness?

It seems we doesn't support KNCNI in LLVM. Am I right?

Correct. We do not support KNCI.

[RKSimon]

That's why I said "completeness" :)

clang/lib/Headers/avx512fintrin.h

9692

Newline between instructions

[LiuChen3]

Address Simon's comments

[craig.topper]

clang/lib/Headers/avx512fintrin.h

9593

"scale them in dst" doesn't make sense.

9611

Use of "dst" here doesn't make sense since it isn't a parameter name so shouldn't be abbreviated.

9636

Same

9654

Same

clang/test/CodeGen/X86/avx512f-builtins.c

10853

Any idea why _mm512_mask_i32loscatter_epi64 is documented as only being KNCI?

[pengfei]

clang/lib/Headers/avx512fintrin.h

9611

Can we arrange "dst" as the abbreviation for the result of the intrinsic?
Intrinsic guide uses such naming conversion and correspondings to its operation.
We are trying to unify the descriptions here and intrinsic guide.

[LiuChen3]

1. Rebase.
2. Adding _mm512_i32loscatter_epi64 and _mm512_mask_i32loscatter_epi64.

[LiuChen3]

Fix format issue.

[LiuChen3]

clang/test/CodeGen/X86/avx512f-builtins.c

10853

After confirmation, we think it’s a intrinsics guide bug. The intrinsics guide will be updated later.

[pengfei]

LGTM.

clang/lib/Headers/avx512fintrin.h

 ///    destination.
 /// \returns A 32-bit signed integer containing the moved value.
 static __inline__ int __DEFAULT_FN_ATTRS512
 _mm512_cvtsi512_si32(__m512i __A) {
   __v16si __b = (__v16si)__A;
   return __b[0];
 }
 
+/// Loads 8 double-precision (64-bit) floating-point elements stored at memory
+/// locations starting at location \a base_addr at packed 32-bit integer indices
+/// stored in the lower half of \a vindex scaled by \a scale them in dst.
+///
+/// This intrinsic corresponds to the <c> VGATHERDPD </c> instructions.
+///
+/// \operation
+/// FOR j := 0 to 7
+///   i := j*64
+///   m := j*32
+///   addr := base_addr + SignExtend64(vindex[m+31:m]) * ZeroExtend64(scale) * 8
+///   dst[i+63:i] := MEM[addr+63:addr]
+/// ENDFOR
+/// dst[MAX:512] := 0
+/// \endoperation
+#define _mm512_i32logather_pd(vindex, base_addr, scale)                        \
+  _mm512_i32gather_pd(_mm512_castsi512_si256(vindex), (base_addr), (scale))
+
+/// Loads 8 double-precision (64-bit) floating-point elements from memory
+/// starting at location \a base_addr at packed 32-bit integer indices stored in
+/// the lower half of \a vindex scaled by \a scale into dst using writemask
+/// \a mask (elements are copied from \a src when the corresponding mask bit is
+/// not set).
+///
+/// This intrinsic corresponds to the <c> VGATHERDPD </c> instructions.
+///
+/// \operation
+/// FOR j := 0 to 7
+///   i := j*64
+///   m := j*32
+///   IF mask[j]
+///     addr := base_addr + SignExtend64(vindex[m+31:m]) * ZeroExtend64(scale) * 8
+///     dst[i+63:i] := MEM[addr+63:addr]
+///   ELSE
+///     dst[i+63:i] := src[i+63:i]
+///   FI
+/// ENDFOR
+/// dst[MAX:512] := 0
+/// \endoperation
+#define _mm512_mask_i32logather_pd(src, mask, vindex, base_addr, scale)        \
+  _mm512_mask_i32gather_pd((src), (mask), _mm512_castsi512_si256(vindex),      \
+                           (base_addr), (scale))
+
+/// Loads 8 64-bit integer elements from memory starting at location \a base_addr
+/// at packed 32-bit integer indices stored in the lower half of \a vindex
+/// scaled by \a scale and stores them in dst.
+///
+/// This intrinsic corresponds to the <c> VPGATHERDQ </c> instructions.
+///
+/// \operation
+/// FOR j := 0 to 7
+///   i := j*64
+///   m := j*32
+///   addr := base_addr + SignExtend64(vindex[m+31:m]) * ZeroExtend64(scale) * 8
+///   dst[i+63:i] := MEM[addr+63:addr]
+/// ENDFOR
+/// dst[MAX:512] := 0
+/// \endoperation
+#define _mm512_i32logather_epi64(vindex, base_addr, scale)                     \
+  _mm512_i32gather_epi64(_mm512_castsi512_si256(vindex), (base_addr), (scale))
+
+/// Loads 8 64-bit integer elements from memory starting at location \a base_addr
+/// at packed 32-bit integer indices stored in the lower half of \a vindex
+/// scaled by \a scale and stores them in dst using writemask \a mask (elements
+/// are copied from \a src when the corresponding mask bit is not set).
+///
+/// This intrinsic corresponds to the <c> VPGATHERDQ </c> instructions.
+///
+/// \operation
+/// FOR j := 0 to 7
+///   i := j*64
+///   m := j*32
+///   IF mask[j]
+///     addr := base_addr + SignExtend64(vindex[m+31:m]) * ZeroExtend64(scale) * 8
+///     dst[i+63:i] := MEM[addr+63:addr]
+///   ELSE
+///     dst[i+63:i] := src[i+63:i]
+///   FI
+/// ENDFOR
+/// dst[MAX:512] := 0
+/// \endoperation
+#define _mm512_mask_i32logather_epi64(src, mask, vindex, base_addr, scale)     \
+  _mm512_mask_i32gather_epi64((src), (mask), _mm512_castsi512_si256(vindex),   \
+                              (base_addr), (scale))
+
+/// Stores 8 packed double-precision (64-bit) floating-point elements in \a v1
+/// and to memory locations starting at location \a base_addr at packed 32-bit
+/// integer indices stored in \a vindex scaled by \a scale.
+///
+/// This intrinsic corresponds to the <c> VSCATTERDPD </c> instructions.
+///
+/// \operation
+/// FOR j := 0 to 7
+///   i := j*64
+///   m := j*32
+///   addr := base_addr + SignExtend64(vindex[m+31:m]) * ZeroExtend64(scale) * 8
+///   MEM[addr+63:addr] := v1[i+63:i]
+/// ENDFOR
+/// \endoperation
+#define _mm512_i32loscatter_pd(base_addr, vindex, v1, scale)                   \
+  _mm512_i32scatter_pd((base_addr), _mm512_castsi512_si256(vindex), (v1), (scale))
+
+/// Stores 8 packed double-precision (64-bit) floating-point elements in \a v1
+/// to memory locations starting at location \a base_addr at packed 32-bit
+/// integer indices stored in \a vindex scaled by \a scale. Only those elements
+/// whose corresponding mask bit is set in writemask \a mask are written to
+/// memory.
+///
+/// This intrinsic corresponds to the <c> VSCATTERDPD </c> instructions.
+///
+/// \operation
+/// FOR j := 0 to 7
+///   i := j*64
+///   m := j*32
+///   IF mask[j]
+///     addr := base_addr + SignExtend64(vindex[m+31:m]) * ZeroExtend64(scale) * 8
+///     MEM[addr+63:addr] := a[i+63:i]
+///   FI
+/// ENDFOR
+/// \endoperation
+#define _mm512_mask_i32loscatter_pd(base_addr, mask, vindex, v1, scale)        \
+  _mm512_mask_i32scatter_pd((base_addr), (mask),                               \
+                            _mm512_castsi512_si256(vindex), (v1), (scale))
+
+/// Stores 8 packed 64-bit integer elements located in \a v1 and stores them in
+/// memory locations starting at location \a base_addr at packed 32-bit integer
+/// indices stored in \a vindex scaled by \a scale.
+///
+/// This intrinsic corresponds to the <c> VPSCATTERDQ </c> instructions.
+///
+/// \operation
+/// FOR j := 0 to 7
+///   i := j*64
+///   m := j*32
+///   addr := base_addr + SignExtend64(vindex[m+31:m]) * ZeroExtend64(scale) * 8
+///   MEM[addr+63:addr] := a[i+63:i]
+/// ENDFOR
+/// \endoperation
+#define _mm512_i32loscatter_epi64(base_addr, vindex, v1, scale)                \
+  _mm512_i32scatter_epi64((base_addr),                                         \
+                          _mm512_castsi512_si256(vindex), (v1), (scale))
+
+/// Stores 8 packed 64-bit integer elements located in a and stores them in
+/// memory locations starting at location \a base_addr at packed 32-bit integer
+/// indices stored in \a vindex scaled by scale using writemask \a mask (elements
+/// whose corresponding mask bit is not set are not written to memory).
+///
+/// This intrinsic corresponds to the <c> VPSCATTERDQ </c> instructions.
+///
+/// \operation
+/// FOR j := 0 to 7
+///   i := j*64
+///   m := j*32
+///   IF mask[j]
+///     addr := base_addr + SignExtend64(vindex[m+31:m]) * ZeroExtend64(scale) * 8
+///     MEM[addr+63:addr] := a[i+63:i]
+///   FI
+/// ENDFOR
+/// \endoperation
+#define _mm512_mask_i32loscatter_epi64(base_addr, mask, vindex, v1, scale)     \
+  _mm512_mask_i32scatter_epi64((base_addr), (mask),                            \
+                               _mm512_castsi512_si256(vindex), (v1), (scale))
+
 #undef __DEFAULT_FN_ATTRS512
 #undef __DEFAULT_FN_ATTRS128
 #undef __DEFAULT_FN_ATTRS
 
 #endif /* __AVX512FINTRIN_H */

clang/test/CodeGen/X86/avx512f-builtins.c

 }
 
 __m512i test_mm512_zextsi256_si512(__m256i A) {
   // CHECK-LABEL: test_mm512_zextsi256_si512
   // CHECK: store <4 x i64> zeroinitializer
   // CHECK: shufflevector <4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   return _mm512_zextsi256_si512(A);
 }
+
+__m512d test_mm512_i32logather_pd(__m512i __index, void const *__addr) {
+  // CHECK-LABEL: @test_mm512_i32logather_pd
+  // CHECK: @llvm.x86.avx512.mask.gather.dpd.512
+  return _mm512_i32logather_pd(__index, __addr, 2);
+}
+
+__m512d test_mm512_mask_i32logather_pd(__m512d __v1_old, __mmask8 __mask, __m512i __index, void const *__addr) {
+  // CHECK-LABEL: @test_mm512_mask_i32logather_pd
+  // CHECK: @llvm.x86.avx512.mask.gather.dpd.512
+  return _mm512_mask_i32logather_pd(__v1_old, __mask, __index, __addr, 2);
+}
+
+void test_mm512_i32loscatter_pd(void *__addr, __m512i __index, __m512d __v1) {
+  // CHECK-LABEL: @test_mm512_i32loscatter_pd
+  // CHECK: @llvm.x86.avx512.mask.scatter.dpd.512
+  return _mm512_i32loscatter_pd(__addr, __index, __v1, 2);
+}
+
+void test_mm512_mask_i32loscatter_pd(void *__addr, __mmask8 __mask, __m512i __index, __m512d __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i32loscatter_pd
+  // CHECK: @llvm.x86.avx512.mask.scatter.dpd.512
+  return _mm512_mask_i32loscatter_pd(__addr, __mask, __index, __v1, 2);
+}
+
+__m512i test_mm512_i32logather_epi64(__m512i __index, void const *__addr) {
+  // CHECK-LABEL: @test_mm512_i32logather_epi64
+  // CHECK: @llvm.x86.avx512.mask.gather.dpq.512
+  return _mm512_i32logather_epi64(__index, __addr, 2);
+}
+
+__m512i test_mm512_mask_i32logather_epi64(__m512i __v1_old, __mmask8 __mask, __m512i __index, void const *__addr) {
+  // CHECK-LABEL: @test_mm512_mask_i32logather_epi64
+  // CHECK: @llvm.x86.avx512.mask.gather.dpq.512
+  return _mm512_mask_i32logather_epi64(__v1_old, __mask, __index, __addr, 2);
+}
+
+void test_mm512_i32loscatter_epi64(void *__addr, __m512i __index, __m512i __v1) {
+  // CHECK-LABEL: @test_mm512_i32loscatter_epi64
+  // CHECK: @llvm.x86.avx512.mask.scatter.dpq.512
+  _mm512_i32loscatter_epi64(__addr, __index, __v1, 2);
+}
+
+void test_mm512_mask_i32loscatter_epi64(void *__addr, __mmask8 __mask, __m512i __index, __m512i __v1) {
+  // CHECK-LABEL: @test_mm512_mask_i32loscatter_epi64
+  // CHECK: @llvm.x86.avx512.mask.scatter.dpq.512
+  _mm512_mask_i32loscatter_epi64(__addr, __mask, __index, __v1, 2);
+}