[X86][AVX512] lowering kunpack intrinsic - clang part

[jina.nahias]

Summary

the llvm part : https://reviews.llvm.org/D39720

[RKSimon]

lib/Headers/avx512bwintrin.h

2045 ↗

(On Diff #121849)

Is this right? The Intel docs says it should be:

k[31:0] := a[31:0]
k[63:32] := b[31:0]
k[MAX:64] := 0

Also, is the cast on __A necessary?

Same for the others.

[jina.nahias]

lib/Headers/avx512bwintrin.h

2045 ↗

(On Diff #121849)

you are right, i fixed it.

[craig.topper]

test/CodeGen/avx512f-builtins.c

6231 ↗

(On Diff #121899)

Does this really produce kunpackb in the backend? The type promotion here makes me skeptic

6235 ↗

(On Diff #121899)

Please add some line breaks to this.

[jina.nahias]

test/CodeGen/avx512f-builtins.c

6231 ↗

(On Diff #121899)

yes,
the code we get:
vpcmpneqd %zmm1, %zmm0, %k0
vpcmpneqd %zmm3, %zmm2, %k1
kunpckbw %k1, %k0, %k1
vpcmpneqd %zmm5, %zmm4, %k0 {%k1}
kmovd %k0, %eax

1. kill: %AX<def> %AX<kill> %EAX<kill>

vzeroupper

[jina.nahias]

ping

cfe/trunk/lib/Headers/avx512bwintrin.h

_mm512_maskz_set1_epi8 (__mmask64 __M, char __A)
   return (__m512i) __builtin_ia32_selectb_512(__M,
                                               (__v64qi) _mm512_set1_epi8(__A),
                                               (__v64qi) _mm512_setzero_si512());
 }
 
 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
 _mm512_kunpackd (__mmask64 __A, __mmask64 __B)
 {
-  return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A,
-                (__mmask64) __B);
+  return (__mmask64)  (( __A  & 0xFFFFFFFF) | ( __B << 32));
 }
 
 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
 _mm512_kunpackw (__mmask32 __A, __mmask32 __B)
 {
-  return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A,
-                (__mmask32) __B);
+return (__mmask32)  (( __A  & 0xFFFF) | ( __B << 16));
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
 _mm512_mask_loadu_epi16 (__m512i __W, __mmask32 __U, void const *__P)
 {
   return (__m512i) __builtin_ia32_loaddquhi512_mask ((__v32hi *) __P,
                  (__v32hi) __W,
                  (__mmask32) __U);

cfe/trunk/lib/Headers/avx512fintrin.h

 _mm512_kortestz (__mmask16 __A, __mmask16 __B)
 {
   return __builtin_ia32_kortestzhi ((__mmask16) __A, (__mmask16) __B);
 }
 
 static __inline__ __mmask16 __DEFAULT_FN_ATTRS
 _mm512_kunpackb (__mmask16 __A, __mmask16 __B)
 {
-  return (__mmask16) __builtin_ia32_kunpckhi ((__mmask16) __A, (__mmask16) __B);
+  return (__mmask16)  (( __A  & 0xFF) | ( __B << 8));
 }
 
 static __inline__ __mmask16 __DEFAULT_FN_ATTRS
 _mm512_kxnor (__mmask16 __A, __mmask16 __B)
 {
   return (__mmask16) __builtin_ia32_kxnorhi ((__mmask16) __A, (__mmask16) __B);
 }
 

cfe/trunk/test/CodeGen/avx512bw-builtins.c

__m512i test_mm512_maskz_set1_epi8(__mmask64 __M, char __A) {
   // CHECK: insertelement <64 x i8> %{{.*}}, i8 %{{.*}}, i32 60
   // CHECK: insertelement <64 x i8> %{{.*}}, i8 %{{.*}}, i32 61
   // CHECK: insertelement <64 x i8> %{{.*}}, i8 %{{.*}}, i32 62
   // CHECK: insertelement <64 x i8> %{{.*}}, i8 %{{.*}}, i32 63
   // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}}
   return _mm512_maskz_set1_epi8(__M, __A); 
 }
 
-__mmask64 test_mm512_kunpackd(__mmask64 __A, __mmask64 __B) {
+__mmask64 test_mm512_kunpackd(__m512i __A, __m512i __B, __m512i __C, __m512i __D, __m512i __E, __m512i __F) {
   // CHECK-LABEL: @test_mm512_kunpackd
-  // CHECK: @llvm.x86.avx512.kunpck.dq
-  return _mm512_kunpackd(__A, __B); 
+  // CHECK: bitcast <64 x i1> %{{.*}} to i64
+  // CHECK: bitcast <64 x i1> %{{.*}} to i64
+  // CHECK: and i64 %{{.*}}, 4294967295
+  // CHECK: shl i64 %{{.*}}, 32
+  // CHECK: or i64 %{{.*}}, %{{.*}}
+  // CHECK: bitcast i64 %{{.*}} to <64 x i1>
+  return _mm512_mask_cmpneq_epu8_mask(_mm512_kunpackd(_mm512_cmpneq_epu8_mask(__B, __A),_mm512_cmpneq_epu8_mask(__C, __D)), __E, __F); 
 }
 
-__mmask32 test_mm512_kunpackw(__mmask32 __A, __mmask32 __B) {
+__mmask32 test_mm512_kunpackw(__m512i __A, __m512i __B, __m512i __C, __m512i __D, __m512i __E, __m512i __F) {
   // CHECK-LABEL: @test_mm512_kunpackw
-  // CHECK: @llvm.x86.avx512.kunpck.wd
-  return _mm512_kunpackw(__A, __B); 
+  // CHECK: bitcast <32 x i1> %{{.*}} to i32
+  // CHECK: bitcast <32 x i1> %{{.*}} to i32
+  // CHECK: and i32 %{{.*}}, 65535
+  // CHECK: shl i32 %{{.*}}, 16
+  // CHECK: or i32 %{{.*}}, %{{.*}}
+  // CHECK: bitcast i32 %{{.*}} to <32 x i1>
+  return _mm512_mask_cmpneq_epu16_mask(_mm512_kunpackw(_mm512_cmpneq_epu16_mask(__B, __A),_mm512_cmpneq_epu16_mask(__C, __D)), __E, __F); 
 }
 
 __m512i test_mm512_mask_loadu_epi16(__m512i __W, __mmask32 __U, void const *__P) {
   // CHECK-LABEL: @test_mm512_mask_loadu_epi16
   // CHECK: @llvm.masked.load.v32i16.p0v32i16(<32 x i16>* %{{.*}}, i32 1, <32 x i1> %{{.*}}, <32 x i16> %{{.*}})
   return _mm512_mask_loadu_epi16(__W, __U, __P); 
 }
 

cfe/trunk/test/CodeGen/avx512f-builtins.c

 }
 
 int test_mm512_kortestz(__mmask16 __A, __mmask16 __B) {
   // CHECK-LABEL: @test_mm512_kortestz
   // CHECK: @llvm.x86.avx512.kortestz.w
   return _mm512_kortestz(__A, __B); 
 }
 
-__mmask16 test_mm512_kunpackb(__mmask16 __A, __mmask16 __B) {
+__mmask16 test_mm512_kunpackb(__m512i __A, __m512i __B, __m512i __C, __m512i __D, __m512i __E, __m512i __F) {
   // CHECK-LABEL: @test_mm512_kunpackb
-  // CHECK: @llvm.x86.avx512.kunpck.bw
-  return _mm512_kunpackb(__A, __B); 
+  // CHECK: bitcast <16 x i1> %{{.*}} to i16
+  // CHECK: bitcast <16 x i1> %{{.*}} to i16
+  // CHECK: and i32 %{{.*}}, 255
+  // CHECK: shl i32 %{{.*}}, 8
+  // CHECK: or i32 %{{.*}}, %{{.*}}
+  // CHECK: bitcast i16 %{{.*}} to <16 x i1>
+  return _mm512_mask_cmpneq_epu32_mask(_mm512_kunpackb(_mm512_cmpneq_epu32_mask(__A, __B),
+                                                       _mm512_cmpneq_epu32_mask(__C, __D)),
+                                                       __E, __F);
 }
 
 __mmask16 test_mm512_kxnor(__mmask16 __A, __mmask16 __B) {
   // CHECK-LABEL: @test_mm512_kxnor
   // CHECK: @llvm.x86.avx512.kxnor.w
   return _mm512_kxnor(__A, __B); 
 }