[AVX512] Adding VPERMB Intrinsics

Differential Revision: http://reviews.llvm.org/D16296

llvm-svn: 258316

4 files changed, 72 insertions, 6 deletions

diff --git a/llvm/include/llvm/IR/IntrinsicsX86.td b/llvm/include/llvm/IR/IntrinsicsX86.td
index adc60612b79..d1879f731ca 100644
--- a/llvm/include/llvm/IR/IntrinsicsX86.td
+++ b/llvm/include/llvm/IR/IntrinsicsX86.td
@@ -4314,14 +4314,14 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
               Intrinsic<[llvm_v32i16_ty], [llvm_v32i16_ty,
                         llvm_v32i16_ty, llvm_v32i16_ty, llvm_i32_ty],  [IntrNoMem]>;
   def int_x86_avx512_mask_permvar_qi_128 : GCCBuiltin<"__builtin_ia32_permvarqi128_mask">,
-              Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty,
-                        llvm_v8i16_ty, llvm_v8i16_ty, llvm_i8_ty],  [IntrNoMem]>;
+              Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty,
+                        llvm_v16i8_ty, llvm_v16i8_ty, llvm_i16_ty],  [IntrNoMem]>;
   def int_x86_avx512_mask_permvar_qi_256 : GCCBuiltin<"__builtin_ia32_permvarqi256_mask">,
-              Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty,
-                        llvm_v16i16_ty, llvm_v16i16_ty, llvm_i16_ty],  [IntrNoMem]>;
+              Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty,
+                        llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty],  [IntrNoMem]>;
   def int_x86_avx512_mask_permvar_qi_512 : GCCBuiltin<"__builtin_ia32_permvarqi512_mask">,
-              Intrinsic<[llvm_v32i16_ty], [llvm_v32i16_ty,
-                        llvm_v32i16_ty, llvm_v32i16_ty, llvm_i32_ty],  [IntrNoMem]>;
+              Intrinsic<[llvm_v64i8_ty], [llvm_v64i8_ty,
+                        llvm_v64i8_ty, llvm_v64i8_ty, llvm_i64_ty],  [IntrNoMem]>;
   def int_x86_avx512_mask_permvar_sf_256 : GCCBuiltin<"__builtin_ia32_permvarsf256_mask">,
               Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty,
                         llvm_v8i32_ty, llvm_v8f32_ty, llvm_i8_ty],  [IntrNoMem]>;
diff --git a/llvm/lib/Target/X86/X86IntrinsicsInfo.h b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
index d5dbe77452b..5b86e948699 100644
--- a/llvm/lib/Target/X86/X86IntrinsicsInfo.h
+++ b/llvm/lib/Target/X86/X86IntrinsicsInfo.h
@@ -1067,6 +1067,9 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_permvar_hi_128, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
   X86_INTRINSIC_DATA(avx512_mask_permvar_hi_256, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
   X86_INTRINSIC_DATA(avx512_mask_permvar_hi_512, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
+  X86_INTRINSIC_DATA(avx512_mask_permvar_qi_128, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
+  X86_INTRINSIC_DATA(avx512_mask_permvar_qi_256, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
+  X86_INTRINSIC_DATA(avx512_mask_permvar_qi_512, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
   X86_INTRINSIC_DATA(avx512_mask_permvar_sf_256, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
   X86_INTRINSIC_DATA(avx512_mask_permvar_sf_512, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
   X86_INTRINSIC_DATA(avx512_mask_permvar_si_256, INTR_TYPE_2OP_MASK, X86ISD::VPERMV, 0),
diff --git a/llvm/test/CodeGen/X86/avx512vbmi-intrinsics.ll b/llvm/test/CodeGen/X86/avx512vbmi-intrinsics.ll
new file mode 100644
index 00000000000..4b6794616a2
--- /dev/null
+++ b/llvm/test/CodeGen/X86/avx512vbmi-intrinsics.ll
@@ -0,0 +1,21 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=skx  -mattr=+avx512vbmi --show-mc-encoding| FileCheck %s
+declare <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
+
+define <64 x i8>@test_int_x86_avx512_mask_permvar_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_permvar_qi_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovq %rdi, %k1 
+; CHECK-NEXT:    vpermb %zmm1, %zmm0, %zmm2 {%k1} 
+; CHECK-NEXT:    vpermb %zmm1, %zmm0, %zmm3 {%k1} {z} 
+; CHECK-NEXT:    vpermb %zmm1, %zmm0, %zmm0 
+; CHECK-NEXT:    vpaddb %zmm3, %zmm2, %zmm1 
+; CHECK-NEXT:    vpaddb %zmm0, %zmm1, %zmm0 
+; CHECK-NEXT:    retq 
+ %res = call <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)
+ %res1 = call <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> zeroinitializer, i64 %x3)
+ %res2 = call <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)
+ %res3 = add <64 x i8> %res, %res1
+ %res4 = add <64 x i8> %res3, %res2
+ ret <64 x i8> %res4
+}
diff --git a/llvm/test/CodeGen/X86/avx512vbmivl-intrinsics.ll b/llvm/test/CodeGen/X86/avx512vbmivl-intrinsics.ll
new file mode 100644
index 00000000000..24c05c436b7
--- /dev/null
+++ b/llvm/test/CodeGen/X86/avx512vbmivl-intrinsics.ll
@@ -0,0 +1,42 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=skx -mattr=+avx512vl -mattr=+avx512vbmi --show-mc-encoding| FileCheck %s
+declare <16 x i8> @llvm.x86.avx512.mask.permvar.qi.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
+
+define <16 x i8>@test_int_x86_avx512_mask_permvar_qi_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_permvar_qi_128:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovw %edi, %k1 
+; CHECK-NEXT:    vpermb %xmm1, %xmm0, %xmm2 {%k1} 
+; CHECK-NEXT:    vpermb %xmm1, %xmm0, %xmm3 {%k1} {z} 
+; CHECK-NEXT:    vpermb %xmm1, %xmm0, %xmm0 
+; CHECK-NEXT:    vpaddb %xmm3, %xmm2, %xmm1 
+; CHECK-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 
+; CHECK-NEXT:    retq 
+  %res = call <16 x i8> @llvm.x86.avx512.mask.permvar.qi.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
+  %res1 = call <16 x i8> @llvm.x86.avx512.mask.permvar.qi.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> zeroinitializer, i16 %x3)
+  %res2 = call <16 x i8> @llvm.x86.avx512.mask.permvar.qi.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
+  %res3 = add <16 x i8> %res, %res1
+  %res4 = add <16 x i8> %res3, %res2
+  ret <16 x i8> %res4
+}
+
+declare <32 x i8> @llvm.x86.avx512.mask.permvar.qi.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
+
+define <32 x i8>@test_int_x86_avx512_mask_permvar_qi_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
+; CHECK-LABEL: test_int_x86_avx512_mask_permvar_qi_256:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    kmovd %edi, %k1 
+; CHECK-NEXT:    vpermb %ymm1, %ymm0, %ymm2 {%k1} 
+; CHECK-NEXT:    vpermb %ymm1, %ymm0, %ymm3 {%k1} {z} 
+; CHECK-NEXT:    vpermb %ymm1, %ymm0, %ymm0 
+; CHECK-NEXT:    vpaddb %ymm3, %ymm2, %ymm1 
+; CHECK-NEXT:    vpaddb %ymm0, %ymm1, %ymm0 
+; CHECK-NEXT:    retq 
+  %res = call <32 x i8> @llvm.x86.avx512.mask.permvar.qi.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
+  %res1 = call <32 x i8> @llvm.x86.avx512.mask.permvar.qi.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> zeroinitializer, i32 %x3)
+  %res2 = call <32 x i8> @llvm.x86.avx512.mask.permvar.qi.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
+  %res3 = add <32 x i8> %res, %res1
+  %res4 = add <32 x i8> %res3, %res2
+  ret <32 x i8> %res4
+}
+