AVX-512: added VPSHUFB instruction - all SKX forms

Added intrinsics and encoding tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@240277 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/IR/IntrinsicsX86.td b/include/llvm/IR/IntrinsicsX86.td
index ded4a7af7bc90947614998ae56ee4cfbe4d1a943..8e1d66849d17deb57c072393c3a68d4c3dba7d63 100644 (file)
--- a/include/llvm/IR/IntrinsicsX86.td
+++ b/include/llvm/IR/IntrinsicsX86.td
@@ -1397,6 +1397,24 @@ let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
           Intrinsic<[llvm_v8i64_ty],
           [llvm_v8i64_ty, llvm_v8i64_ty, llvm_v8i64_ty, llvm_i8_ty],
           [IntrNoMem]>;
+
+  def int_x86_avx512_mask_pshuf_b_128 : 
+        GCCBuiltin<"__builtin_ia32_pshufb128_mask">,
+          Intrinsic<[llvm_v16i8_ty],
+          [llvm_v16i8_ty, llvm_v16i8_ty, llvm_v16i8_ty,  llvm_i16_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_pshuf_b_256 : 
+        GCCBuiltin<"__builtin_ia32_pshufb256_mask">,
+          Intrinsic<[llvm_v32i8_ty],
+          [llvm_v32i8_ty, llvm_v32i8_ty, llvm_v32i8_ty,  llvm_i32_ty],
+          [IntrNoMem]>;
+
+  def int_x86_avx512_mask_pshuf_b_512 : 
+        GCCBuiltin<"__builtin_ia32_pshufb512_mask">,
+          Intrinsic<[llvm_v64i8_ty],
+          [llvm_v64i8_ty, llvm_v64i8_ty, llvm_v64i8_ty,  llvm_i64_ty],
+          [IntrNoMem]>;
 }
 
 // Vector blend

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index c60b11cbe3162cc5a4bdb32c3d1c3c857e76ebeb..56c5a63987574c1af94b7534fde051cee6f11af6 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -3870,6 +3870,19 @@ defm VPSHUFH : avx512_shift_rmi_w<0x70, MRMSrcReg, MRMSrcMem, "vpshufhw",
                                   X86PShufhw>, EVEX, AVX512XSIi8Base, VEX_W;
 defm VPSHUFL : avx512_shift_rmi_w<0x70, MRMSrcReg, MRMSrcMem, "vpshuflw",
                                   X86PShuflw>, EVEX, AVX512XDIi8Base, VEX_W;
+                                  
+multiclass avx512_pshufb_sizes<bits<8> opc, string OpcodeStr, SDNode OpNode> {
+  let Predicates = [HasBWI] in
+  defm Z:    avx512_var_shift<opc, OpcodeStr, OpNode, v64i8_info>, EVEX_V512;
+
+  let Predicates = [HasVLX, HasBWI] in {
+  defm Z256: avx512_var_shift<opc, OpcodeStr, OpNode, v32i8x_info>, EVEX_V256;
+  defm Z128: avx512_var_shift<opc, OpcodeStr, OpNode, v16i8x_info>, EVEX_V128;
+  }
+}
+
+defm VPSHUFB: avx512_pshufb_sizes<0x00, "vpshufb", X86pshufb>;
+
 //===----------------------------------------------------------------------===//
 // AVX-512 - MOVDDUP
 //===----------------------------------------------------------------------===//

diff --git a/lib/Target/X86/X86IntrinsicsInfo.h b/lib/Target/X86/X86IntrinsicsInfo.h
index b3007906886e04d1d64454958d9add12c47eb14a..b03f2b3e38e1c456ad801b5bf2f829948817d9bc 100644 (file)
--- a/lib/Target/X86/X86IntrinsicsInfo.h
+++ b/lib/Target/X86/X86IntrinsicsInfo.h
@@ -582,6 +582,12 @@ static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_por_q_128, INTR_TYPE_2OP_MASK, ISD::OR, 0),
   X86_INTRINSIC_DATA(avx512_mask_por_q_256, INTR_TYPE_2OP_MASK, ISD::OR, 0),
   X86_INTRINSIC_DATA(avx512_mask_por_q_512, INTR_TYPE_2OP_MASK, ISD::OR, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pshuf_b_128, INTR_TYPE_2OP_MASK,
+                     X86ISD::PSHUFB, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pshuf_b_256, INTR_TYPE_2OP_MASK,
+                    X86ISD::PSHUFB, 0),
+  X86_INTRINSIC_DATA(avx512_mask_pshuf_b_512, INTR_TYPE_2OP_MASK,
+                    X86ISD::PSHUFB, 0),
   X86_INTRINSIC_DATA(avx512_mask_psll_d,        INTR_TYPE_2OP_MASK, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(avx512_mask_psll_q,        INTR_TYPE_2OP_MASK, X86ISD::VSHL, 0),
   X86_INTRINSIC_DATA(avx512_mask_pslli_d,       VSHIFT_MASK, X86ISD::VSHLI, 0),

diff --git a/test/CodeGen/X86/avx512bw-intrinsics.ll b/test/CodeGen/X86/avx512bw-intrinsics.ll
index 06d44872d1e9db4f65a49bcf664a9b225664e26e..7c82d5f724f0a812f69e0856e5dc731cbe9e98fd 100644 (file)
--- a/test/CodeGen/X86/avx512bw-intrinsics.ll
+++ b/test/CodeGen/X86/avx512bw-intrinsics.ll
@@ -957,3 +957,16 @@ define <32 x i16>@test_int_x86_avx512_mask_pavg_w_512(<32 x i16> %x0, <32 x i16>
   %res2 = add <32 x i16> %res, %res1
   ret <32 x i16> %res2
 }
+
+declare <64 x i8> @llvm.x86.avx512.mask.pshuf.b.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pshuf_b_512
+; CHECK-NOT: call 
+; CHECK: kmov 
+; CHECK: vpshufb %zmm{{.*}}{%k1} 
+define <64 x i8>@test_int_x86_avx512_mask_pshuf_b_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {
+  %res = call <64 x i8> @llvm.x86.avx512.mask.pshuf.b.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)
+  %res1 = call <64 x i8> @llvm.x86.avx512.mask.pshuf.b.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)
+  %res2 = add <64 x i8> %res, %res1
+  ret <64 x i8> %res2
+}
\ No newline at end of file

diff --git a/test/CodeGen/X86/avx512bwvl-intrinsics.ll b/test/CodeGen/X86/avx512bwvl-intrinsics.ll
index 6cd6c1de805d02d7db165af656319bc2492d663e..6272cc86c12042afccf0bbe8b003d770eb53f475 100644 (file)
--- a/test/CodeGen/X86/avx512bwvl-intrinsics.ll
+++ b/test/CodeGen/X86/avx512bwvl-intrinsics.ll
@@ -3007,3 +3007,29 @@ define <16 x i16>@test_int_x86_avx512_mask_pavg_w_256(<16 x i16> %x0, <16 x i16>
   %res2 = add <16 x i16> %res, %res1
   ret <16 x i16> %res2
 }
+
+declare <16 x i8> @llvm.x86.avx512.mask.pshuf.b.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pshuf_b_128
+; CHECK-NOT: call 
+; CHECK: kmov 
+; CHECK: vpshufb %xmm{{.*}}{%k1} 
+define <16 x i8>@test_int_x86_avx512_mask_pshuf_b_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
+  %res = call <16 x i8> @llvm.x86.avx512.mask.pshuf.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3)
+  %res1 = call <16 x i8> @llvm.x86.avx512.mask.pshuf.b.128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 -1)
+  %res2 = add <16 x i8> %res, %res1
+  ret <16 x i8> %res2
+}
+
+declare <32 x i8> @llvm.x86.avx512.mask.pshuf.b.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
+
+; CHECK-LABEL: @test_int_x86_avx512_mask_pshuf_b_256
+; CHECK-NOT: call 
+; CHECK: kmov 
+; CHECK: vpshufb %ymm{{.*}}{%k1} 
+define <32 x i8>@test_int_x86_avx512_mask_pshuf_b_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
+  %res = call <32 x i8> @llvm.x86.avx512.mask.pshuf.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3)
+  %res1 = call <32 x i8> @llvm.x86.avx512.mask.pshuf.b.256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 -1)
+  %res2 = add <32 x i8> %res, %res1
+  ret <32 x i8> %res2
+}

diff --git a/test/MC/X86/x86-64-avx512bw.s b/test/MC/X86/x86-64-avx512bw.s
index 45e746308cbe89895b5121483c7b5bd0603d9ad5..5473124540eb08ac7264f24c665cc906f9e161d2 100644 (file)
--- a/test/MC/X86/x86-64-avx512bw.s
+++ b/test/MC/X86/x86-64-avx512bw.s
@@ -3560,3 +3560,38 @@
 // CHECK:  encoding: [0x62,0x61,0x15,0x40,0xe3,0xaa,0xc0,0xdf,0xff,0xff]
           vpavgw -8256(%rdx), %zmm29, %zmm29
 
+// CHECK: vpshufb %zmm20, %zmm26, %zmm22
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x40,0x00,0xf4]
+          vpshufb %zmm20, %zmm26, %zmm22
+
+// CHECK: vpshufb %zmm20, %zmm26, %zmm22 {%k7}
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x47,0x00,0xf4]
+          vpshufb %zmm20, %zmm26, %zmm22 {%k7}
+
+// CHECK: vpshufb %zmm20, %zmm26, %zmm22 {%k7} {z}
+// CHECK:  encoding: [0x62,0xa2,0x2d,0xc7,0x00,0xf4]
+          vpshufb %zmm20, %zmm26, %zmm22 {%k7} {z}
+
+// CHECK: vpshufb (%rcx), %zmm26, %zmm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x40,0x00,0x31]
+          vpshufb (%rcx), %zmm26, %zmm22
+
+// CHECK: vpshufb 291(%rax,%r14,8), %zmm26, %zmm22
+// CHECK:  encoding: [0x62,0xa2,0x2d,0x40,0x00,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vpshufb 291(%rax,%r14,8), %zmm26, %zmm22
+
+// CHECK: vpshufb 8128(%rdx), %zmm26, %zmm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x40,0x00,0x72,0x7f]
+          vpshufb 8128(%rdx), %zmm26, %zmm22
+
+// CHECK: vpshufb 8192(%rdx), %zmm26, %zmm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x40,0x00,0xb2,0x00,0x20,0x00,0x00]
+          vpshufb 8192(%rdx), %zmm26, %zmm22
+
+// CHECK: vpshufb -8192(%rdx), %zmm26, %zmm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x40,0x00,0x72,0x80]
+          vpshufb -8192(%rdx), %zmm26, %zmm22
+
+// CHECK: vpshufb -8256(%rdx), %zmm26, %zmm22
+// CHECK:  encoding: [0x62,0xe2,0x2d,0x40,0x00,0xb2,0xc0,0xdf,0xff,0xff]
+          vpshufb -8256(%rdx), %zmm26, %zmm22

diff --git a/test/MC/X86/x86-64-avx512bw_vl.s b/test/MC/X86/x86-64-avx512bw_vl.s
index 991c6102ebdfe27d2690fa174bc69969d06b075b..14a87df1ea835caad68fe9befe3b6299d858ec1c 100644 (file)
--- a/test/MC/X86/x86-64-avx512bw_vl.s
+++ b/test/MC/X86/x86-64-avx512bw_vl.s
@@ -6510,3 +6510,76 @@
 // CHECK: vpavgw -4128(%rdx), %ymm23, %ymm21
 // CHECK:  encoding: [0x62,0xe1,0x45,0x20,0xe3,0xaa,0xe0,0xef,0xff,0xff]
           vpavgw -4128(%rdx), %ymm23, %ymm21
+
+// CHECK: vpshufb %xmm27, %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0x82,0x3d,0x00,0x00,0xfb]
+          vpshufb %xmm27, %xmm24, %xmm23
+
+// CHECK: vpshufb %xmm27, %xmm24, %xmm23 {%k4}
+// CHECK:  encoding: [0x62,0x82,0x3d,0x04,0x00,0xfb]
+          vpshufb %xmm27, %xmm24, %xmm23 {%k4}
+
+// CHECK: vpshufb %xmm27, %xmm24, %xmm23 {%k4} {z}
+// CHECK:  encoding: [0x62,0x82,0x3d,0x84,0x00,0xfb]
+          vpshufb %xmm27, %xmm24, %xmm23 {%k4} {z}
+
+// CHECK: vpshufb (%rcx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x00,0x39]
+          vpshufb (%rcx), %xmm24, %xmm23
+
+// CHECK: vpshufb 291(%rax,%r14,8), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xa2,0x3d,0x00,0x00,0xbc,0xf0,0x23,0x01,0x00,0x00]
+          vpshufb 291(%rax,%r14,8), %xmm24, %xmm23
+
+// CHECK: vpshufb 2032(%rdx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x00,0x7a,0x7f]
+          vpshufb 2032(%rdx), %xmm24, %xmm23
+
+// CHECK: vpshufb 2048(%rdx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x00,0xba,0x00,0x08,0x00,0x00]
+          vpshufb 2048(%rdx), %xmm24, %xmm23
+
+// CHECK: vpshufb -2048(%rdx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x00,0x7a,0x80]
+          vpshufb -2048(%rdx), %xmm24, %xmm23
+
+// CHECK: vpshufb -2064(%rdx), %xmm24, %xmm23
+// CHECK:  encoding: [0x62,0xe2,0x3d,0x00,0x00,0xba,0xf0,0xf7,0xff,0xff]
+          vpshufb -2064(%rdx), %xmm24, %xmm23
+
+// CHECK: vpshufb %ymm17, %ymm18, %ymm19
+// CHECK:  encoding: [0x62,0xa2,0x6d,0x20,0x00,0xd9]
+          vpshufb %ymm17, %ymm18, %ymm19
+
+// CHECK: vpshufb %ymm17, %ymm18, %ymm19 {%k4}
+// CHECK:  encoding: [0x62,0xa2,0x6d,0x24,0x00,0xd9]
+          vpshufb %ymm17, %ymm18, %ymm19 {%k4}
+
+// CHECK: vpshufb %ymm17, %ymm18, %ymm19 {%k4} {z}
+// CHECK:  encoding: [0x62,0xa2,0x6d,0xa4,0x00,0xd9]
+          vpshufb %ymm17, %ymm18, %ymm19 {%k4} {z}
+
+// CHECK: vpshufb (%rcx), %ymm18, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x6d,0x20,0x00,0x19]
+          vpshufb (%rcx), %ymm18, %ymm19
+
+// CHECK: vpshufb 291(%rax,%r14,8), %ymm18, %ymm19
+// CHECK:  encoding: [0x62,0xa2,0x6d,0x20,0x00,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vpshufb 291(%rax,%r14,8), %ymm18, %ymm19
+
+// CHECK: vpshufb 4064(%rdx), %ymm18, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x6d,0x20,0x00,0x5a,0x7f]
+          vpshufb 4064(%rdx), %ymm18, %ymm19
+
+// CHECK: vpshufb 4096(%rdx), %ymm18, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x6d,0x20,0x00,0x9a,0x00,0x10,0x00,0x00]
+          vpshufb 4096(%rdx), %ymm18, %ymm19
+
+// CHECK: vpshufb -4096(%rdx), %ymm18, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x6d,0x20,0x00,0x5a,0x80]
+          vpshufb -4096(%rdx), %ymm18, %ymm19
+
+// CHECK: vpshufb -4128(%rdx), %ymm18, %ymm19
+// CHECK:  encoding: [0x62,0xe2,0x6d,0x20,0x00,0x9a,0xe0,0xef,0xff,0xff]
+          vpshufb -4128(%rdx), %ymm18, %ymm19
+