[X86][AVX512] Remove constant pool shuffle decoding from SelectionDAG

[RKSimon]

Summary

This patch removes the last use of the constant pool shuffle decode helper and consistently uses the 'getTargetShuffleMaskIndices' versions instead. The constant pool versions are now purely used for assembly comments.

The avx512vbmi instrinsic upgrades had to be altered as they were being decoded as broadcasts, similar to what I fixed in rL346032. I don't think the change is critical - although its annoying that we lose the {k}{z} instruction, on vpermi2 it's not that useful....

[RKSimon]

ping?

[craig.topper]

LGTM

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

 
 #include "X86ISelLowering.h"
 #include "Utils/X86ShuffleDecode.h"
 #include "X86CallingConv.h"
 #include "X86FrameLowering.h"
 #include "X86InstrBuilder.h"
 #include "X86IntrinsicsInfo.h"
 #include "X86MachineFunctionInfo.h"
-#include "X86ShuffleDecodeConstantPool.h"
 #include "X86TargetMachine.h"
 #include "X86TargetObjectFile.h"
 #include "llvm/ADT/SmallBitVector.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/ADT/StringExtras.h"
 #include "llvm/ADT/StringSwitch.h"
 #include "llvm/Analysis/EHPersonalities.h"
static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
   case X86ISD::VPERMV3: {
     assert(N->getOperand(0).getValueType() == VT && "Unexpected value type");
     assert(N->getOperand(2).getValueType() == VT && "Unexpected value type");
     IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(2);
     // Unlike most shuffle nodes, VPERMV3's mask operand is the middle one.
     Ops.push_back(N->getOperand(0));
     Ops.push_back(N->getOperand(2));
     SDValue MaskNode = N->getOperand(1);
-    if (auto *C = getTargetConstantFromNode(MaskNode)) {
-      DecodeVPERMV3Mask(C, MaskEltSize, VT.getSizeInBits(), Mask);
+    if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask,
+                                    RawUndefs)) {
+      DecodeVPERMV3Mask(RawMask, RawUndefs, Mask);
       break;
     }
     return false;
   }
   default: llvm_unreachable("unknown target shuffle node");
   }
 
   // Empty mask indicates the decode failed.

llvm/trunk/test/CodeGen/X86/avx512vbmi-intrinsics-upgrade.ll

 
 define <64 x i8>@test_int_x86_avx512_mask_vpermi2var_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {
 ; X86-LABEL: test_int_x86_avx512_mask_vpermi2var_qi_512:
 ; X86:       # %bb.0:
 ; X86-NEXT:    vmovdqa64 %zmm0, %zmm3 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xd8]
 ; X86-NEXT:    vpermt2b %zmm2, %zmm1, %zmm3 # encoding: [0x62,0xf2,0x75,0x48,0x7d,0xda]
 ; X86-NEXT:    kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]
 ; X86-NEXT:    vpermi2b %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x75,0xca]
-; X86-NEXT:    vpxor %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd9,0xef,0xe4]
-; X86-NEXT:    vpermi2b %zmm2, %zmm0, %zmm4 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xc9,0x75,0xe2]
+; X86-NEXT:    vmovdqa64 %zmm1, %zmm4 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xe1]
+; X86-NEXT:    vpermi2b %zmm2, %zmm0, %zmm4 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x75,0xe2]
 ; X86-NEXT:    vpaddb %zmm3, %zmm4, %zmm0 # encoding: [0x62,0xf1,0x5d,0x48,0xfc,0xc3]
 ; X86-NEXT:    vpaddb %zmm0, %zmm1, %zmm0 # encoding: [0x62,0xf1,0x75,0x48,0xfc,0xc0]
 ; X86-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-LABEL: test_int_x86_avx512_mask_vpermi2var_qi_512:
 ; X64:       # %bb.0:
 ; X64-NEXT:    vmovdqa64 %zmm0, %zmm3 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xd8]
 ; X64-NEXT:    vpermt2b %zmm2, %zmm1, %zmm3 # encoding: [0x62,0xf2,0x75,0x48,0x7d,0xda]
 ; X64-NEXT:    kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
 ; X64-NEXT:    vpermi2b %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x75,0xca]
-; X64-NEXT:    vpxor %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd9,0xef,0xe4]
-; X64-NEXT:    vpermi2b %zmm2, %zmm0, %zmm4 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xc9,0x75,0xe2]
+; X64-NEXT:    vmovdqa64 %zmm1, %zmm4 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xe1]
+; X64-NEXT:    vpermi2b %zmm2, %zmm0, %zmm4 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x75,0xe2]
 ; X64-NEXT:    vpaddb %zmm3, %zmm4, %zmm0 # encoding: [0x62,0xf1,0x5d,0x48,0xfc,0xc3]
 ; X64-NEXT:    vpaddb %zmm0, %zmm1, %zmm0 # encoding: [0x62,0xf1,0x75,0x48,0xfc,0xc0]
 ; X64-NEXT:    retq # encoding: [0xc3]
   %res = call <64 x i8> @llvm.x86.avx512.mask.vpermi2var.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.vpermi2var.qi.512(<64 x i8> %x0, <64 x i8> zeroinitializer, <64 x i8> %x2, i64 %x3)
+  %res1 = call <64 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.512(<64 x i8> %x0, <64 x i8> %res, <64 x i8> %x2, i64 %x3)
   %res2 = call <64 x i8> @llvm.x86.avx512.mask.vpermi2var.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
 }
 
 declare <64 x i8> @llvm.x86.avx512.mask.vpermt2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)
 

llvm/trunk/test/CodeGen/X86/avx512vbmivl-intrinsics-upgrade.ll

 
 define <16 x i8>@test_int_x86_avx512_mask_vpermi2var_qi_128(<16 x i8> %x0, <16 x i8> %x1, <16 x i8> %x2, i16 %x3) {
 ; X86-LABEL: test_int_x86_avx512_mask_vpermi2var_qi_128:
 ; X86:       # %bb.0:
 ; X86-NEXT:    vmovdqa %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xd8]
 ; X86-NEXT:    vpermt2b %xmm2, %xmm1, %xmm3 # encoding: [0x62,0xf2,0x75,0x08,0x7d,0xda]
 ; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1 # encoding: [0xc5,0xf8,0x90,0x4c,0x24,0x04]
 ; X86-NEXT:    vpermi2b %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x75,0xca]
-; X86-NEXT:    vpxor %xmm4, %xmm4, %xmm4 # EVEX TO VEX Compression encoding: [0xc5,0xd9,0xef,0xe4]
-; X86-NEXT:    vpermi2b %xmm2, %xmm0, %xmm4 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x75,0xe2]
+; X86-NEXT:    vmovdqa %xmm1, %xmm4 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xe1]
+; X86-NEXT:    vpermi2b %xmm2, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x75,0xe2]
 ; X86-NEXT:    vpaddb %xmm3, %xmm4, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xd9,0xfc,0xc3]
 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfc,0xc0]
 ; X86-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-LABEL: test_int_x86_avx512_mask_vpermi2var_qi_128:
 ; X64:       # %bb.0:
 ; X64-NEXT:    vmovdqa %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xd8]
 ; X64-NEXT:    vpermt2b %xmm2, %xmm1, %xmm3 # encoding: [0x62,0xf2,0x75,0x08,0x7d,0xda]
 ; X64-NEXT:    kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
 ; X64-NEXT:    vpermi2b %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x75,0xca]
-; X64-NEXT:    vpxor %xmm4, %xmm4, %xmm4 # EVEX TO VEX Compression encoding: [0xc5,0xd9,0xef,0xe4]
-; X64-NEXT:    vpermi2b %xmm2, %xmm0, %xmm4 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x75,0xe2]
+; X64-NEXT:    vmovdqa %xmm1, %xmm4 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xe1]
+; X64-NEXT:    vpermi2b %xmm2, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x75,0xe2]
 ; X64-NEXT:    vpaddb %xmm3, %xmm4, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xd9,0xfc,0xc3]
 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfc,0xc0]
 ; X64-NEXT:    retq # encoding: [0xc3]
   %res = call <16 x i8> @llvm.x86.avx512.mask.vpermi2var.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.vpermi2var.qi.128(<16 x i8> %x0, <16 x i8> zeroinitializer, <16 x i8> %x2, i16 %x3)
+  %res1 = call <16 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.128(<16 x i8> %x0, <16 x i8> %res, <16 x i8> %x2, i16 %x3)
   %res2 = call <16 x i8> @llvm.x86.avx512.mask.vpermi2var.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.vpermi2var.qi.256(<32 x i8>, <32 x i8>, <32 x i8>, i32)
 
 define <32 x i8>@test_int_x86_avx512_mask_vpermi2var_qi_256(<32 x i8> %x0, <32 x i8> %x1, <32 x i8> %x2, i32 %x3) {
 ; X86-LABEL: test_int_x86_avx512_mask_vpermi2var_qi_256:
 ; X86:       # %bb.0:
 ; X86-NEXT:    vmovdqa %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xd8]
 ; X86-NEXT:    vpermt2b %ymm2, %ymm1, %ymm3 # encoding: [0x62,0xf2,0x75,0x28,0x7d,0xda]
 ; X86-NEXT:    kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
 ; X86-NEXT:    vpermi2b %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x75,0xca]
-; X86-NEXT:    vpxor %xmm4, %xmm4, %xmm4 # EVEX TO VEX Compression encoding: [0xc5,0xd9,0xef,0xe4]
-; X86-NEXT:    vpermi2b %ymm2, %ymm0, %ymm4 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x75,0xe2]
+; X86-NEXT:    vmovdqa %ymm1, %ymm4 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xe1]
+; X86-NEXT:    vpermi2b %ymm2, %ymm0, %ymm4 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x75,0xe2]
 ; X86-NEXT:    vpaddb %ymm3, %ymm4, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xdd,0xfc,0xc3]
 ; X86-NEXT:    vpaddb %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfc,0xc0]
 ; X86-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-LABEL: test_int_x86_avx512_mask_vpermi2var_qi_256:
 ; X64:       # %bb.0:
 ; X64-NEXT:    vmovdqa %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xd8]
 ; X64-NEXT:    vpermt2b %ymm2, %ymm1, %ymm3 # encoding: [0x62,0xf2,0x75,0x28,0x7d,0xda]
 ; X64-NEXT:    kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
 ; X64-NEXT:    vpermi2b %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x75,0xca]
-; X64-NEXT:    vpxor %xmm4, %xmm4, %xmm4 # EVEX TO VEX Compression encoding: [0xc5,0xd9,0xef,0xe4]
-; X64-NEXT:    vpermi2b %ymm2, %ymm0, %ymm4 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x75,0xe2]
+; X64-NEXT:    vmovdqa %ymm1, %ymm4 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xe1]
+; X64-NEXT:    vpermi2b %ymm2, %ymm0, %ymm4 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x75,0xe2]
 ; X64-NEXT:    vpaddb %ymm3, %ymm4, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xdd,0xfc,0xc3]
 ; X64-NEXT:    vpaddb %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf5,0xfc,0xc0]
 ; X64-NEXT:    retq # encoding: [0xc3]
   %res = call <32 x i8> @llvm.x86.avx512.mask.vpermi2var.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.vpermi2var.qi.256(<32 x i8> %x0, <32 x i8> zeroinitializer, <32 x i8> %x2, i32 %x3)
+  %res1 = call <32 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.256(<32 x i8> %x0, <32 x i8> %res, <32 x i8> %x2, i32 %x3)
   %res2 = call <32 x i8> @llvm.x86.avx512.mask.vpermi2var.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
 }
 
 declare <16 x i8> @llvm.x86.avx512.mask.vpermt2var.qi.128(<16 x i8>, <16 x i8>, <16 x i8>, i16)