[X86] Fix decoding for vinsertps immediate operand

The relevant bit from the Intel SDM for vinsertps semantics:
```
IF (SRC = REG) THEN COUNT_S := imm8[7:6] ELSE COUNT_S := 0
```

This is now taken into account.

Author: omern1

llvm/lib/Target/X86/MCTargetDesc/X86InstComments.cpp

@@ -1122,15 +1122,21 @@ bool llvm::EmitAnyX86InstComments(const MCInst *MI, raw_ostream &OS,
   case X86::VINSERTPSrri:
   case X86::VINSERTPSZrri:
     Src2Name = getRegName(MI->getOperand(2).getReg());
-    [[fallthrough]];
+    DestName = getRegName(MI->getOperand(0).getReg());
+    Src1Name = getRegName(MI->getOperand(1).getReg());
+    if (MI->getOperand(NumOperands - 1).isImm())
+      DecodeINSERTPSMask(MI->getOperand(NumOperands - 1).getImm(),
+                         ShuffleMask, false);
+    break;
+
   case X86::INSERTPSrmi:
   case X86::VINSERTPSrmi:
   case X86::VINSERTPSZrmi:
     DestName = getRegName(MI->getOperand(0).getReg());
     Src1Name = getRegName(MI->getOperand(1).getReg());
     if (MI->getOperand(NumOperands - 1).isImm())
       DecodeINSERTPSMask(MI->getOperand(NumOperands - 1).getImm(),
-                         ShuffleMask);
+                         ShuffleMask, true);
     break;
 
   case X86::MOVLHPSrr:

llvm/lib/Target/X86/MCTargetDesc/X86ShuffleDecode.cpp

@@ -23,7 +23,7 @@
 
 namespace llvm {
 
-void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
+void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask, bool SrcIsMem) {
   // Defaults the copying the dest value.
   ShuffleMask.push_back(0);
   ShuffleMask.push_back(1);
@@ -33,7 +33,7 @@ void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
   // Decode the immediate.
   unsigned ZMask = Imm & 15;
   unsigned CountD = (Imm >> 4) & 3;
-  unsigned CountS = (Imm >> 6) & 3;
+  unsigned CountS = SrcIsMem ? 0 : (Imm >> 6) & 3;
 
   // CountS selects which input element to use.
   unsigned InVal = 4 + CountS;

llvm/lib/Target/X86/MCTargetDesc/X86ShuffleDecode.h

@@ -28,7 +28,7 @@ template <typename T> class SmallVectorImpl;
 enum { SM_SentinelUndef = -1, SM_SentinelZero = -2 };
 
 /// Decode a 128-bit INSERTPS instruction as a v4f32 shuffle mask.
-void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask);
+void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask, bool SrcIsMem);
 
 // Insert the bottom Len elements from a second source into a vector starting at
 // element Idx.

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -5362,7 +5362,7 @@ static bool getTargetShuffleMask(SDValue N, bool AllowSentinelZero,
     assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
     assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
     ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
-    DecodeINSERTPSMask(ImmN, Mask);
+    DecodeINSERTPSMask(ImmN, Mask, false);
     IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
     break;
   case X86ISD::EXTRQI:

llvm/test/MC/X86/vinsertps_decode.s

@@ -0,0 +1,7 @@
+# RUN: llvm-mc -triple x86_64-unknown-unknown %s | FileCheck %s
+
+.intel_syntax
+
+# CHECK:  vinsertps {{.*}} # xmm2 = xmm2[0,1,2],mem[0]
+
+vinsertps xmm2,xmm2,dword ptr [r14+rdi*8+0x4C],0x0B0