diff --git a/src/coreclr/jit/emitxarch.cpp b/src/coreclr/jit/emitxarch.cpp
index f776d7defe4d1a..30b5ea482838b8 100644
--- a/src/coreclr/jit/emitxarch.cpp
+++ b/src/coreclr/jit/emitxarch.cpp
@@ -5195,7 +5195,11 @@ inline UNATIVE_OFFSET emitter::emitInsSizeSVCalcDisp(instrDesc* id, code_t code,
         {
             ssize_t compressedDsp;
 
-            if (TryEvexCompressDisp8Byte(id, dsp, &compressedDsp, &dspInByte))
+            // Only the scaling factor of the original EVEX instructions can be changed by embedded broadcast.
+            // If the instruction does not have tuple type info, say extended EVEX from APX, the scaling factor is
+            // constantly 1, then this optimization cannot be performed, and whether disp8 or disp32 should be applied
+            // only depends dspInByte.
+            if (TryEvexCompressDisp8Byte(id, dsp, &compressedDsp, &dspInByte) && hasTupleTypeInfo(ins))
             {
                 SetEvexCompressedDisplacement(id);
             }
@@ -5368,7 +5372,7 @@ UNATIVE_OFFSET emitter::emitInsSizeAM(instrDesc* id, code_t code)
     {
         ssize_t compressedDsp;
 
-        if (TryEvexCompressDisp8Byte(id, dsp, &compressedDsp, &dspInByte))
+        if (TryEvexCompressDisp8Byte(id, dsp, &compressedDsp, &dspInByte) && hasTupleTypeInfo(ins))
         {
             SetEvexCompressedDisplacement(id);
         }
@@ -14672,13 +14676,16 @@ BYTE* emitter::emitOutputAM(BYTE* dst, instrDesc* id, code_t code, CnsVal* addc)
             assert(isCompressed && dspInByte);
             dsp = compressedDsp;
         }
-        else if (TakesEvexPrefix(id) || TakesApxExtendedEvexPrefix(id))
+        else if (TakesEvexPrefix(id) && !IsBMIInstruction(ins))
         {
-            assert(!TryEvexCompressDisp8Byte(id, dsp, &compressedDsp, &dspInByte));
+            assert(!(TryEvexCompressDisp8Byte(id, dsp, &compressedDsp, &dspInByte) && hasTupleTypeInfo(ins)));
             dspInByte = false;
         }
         else
         {
+            // TODO-XArch-APX: for now, Extended Evex instruction will not have compressed displacement, or more
+            // accurately, extended evex may not have compressed displacement optimization as the scaling factor is
+            // constantly 1.
             dspInByte = ((signed char)dsp == (ssize_t)dsp);
         }
         dspIsZero = (dsp == 0);
@@ -15556,7 +15563,7 @@ BYTE* emitter::emitOutputSV(BYTE* dst, instrDesc* id, code_t code, CnsVal* addc)
             assert(isCompressed && dspInByte);
             dsp = (int)compressedDsp;
         }
-        else if (TakesEvexPrefix(id) || TakesApxExtendedEvexPrefix(id))
+        else if (TakesEvexPrefix(id) && !IsBMIInstruction(ins))
         {
 #if FEATURE_FIXED_OUT_ARGS
             // TODO-AMD64-CQ: We should be able to accurately predict this when FEATURE_FIXED_OUT_ARGS
@@ -17904,6 +17911,7 @@ BYTE* emitter::emitOutputLJ(insGroup* ig, BYTE* dst, instrDesc* i)
             idAmd->idCodeSize(sz);
 
             code = insCodeRM(ins);
+            code = AddX86PrefixIfNeeded(id, code, id->idOpSize());
             code |= (insEncodeReg345(id, id->idReg1(), EA_PTRSIZE, &code) << 8);
 
             dst = emitOutputAM(dst, idAmd, code, nullptr);
diff --git a/src/coreclr/jit/instr.cpp b/src/coreclr/jit/instr.cpp
index e9cf023943f493..844a621b5cb563 100644
--- a/src/coreclr/jit/instr.cpp
+++ b/src/coreclr/jit/instr.cpp
@@ -692,7 +692,7 @@ void CodeGen::inst_SET(emitJumpKind condition, regNumber reg, insOpts instOption
         assert(INS_setge == (INS_setge_apx + offset));
         assert(INS_setle == (INS_setle_apx + offset));
         assert(INS_setg == (INS_setg_apx + offset));
-        ins = (instruction)(ins + offset);
+        ins = (instruction)(ins - offset);
     }
 #endif
 
diff --git a/src/coreclr/jit/lsraxarch.cpp b/src/coreclr/jit/lsraxarch.cpp
index 47c232540580d9..a947e5546e58c3 100644
--- a/src/coreclr/jit/lsraxarch.cpp
+++ b/src/coreclr/jit/lsraxarch.cpp
@@ -1157,12 +1157,12 @@ int LinearScan::BuildShiftRotate(GenTree* tree)
 #endif
         if (!source->isContained())
     {
-        tgtPrefUse = BuildUse(source, srcCandidates);
+        tgtPrefUse = BuildUse(source, ForceLowGprForApxIfNeeded(source, srcCandidates, getEvexIsSupported()));
         srcCount++;
     }
     else
     {
-        srcCount += BuildOperandUses(source, srcCandidates);
+        srcCount += BuildOperandUses(source, ForceLowGprForApxIfNeeded(source, srcCandidates, getEvexIsSupported()));
     }
 
     if (!tree->isContained())
@@ -1172,6 +1172,9 @@ int LinearScan::BuildShiftRotate(GenTree* tree)
             srcCount += BuildDelayFreeUses(shiftBy, source, SRBM_RCX);
             buildKillPositionsForNode(tree, currentLoc + 1, SRBM_RCX);
         }
+        dstCandidates = (tree->GetRegNum() == REG_NA)
+                            ? ForceLowGprForApxIfNeeded(tree, dstCandidates, getEvexIsSupported())
+                            : dstCandidates;
         BuildDef(tree, dstCandidates);
     }
     else
@@ -3280,8 +3283,8 @@ int LinearScan::BuildMul(GenTree* tree)
             srcCandidates1 = SRBM_RDX;
         }
 
-        srcCount = BuildOperandUses(op1, srcCandidates1);
-        srcCount += BuildOperandUses(op2, srcCandidates2);
+        srcCount = BuildOperandUses(op1, ForceLowGprForApxIfNeeded(op1, srcCandidates1, getEvexIsSupported()));
+        srcCount += BuildOperandUses(op2, ForceLowGprForApxIfNeeded(op2, srcCandidates2, getEvexIsSupported()));
 
 #if defined(TARGET_X86)
         if (tree->OperIs(GT_MUL_LONG))