Fix predicated store sub-DW value handling

This change addresses the handling of predicated
stores for sub-DW values with non-uniform stored values.
Predicate alone is not enough to calculate the correct
offset. So, we use `EMASK & Predicate` to determine the
correct offset.

Author: YuriPlyakhin

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -12368,12 +12368,17 @@ CVariable *EmitPass::UniformCopy(CVariable *var, bool doSub) {
 
 /// Uniform copy allowing to reuse the off calculated by a previous call
 /// This allow avoiding redundant code
-CVariable *EmitPass::UniformCopy(CVariable *var, CVariable *&off, CVariable *eMask, bool doSub, bool safeGuard) {
+CVariable *EmitPass::UniformCopy(CVariable *var, CVariable *&off, CVariable *eMask, bool doSub, bool safeGuard,
+                                 CVariable *predicate) {
   IGC_ASSERT_MESSAGE(!var->IsUniform(), "Expect non-uniform source!");
 
   if (eMask == nullptr) {
     eMask = GetExecutionMask();
   }
+  if (predicate != nullptr) {
+    m_encoder->And(eMask, eMask, predicate);
+    m_encoder->Push();
+  }
   if (off == nullptr) {
     // Get offset to any 1s. For simplicity, use 'fbl' to find the lowest 1s.
     off = m_currShader->GetNewVariable(1, ISA_TYPE_UD, EALIGN_DWORD, true, CName::NONE);
@@ -17762,8 +17767,8 @@ void EmitPass::emitLSCVectorStore_subDW(LSC_CACHE_OPTS CacheOpts, bool UseA32, R
     if (!srcUniform) {
       if (predicate) {
         CVariable *offset = nullptr;
-        CVariable *eMask = CastFlagToVariable(predicateVar);
-        stVar = UniformCopy(stVar, offset, eMask);
+        CVariable *pMask = CastFlagToVariable(predicateVar);
+        stVar = UniformCopy(stVar, offset, nullptr, false, false, pMask);
       } else {
         stVar = UniformCopy(stVar);
       }

IGC/Compiler/CISACodeGen/EmitVISAPass.hpp

@@ -653,7 +653,7 @@ class EmitPass : public llvm::FunctionPass {
   CVariable *GetHalfExecutionMask();
   CVariable *UniformCopy(CVariable *var, bool doSub = false);
   CVariable *UniformCopy(CVariable *var, CVariable *&LaneOffset, CVariable *eMask = nullptr, bool doSub = false,
-                         bool safeGuard = false);
+                         bool safeGuard = false, CVariable *predicate = nullptr);
 
   // generate loop header to process sample instruction with varying resource/sampler
   bool ResourceLoopHeader(const CVariable *destination, ResourceDescriptor &resource, SamplerDescriptor &sampler,

IGC/Compiler/tests/EmitVISAPass/predicated-store-subdw.ll

@@ -24,7 +24,11 @@
 ; CHECK: cmp.eq (M1, 32) [[F_LOC_IDX_MATCH:.*]] [[LOCAL_IDX]](0,0)<1;1,0> loc_idx(0,0)<0;1,0>
 ; CHECK: and (M1, 32) [[F_LOC_IDX_MATCH]] [[F_LOC_IDX_MATCH]] [[F_LOC_IDYZ]]
 ; CHECK: mov (M1_NM, 1) [[FVAR:.*]](0,0)<1> [[F_LOC_IDX_MATCH]]
-; CHECK: fbl (M1_NM, 1) [[OFFSET_ALIAS]](0,0)<1> [[FVAR]](0,0)<0;1,0>
+; CHECK: setp (M1_NM, 32) [[F_EMASK:.*]] 0x0:ud
+; CHECK: cmp.eq (M1, 32) [[F_EMASK]] [[DUMMY:.*]](0,0)<0;1,0> [[DUMMY]](0,0)<0;1,0>
+; CHECK: mov (M1_NM, 1) [[VAR_EMASK:.*]](0,0)<1> [[F_EMASK]]
+; CHECK: and (M1_NM, 1) [[VAR_EMASK]](0,0)<1> [[VAR_EMASK]](0,0)<0;1,0> [[FVAR]](0,0)<0;1,0>
+; CHECK: fbl (M1_NM, 1) [[OFFSET_ALIAS]](0,0)<1> [[VAR_EMASK]](0,0)<0;1,0>
 ; CHECK: shl (M1_NM, 1) [[OFFSET_ALIAS]](0,0)<1> [[OFFSET_ALIAS]](0,0)<0;1,0> 0x1:w
 ; CHECK: addr_add (M1_NM, 1) [[ADDR:.*]](0)<1> &{{.*}} [[OFFSET]](0,0)<0;1,0>
 ; CHECK: mov (M1_NM, 1) [[DATA_ALIAS]](0,0)<1> r[[[ADDR]](0),0]<0;1,0>:hf