Reset sub_group_size when using CallWA

Stackcalls now supports SIMD32 by default, but if the EUFusion Call WA is required, we still need to default stackcall compilation to SIMD16.
However, when the Call WA is enabled and sub_group_size(32) is set, we need to also reset to sub_group_size(16) so that SIMD16 compilation will succeed.

Author: dlei6g

IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp

@@ -3817,15 +3817,6 @@ namespace IGC
             }
         }
 
-        {
-            // If stack calls are present, disable simd32 in order to do wa in visa
-            bool needCallWA = (IGC_IS_FLAG_ENABLED(EnableCallWA) && m_Context->platform.hasFusedEU());
-            if (needCallWA && simdMode == SIMDMode::SIMD32 && HasStackCalls())
-            {
-                return false;
-            }
-        }
-
         if (!m_Context->m_retryManager.IsFirstTry())
         {
             m_Context->ClearSIMDInfo(simdMode, ShaderDispatchMode::NOT_APPLICABLE);
@@ -4093,6 +4084,24 @@ namespace IGC
         bool isIndirectGroup = m_FGA && m_FGA->getGroup(&F) && IGC::isIntelSymbolTableVoidProgram(m_FGA->getGroupHead(&F));
         bool hasSubroutine = m_FGA && m_FGA->getGroup(&F) && !m_FGA->getGroup(&F)->isSingle() && !hasStackCall && !isIndirectGroup;
 
+        // If stack calls are present, disable simd32 in order to do CallWA in visa
+        if (IGC_IS_FLAG_ENABLED(EnableCallWA) &&
+            pCtx->platform.hasFusedEU() &&
+            pCtx->platform.getWATable().Wa_14016243945 == false &&
+            simdMode == SIMDMode::SIMD32 &&
+            (hasStackCall || isIndirectGroup))
+        {
+            // If sub_group_size is set to 32, resize it to 16 so SIMD16 compilation will still succeed
+            if (simd_size == 32)
+            {
+                llvm::Function* Kernel = m_FGA->getGroup(&F)->getHead();
+                funcInfoMD = pMdUtils->getFunctionsInfoItem(Kernel);
+                funcInfoMD->getSubGroupSize()->setSIMD_size(16);
+            }
+            pCtx->SetSIMDInfo(SIMD_SKIP_HW, simdMode, ShaderDispatchMode::NOT_APPLICABLE);
+            return SIMDStatus::SIMD_FUNC_FAIL;
+        }
+
         if (simd_size == 0)
         {
             // Default to lowest SIMD mode for stack calls/indirect calls