GEP canon is on only for OCL

Revert "turning GEP canon off" for non OCL shaders for now, as turning off
GEP canon has perf regression on non-ocl shaders.

This is partial revert and it remains on for OCL

Author: jgu222

IGC/Compiler/CISACodeGen/MemOpt.cpp

@@ -321,6 +321,22 @@ namespace {
             return true;
         }
 
+        bool EnableCanonicalizeGEP() const {
+            switch (IGC_GET_FLAG_VALUE(MemOptGEPCanon)) {
+            case 1:
+                return false;
+            case 2:
+            {
+                if (CGC && CGC->type == ShaderType::OPENCL_SHADER)
+                    return false;
+                break;
+            }
+            default:
+                break;
+            }
+            return true;
+        }
+
         /// Canonicalize the calculation of 64-bit pointer by performing the
         /// following transformations to help SCEV to identify the constant offset
         /// between pointers.
@@ -526,7 +542,7 @@ bool MemOpt::runOnFunction(Function& F) {
         if (MemRefs.size() < 2)
             continue;
 
-        if (IGC_IS_FLAG_ENABLED(EnableMemOptGEPCanon)) {
+        if (EnableCanonicalizeGEP()) {
             // Canonicalize 64-bit GEP to help SCEV find constant offset by
             // distributing `zext`/`sext` over safe expressions.
             for (auto& M : MemRefs)
@@ -553,7 +569,7 @@ bool MemOpt::runOnFunction(Function& F) {
             }
         }
 
-        if (IGC_IS_FLAG_ENABLED(EnableMemOptGEPCanon)) {
+        if (EnableCanonicalizeGEP()) {
             // Optimize 64-bit GEP to reduce strength by factoring out `zext`/`sext`
             // over safe expressions.
             for (auto I : MemRefsToOptimize)
@@ -1137,7 +1153,7 @@ bool MemOpt::mergeLoad(LoadInst* LeadingLoad,
         return false;
     const SCEV* LeadingLastIdx = nullptr; // set on-demand
     bool DoCmpOnLastIdx = false;
-    if (IGC_IS_FLAG_DISABLED(EnableMemOptGEPCanon)) {
+    if (!EnableCanonicalizeGEP()) {
         auto aGEP = dyn_cast<GetElementPtrInst>(LeadingLoad->getPointerOperand());
         if (aGEP && aGEP->hasIndices()) {
             // index starts from 1
@@ -1305,8 +1321,8 @@ bool MemOpt::mergeLoad(LoadInst* LeadingLoad,
             MaxElts = profitVec[k++];
         }
 
-        if (IGC_IS_FLAG_ENABLED(EnableMemOptGEPCanon)) {
-            // Guard under the key to distinguish new code (EnableMemOptGEPCanon=0) from the old.
+        if (EnableCanonicalizeGEP()) {
+            // Guard under the key to distinguish new code (GEPCanon is off) from the old.
             //    Note: not sure about the reason for the following check.
             if (NumElts == 3 && (LeadingLoadScalarType->isIntegerTy(16) || LeadingLoadScalarType->isHalfTy())) {
                 return false;

IGC/common/igc_flags.h

@@ -436,7 +436,7 @@ DECLARE_IGC_REGKEY(bool, EnableHSSinglePatchDispatch,   false, "Setting this to
 DECLARE_IGC_REGKEY(bool, DisableGPGPUIndirectPayload,   false, "Disable OCL indirect GPGPU payload", false)
 DECLARE_IGC_REGKEY(bool, DisableDSDualPatch,            false, "Setting it to true with enable Single and Dual Patch dispatch mode for Domain Shader", false)
 DECLARE_IGC_REGKEY(bool, DisableMemOpt,                 false, "Disable MemOpt, merging load/store", true)
-DECLARE_IGC_REGKEY(bool, EnableMemOptGEPCanon,          false, "[test] Enable GEP canonicalization in MemOpt", true)
+DECLARE_IGC_REGKEY(DWORD, MemOptGEPCanon,               2,     "[test] GEP canonicalization in MemOpt. 0 : enable; 1: disable; 2: disable only for OCL;", true)
 DECLARE_IGC_REGKEY(bool, DisableMemOpt2,                false, "Disable MemOpt2", false)
 DECLARE_IGC_REGKEY(DWORD, EnableLdStCombine,            1,     "Enable load/store combine pass if set to 1 or 2 (intend to replace memopt)", true)
 DECLARE_IGC_REGKEY(DWORD, MaxStoreVectorSizeInBytes,    0,     "[LdStCombine] the max non-uniform vector size for the coalesced store. 0: compiler choice (default, 16(4DW)); others: 4/8/16/32", true)