make code assumption in get_global_id optional

Stateless to stateful optimization checks if the offset of GEP
instruction is positive. This is done with the assumption that the sign
bit of global_id will be always off. This assume interferes with
instcombine pass, so it is skipped unless really needed, that is:
  1) StatelessToStateful pass is enabled, AND
  2) Buffers don't use implicit bufferOffsetArg.

IGC/BiFModule/Headers/bif_flag_controls.h

@@ -35,6 +35,7 @@ BIF_FLAG_CONTROL(bool, ForceL1Prefetch)
 BIF_FLAG_CONTROL(bool, UseHighAccuracyMath)
 BIF_FLAG_CONTROL(bool, EnableSWSrgbWrites)
 BIF_FLAG_CONTROL(int, MaxHWThreadIDPerSubDevice)
+BIF_FLAG_CONTROL(bool, UseAssumeInGetGlobalId)
 BIF_FLAG_CONTROL(int, JointMatrixLoadStoreOpt)
 BIF_FLAG_CONTROL(bool, UseOOBChecks)
 #endif // __BIF_FLAG_CONTROL_H__

IGC/BiFModule/Implementation/workitem.cl

@@ -113,7 +113,8 @@ size_t OVERLOADABLE __intel_GlobalInvocationId(uint dim)
      // We want to show, that the value is positive.
      // On LLVM level, where the signedness of the type is lost, the only way to prove
      // that the value is positive is to check the sign bit.
-     __builtin_assume((v & 0x8000000000000000ULL) == 0);
+     if (BIF_FLAG_CTRL_GET(UseAssumeInGetGlobalId))
+        __builtin_assume((v & 0x8000000000000000ULL) == 0);
 #endif
 
     return v;

IGC/Compiler/Builtins/BIFFlagCtrl/BIFFlagCtrlResolution.cpp

@@ -92,6 +92,18 @@ void BIFFlagCtrlResolution::FillFlagCtrl() {
     BIF_FLAG_CTRL_SET(UseHighAccuracyMath, false);
   }
 
+  // Stateless to stateful optimization checks if the offset of GEP instruction is positive.
+  // This is done with the assumption that the sign bit of global_id will be always off.
+  // This assume interferes with instcombine pass, so it is skipped unless really needed, that is:
+  //   1) StatelessToStateful pass is enabled, AND
+  //   2) Buffers don't use implicit bufferOffsetArg.
+  bool useAssumeInGetGlobalId = PtrCGC->m_DriverInfo.SupportsStatelessToStatefulBufferTransformation() &&
+      !PtrCGC->getModuleMetaData()->compOpt.GreaterThan4GBBufferRequired &&
+      IGC_IS_FLAG_ENABLED(EnableStatelessToStateful) &&
+      !((IGC_IS_FLAG_ENABLED(EnableSupportBufferOffset) || PtrCGC->getModuleMetaData()->compOpt.HasBufferOffsetArg)) &&
+      !PtrCGC->getModuleMetaData()->compOpt.OptDisable;
+  BIF_FLAG_CTRL_SET(UseAssumeInGetGlobalId, useAssumeInGetGlobalId);
+
   BIF_FLAG_CTRL_SET(EnableSWSrgbWrites,
                     IGC_GET_FLAG_VALUE(cl_khr_srgb_image_writes));
   BIF_FLAG_CTRL_SET(MaxHWThreadIDPerSubDevice,

IGC/Compiler/Builtins/BIFFlagCtrl/BIFFlagCtrlResolution.hpp

@@ -22,7 +22,7 @@ SPDX-License-Identifier: MIT
   BIF_FLAG_CTRL_N_S(BIF_FLAG_NAME),
 
 #define BIF_FLAG_CTRL_SET(BIF_FLAG_NAME, BIF_FLAG_VALUE)                          \
-  ListDelegates.emplace(BIF_FLAG_CTRL_N_S(BIF_FLAG_NAME), [this]() -> bool {      \
+  ListDelegates.emplace(BIF_FLAG_CTRL_N_S(BIF_FLAG_NAME), [&]() -> bool {      \
     return replace(BIF_FLAG_VALUE,                                                \
                    pModule->getGlobalVariable(BIF_FLAG_CTRL_N_S(BIF_FLAG_NAME))); \
   })

IGC/VectorCompiler/lib/GenXOpts/CMTrans/GenXBIFFlagCtrlResolution.cpp

@@ -77,6 +77,7 @@ void GenXBIFFlagCtrlResolution::FillFlagCtrl() {
   BIF_FLAG_CTRL_SET(ProfilingTimerResolution, 0.0f);
   BIF_FLAG_CTRL_SET(UseMathWithLUT, false);
   BIF_FLAG_CTRL_SET(UseHighAccuracyMath, false);
+  BIF_FLAG_CTRL_SET(UseAssumeInGetGlobalId, true);
   // FIXME: target specific, but subtarget cannot be reached in middle-end.
   BIF_FLAG_CTRL_SET(HasInt64SLMAtomicCAS, false);
   BIF_FLAG_CTRL_SET(JointMatrixLoadStoreOpt, 3);