run early MemCpy Optimization pass

When input SPIR-V does memcpy with src pointing to zeroinitialized
memory, legacy SPIR-V translator did early optimization by replacing
memcpy with memset.

Khronos SPIR-V translator emits memcpy intact. This impacts
AlignmentAnalysis pass, which produces different dst alignment for
memcpy and memset, impacting following optimizations.

Run LLVM's MemCpy Optimization pass before AlignmentAnalysis to replace
memcpy with memset, producing result closer to legacy translator.

(cherry picked from commit 8bc3ee1)

Author: pkwasnie-intel

IGC/AdaptorOCL/UnifyIROCL.cpp

@@ -331,6 +331,7 @@ static void CommonOCLBasedPasses(
 #ifdef IGC_SCALAR_USE_KHRONOS_SPIRV_TRANSLATOR
     mpmSPIR.add(new PreprocessSPVIR());
     mpmSPIR.add(new PromoteBools());
+    mpmSPIR.add(llvm::createMemCpyOptPass());
 #endif // IGC_SCALAR_USE_KHRONOS_SPIRV_TRANSLATOR
     mpmSPIR.add(new TypesLegalizationPass());
     mpmSPIR.add(new TargetLibraryInfoWrapperPass());

IGC/ocloc_tests/SPIRV-Asm/OpCopyMemorySized_alignment.spvasm

@@ -0,0 +1,60 @@
+; Input spv:
+;   1. Defines struct struct.complex = { double, double }
+;   2. Defines local array variable [5 x [6 x struct.complex ]]
+;   3. Defines global array variable [16 x i8] = { 0 }
+;   4. Kernel sets first struct in array to zero by doing memcpy of global i8 array.
+;
+; Legacy translator does early optimization by replacing memcpy to memset.
+; Khronos translator emits memcpy as-it, which is later optimized to memset by compiler.
+; Test if in both cases AlignmentAnalysis pass assigns correct allignment to pointer.
+
+; REQUIRES: regkeys, spirv-as, pvc-supported
+; RUN: spirv-as --target-env spv1.0 -o %t.spv %s
+; RUN: ocloc compile -spirv_input -file %t.spv -device pvc -options " -igc_opts 'PrintToConsole=1 PrintAfter=AlignmentAnalysisPass'" 2>&1 | FileCheck %s
+
+               OpCapability Addresses
+               OpCapability Kernel
+               OpMemoryModel Physical32 OpenCL
+               OpEntryPoint Kernel %1 "test"
+               OpName %cpx "struct.complex" ; 0x00000884
+       %void = OpTypeVoid
+     %double = OpTypeFloat 64
+      %uchar = OpTypeInt 8 0
+       %uint = OpTypeInt 32 0
+      %ulong = OpTypeInt 64 0
+     %uint_0 = OpConstant %uint 0
+    %ulong_0 = OpConstant %ulong 0
+    %ulong_5 = OpConstant %ulong 5
+    %ulong_6 = OpConstant %ulong 6
+   %ulong_16 = OpConstant %ulong 16
+  %ptr_uchar = OpTypePointer UniformConstant %uchar
+  %arr_uchar = OpTypeArray %uchar %ulong_16
+%ptr_arr_uchar = OpTypePointer UniformConstant %arr_uchar
+        %cpx = OpTypeStruct %double %double
+    %arr_cpx = OpTypeArray %cpx %ulong_6
+%arr_arr_cpx = OpTypeArray %arr_cpx %ulong_5
+%ptr_arr_arr_cpx = OpTypePointer Function %arr_arr_cpx
+    %ptr_cpx = OpTypePointer Function %cpx
+ %ptr_double = OpTypePointer Function %double
+   %ptr_char = OpTypePointer Function %uchar
+               OpDecorate %349 Constant
+        %347 = OpConstantNull %arr_uchar
+        %349 = OpVariable %ptr_arr_uchar UniformConstant %347
+          %7 = OpTypeFunction %void
+          %1 = OpFunction %void None %7
+         %10 = OpLabel
+        %317 = OpVariable %ptr_arr_arr_cpx Function
+        %334 = OpInBoundsPtrAccessChain %ptr_cpx %317 %ulong_0 %ulong_0 %ulong_0
+        %339 = OpInBoundsPtrAccessChain %ptr_double %334 %ulong_0 %uint_0
+        %345 = OpBitcast %ptr_char %339
+        %351 = OpBitcast %ptr_uchar %349
+               OpCopyMemorySized %345 %351 %ulong_16 Aligned 0
+               OpReturn
+               OpFunctionEnd
+
+; CHECK-LABEL: @test(
+; CHECK-NOT:     call void @llvm.memcpy
+; CHECK:         [[ALLOCA:%.*]] = alloca %struct.complex
+; CHECK:         [[PTR:%.*]] = bitcast %struct.complex* [[ALLOCA]] to i8*
+; CHECK:         call void @llvm.memset.p0i8.i64(i8* align 8 [[PTR]], i8 0, i64 16, i1 false)
+; CHECK:         ret void