Enable trampoline insertion pass by default

TrampolineInsertion pass must be enabled by default to make VISA and
binary linking possible.

Author: Anton Sidorenko

IGC/VectorCompiler/include/vc/Utils/GenX/KernelInfo.h

@@ -9,7 +9,9 @@ SPDX-License-Identifier: MIT
 #ifndef VC_UTILS_GENX_KERNELINFO_H
 #define VC_UTILS_GENX_KERNELINFO_H
 
+#include "vc/InternalIntrinsics/InternalIntrinsics.h"
 #include "vc/Utils/GenX/InternalMetadata.h"
+#include "vc/Utils/GenX/Intrinsics.h"
 #include "vc/Utils/GenX/RegCategory.h"
 
 #include "Probe/Assertion.h"
@@ -88,13 +90,18 @@ inline bool isIndirect(const llvm::Function *F) {
 // structure types is fixed for intrinsics.
   if (llvm::GenXIntrinsic::isAnyNonTrivialIntrinsic(F))
     return false;
-  // FIXME: The condition of which function is considered to be indirectly
-  // called will be changed soon.
-  bool IsIndirect = F->hasAddressTaken();
+  if (vc::InternalIntrinsic::isInternalNonTrivialIntrinsic(F))
+    return false;
+  if (vc::isKernel(F))
+    return false;
+  if (vc::isEmulationFunction(*F))
+    return false;
+  if (!F->hasAddressTaken() && F->hasLocalLinkage())
+    return false;
   IGC_ASSERT_MESSAGE(
-      !IsIndirect || requiresStackCall(F),
+      requiresStackCall(F),
       "The indirectly-called function is expected to be a stack call");
-  return IsIndirect;
+  return true;
 }
 
 inline bool isIndirect(const llvm::Function &F) { return isIndirect(&F); }

IGC/VectorCompiler/lib/GenXCodeGen/GenXCisaBuilder.cpp

@@ -5431,10 +5431,6 @@ void GenXKernelBuilder::buildInlineAsm(CallInst *CI) {
 void GenXKernelBuilder::buildCall(CallInst *CI, const DstOpndDesc &DstDesc) {
   LLVM_DEBUG(dbgs() << CI << "\n");
   Function *Callee = CI->getCalledFunction();
-  IGC_ASSERT_MESSAGE(
-      !Callee || !Callee->isDeclaration(),
-      "Currently VC backend does not support modules with external functions");
-
   if (!Callee || vc::requiresStackCall(Callee)) {
     if (UseNewStackBuilder)
       buildStackCallLight(CI, DstDesc);

IGC/VectorCompiler/lib/GenXCodeGen/GenXGlobalValueLowering.cpp

@@ -208,10 +208,8 @@ void GenXGlobalValueLowering::fillWorkListForGVInstUse(Use &GVUse) {
                          GenXIntrinsic::genx_gaddr,
                      "llvm.gaddr must be inserted by this pass, but someone "
                      "inserted it before");
-  // Skipping direct calls. Relocating will make them indirect.
-  if (isa<Function>(ConstWithGV) && isa<CallInst>(Usr) &&
-      cast<CallInst>(Usr)->getCalledFunction() == cast<Function>(ConstWithGV))
-    return;
+  // Do not skip direct calls as an indirectly-called function can be called
+  // only indirectly. See GenXTrampolineInsertion
   Function *Func = Usr->getParent()->getParent();
   WorkList[Func].Users.insert({Usr, GVUse.getOperandNo()});
   WorkList[Func].Replacement[ConstWithGV] = nullptr;

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

@@ -52,6 +52,7 @@ SPDX-License-Identifier: MIT
 
 #include "llvmWrapper/IR/Value.h"
 #include "vc/GenXOpts/GenXOpts.h"
+#include "vc/InternalIntrinsics/InternalIntrinsics.h"
 #include "vc/Support/BackendConfig.h"
 #include "vc/Utils/GenX/Intrinsics.h"
 #include "vc/Utils/GenX/KernelInfo.h"
@@ -67,7 +68,7 @@ using namespace llvm;
 
 static cl::opt<bool> EnableTrampolineInsertion(
     "vc-enable-trampoline-insertion",
-    llvm::cl::desc("Enable/disable GenXTrampolineInsertion"), cl::init(false),
+    llvm::cl::desc("Enable/disable GenXTrampolineInsertion"), cl::init(true),
     cl::Hidden);
 
 namespace {
@@ -99,6 +100,8 @@ class GenXTrampolineInsertion : public ModulePass,
 void GenXTrampolineInsertion::visitFunction(Function &F) {
   if (GenXIntrinsic::isAnyNonTrivialIntrinsic(&F))
     return;
+  if (vc::InternalIntrinsic::isInternalNonTrivialIntrinsic(&F))
+    return;
   if (vc::isEmulationFunction(F))
     return;
   if (vc::isKernel(&F))

IGC/common/igc_flags.h

@@ -767,7 +767,7 @@ DECLARE_IGC_GROUP("VectorCompiler Options")
     DECLARE_IGC_REGKEY(DWORD, VCNoOptFinalizerControl, 0, "Controls if finalizer is invoked with -debug flag", true)
     DECLARE_IGC_REGKEY(DWORD, VCDisableLRCoalescingControl, 0, "Controls if LR coalescing", true)
     DECLARE_IGC_REGKEY(DWORD, VCDisableExtraCoalescing, 0, "Disable extra coalescing", true)
-    DECLARE_IGC_REGKEY(bool, VCSaveStackCallLinkage, false,
+    DECLARE_IGC_REGKEY(bool, VCSaveStackCallLinkage, true,
                        "Do not override stack calls linkage as internal", true)
     DECLARE_IGC_REGKEY(bool, VCDirectCallsOnly, false, "Generate code under the assumption all unknown calls are direct", true)
     DECLARE_IGC_REGKEY(DWORD, VCLoopUnrollThreshold, 0, "Set the loop unroll threshold for VC. Value 0 will use the default threshold.", true)