[SYCL] [ESIMD] Aggressively inline every function called by ESIMD kernel (#3261)

This patch is created to overcome a current limitation of Intel
GPU vector backend that requires kernel functions to be inlined
into the kernel itself.

I removed later inlining passes since AlwaysInline pass is supposed
to inline everything that's possible.

Author: DenisBakhvalov

llvm/lib/SYCLLowerIR/LowerESIMD.cpp

@@ -1244,6 +1244,14 @@ PreservedAnalyses SYCLLowerESIMDPass::run(Module &M,
 
 size_t SYCLLowerESIMDPass::runOnFunction(Function &F,
                                          SmallPtrSet<Type *, 4> &GVTS) {
+  // There is a current limitation of GPU vector backend that requires kernel
+  // functions to be inlined into the kernel itself. To overcome this
+  // limitation, mark every function called from ESIMD kernel with
+  // 'alwaysinline' attribute.
+  if ((F.getCallingConv() != CallingConv::SPIR_KERNEL) &&
+      !F.hasFnAttribute(Attribute::AlwaysInline))
+    F.addFnAttr(Attribute::AlwaysInline);
+
   SmallVector<CallInst *, 32> ESIMDIntrCalls;
   SmallVector<Instruction *, 8> ESIMDToErases;
 

llvm/test/SYCLLowerIR/esimd_lower_inline_hint.ll

@@ -0,0 +1,39 @@
+; RUN: opt -LowerESIMD -S < %s | FileCheck %s
+
+; This test checks that LowerESIMD pass sets the 'alwaysinline'
+; attribute for all non-kernel functions.
+
+define spir_kernel void @EsimdKernel1() {
+; CHECK: @EsimdKernel1(
+; CHECK-NEXT:    call void @foo()
+; CHECK-NEXT:    call void @bar()
+  call void @foo()
+  call void @bar()
+  ret void
+}
+
+define spir_kernel void @EsimdKernel2() {
+; CHECK: @EsimdKernel2(
+; CHECK-NEXT:    call void @foobar()
+  call void @foobar()
+  ret void
+}
+
+define spir_func void @foo() {
+; CHECK: @foo() #[[ATTR:[0-9]+]]
+  ret void
+}
+
+define spir_func void @bar() {
+; CHECK: @bar() #[[ATTR]]
+; CHECK-NEXT:    call void @foobar
+  call void @foobar()
+  ret void
+}
+
+define spir_func void @foobar() {
+; CHECK: @foobar() #[[ATTR]]
+  ret void
+}
+
+; CHECK: attributes #[[ATTR]] = { alwaysinline }

llvm/tools/sycl-post-link/sycl-post-link.cpp

@@ -37,6 +37,7 @@
 #include "llvm/Support/SystemUtils.h"
 #include "llvm/Support/WithColor.h"
 #include "llvm/Transforms/IPO.h"
+#include "llvm/Transforms/IPO/AlwaysInliner.h"
 #include "llvm/Transforms/IPO/GlobalDCE.h"
 #include "llvm/Transforms/InstCombine/InstCombine.h"
 #include "llvm/Transforms/Scalar.h"
@@ -489,40 +490,15 @@ static string_vector saveResultSymbolsLists(string_vector &ResSymbolsLists,
     }                                                                          \
   }
 
-// Helper function for creating Inliner pass.
-// The approach is taken from opt tool.
-static Pass *createFunctionInliningPassHelper() {
-  if (OptLevelO0)
-    return createFunctionInliningPass(0, 0, false);
-
-  if (OptLevelO1)
-    return createFunctionInliningPass(1, 0, false);
-
-  if (OptLevelO2)
-    return createFunctionInliningPass(2, 0, false);
-
-  if (OptLevelOs)
-    return createFunctionInliningPass(2, 1, false);
-
-  if (OptLevelOz)
-    return createFunctionInliningPass(2, 2, false);
-
-  if (OptLevelO3)
-    return createFunctionInliningPass(3, 0, false);
-
-  return createFunctionInliningPass();
-}
-
 // When ESIMD code was separated from the regular SYCL code,
 // we can safely process ESIMD part.
 // TODO: support options like -debug-pass, -print-[before|after], and others
 static void LowerEsimdConstructs(Module &M) {
   legacy::PassManager MPM;
   MPM.add(createSYCLLowerESIMDPass());
   if (!OptLevelO0) {
-    // Inlining and SROA passes are required to make
-    // ESIMD/accessor_gather_scatter.cpp test work.
-    MPM.add(createFunctionInliningPassHelper());
+    // Force-inline all functions marked 'alwaysinline' by the LowerESIMD pass.
+    MPM.add(createAlwaysInlinerLegacyPass());
     MPM.add(createSROAPass());
   }
   MPM.add(createESIMDLowerVecArgPass());
@@ -532,7 +508,7 @@ static void LowerEsimdConstructs(Module &M) {
     MPM.add(createEarlyCSEPass(true));
     MPM.add(createInstructionCombiningPass());
     MPM.add(createDeadCodeEliminationPass());
-    MPM.add(createFunctionInliningPassHelper());
+    // TODO: maybe remove some passes below that don't affect code quality
     MPM.add(createSROAPass());
     MPM.add(createEarlyCSEPass(true));
     MPM.add(createInstructionCombiningPass());