[AMDGPU] Constrain the inliner on maximum number of basic blocks in a caller function

[vpykhtin]

Summary

This patch improves compile time performance: luxmark hotel scene compilation time is reduced from 250 to 6 sec.

The actual check is performed last to give a change to inline what have to be always inlined.

[rampitec]

lib/Target/AMDGPU/AMDGPUInline.cpp

53 ↗

(On Diff #203187)

Use prefix amdgpu.

[arsenm]

lib/Target/AMDGPU/AMDGPUInline.cpp

53 ↗

(On Diff #203187)

Should add a comment explaining this is a compile time hack

test/CodeGen/AMDGPU/inline-maxbb.ll

1–2 ↗

(On Diff #203187)

Should only run the amdgpu-inline pass, not all of -O3

[vpykhtin]

updated diff with per review issues done.

[rampitec]

LGTM

llvm/trunk/lib/Target/AMDGPU/AMDGPUInline.cpp

 
 // If the amount of scratch memory to eliminate exceeds our ability to allocate
 // it into registers we gain nothing by aggressively inlining functions for that
 // heuristic.
 static cl::opt<unsigned>
 ArgAllocaCutoff("amdgpu-inline-arg-alloca-cutoff", cl::Hidden, cl::init(256),
                 cl::desc("Maximum alloca size to use for inline cost"));
 
+// Inliner constraint to achieve reasonable compilation time
+static cl::opt<size_t>
+MaxBB("amdgpu-inline-max-bb", cl::Hidden, cl::init(300),
+      cl::desc("Maximum BB number allowed in a function after inlining"
+               " (compile time constraint)"));
+
 namespace {
 
 class AMDGPUInliner : public LegacyInlinerBase {
 
 public:
   AMDGPUInliner() : LegacyInlinerBase(ID) {
     initializeAMDGPUInlinerPass(*PassRegistry::getPassRegistry());
     Params = getInlineParams();
InlineCost AMDGPUInliner::getInlineCost(CallSite CS) {
   }
 
   OptimizationRemarkEmitter ORE(Caller);
   std::function<AssumptionCache &(Function &)> GetAssumptionCache =
       [this](Function &F) -> AssumptionCache & {
     return ACT->getAssumptionCache(F);
   };
 
-  return llvm::getInlineCost(cast<CallBase>(*CS.getInstruction()), Callee,
+  auto IC = llvm::getInlineCost(cast<CallBase>(*CS.getInstruction()), Callee,
                              LocalParams, TTI, GetAssumptionCache, None, PSI,
                              RemarksEnabled ? &ORE : nullptr);
+
+  if (IC && !IC.isAlways()) {
+    // Single BB does not increase total BB amount, thus subtract 1
+    size_t Size = Caller->size() + Callee->size() - 1;
+    if (MaxBB && Size > MaxBB)
+      return llvm::InlineCost::getNever("max number of bb exceeded");
+  }
+  return IC;
 }

llvm/trunk/test/CodeGen/AMDGPU/inline-maxbb.ll

+; RUN: opt -mtriple=amdgcn-- --amdgpu-inline -S -amdgpu-inline-max-bb=2 %s | FileCheck %s --check-prefix=NOINL
+; RUN: opt -mtriple=amdgcn-- --amdgpu-inline -S -amdgpu-inline-max-bb=3 %s | FileCheck %s --check-prefix=INL
+
+define i32 @callee(i32 %x) {
+entry:
+  %cc = icmp eq i32 %x, 1
+  br i1 %cc, label %ret_res, label %mulx
+
+mulx:
+  %mul1 = mul i32 %x, %x
+  %mul2 = mul i32 %mul1, %x
+  %mul3 = mul i32 %mul1, %mul2
+  %mul4 = mul i32 %mul3, %mul2
+  %mul5 = mul i32 %mul4, %mul3
+  br label %ret_res
+
+ret_res:
+  %r = phi i32 [ %mul5, %mulx ], [ %x, %entry ]
+  ret i32 %r
+}
+
+; INL-LABEL: @caller
+; NOINL-LABEL: @caller
+; INL: mul i32
+; INL-NOT: call i32
+; NOINL-NOT: mul i32
+; NOINL: call i32
+
+define amdgpu_kernel void @caller(i32 %x) {
+  %res = call i32 @callee(i32 %x)
+  store volatile i32 %res, i32 addrspace(1)* undef
+  ret void
+}