[CostModel] Make target intrinsics cheap by default

[dmgreen]

Summary

This patch changes the intrinsics cost model to assume that by default target intrinsics are cheap. This didn't seem to be the case for all intrinsics, and is potentially an MVE problem due to our scalarization overheads. Cheap seems to be a good default in general though.

Found from D90554, which would also end up changing the size costs which then has an effect on loop unrolling and unswitch.

[spatel]

llvm/include/llvm/CodeGen/BasicTTIImpl.h

1135–1140

Do you see any differences if we add the check here?
I've been staring at and stepping through this code for days, and I'm still not exactly sure how it works...
But it seems like if we can stub something out at this level, it's better than waiting for the type-based specialization.

[dmgreen]

Yeah that makes sense. My tests were not unhappy about moving it.

[spatel]

LGTM

llvm/include/llvm/CodeGen/BasicTTIImpl.h

public:
   }
 
   /// Get intrinsic cost based on arguments.
   unsigned getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
                                  TTI::TargetCostKind CostKind) {
     // Check for generically free intrinsics.
     if (BaseT::getIntrinsicInstrCost(ICA, CostKind) == 0)
       return 0;
 
+    // Assume that target intrinsics are cheap
+    Intrinsic::ID IID = ICA.getID();
+    if (Function::isTargetIntrinsic(IID))
+      return TargetTransformInfo::TCC_Basic;
+
     if (ICA.isTypeBasedOnly())
       return getTypeBasedIntrinsicInstrCost(ICA, CostKind);
 
     // TODO: Handle scalable vectors?
     Type *RetTy = ICA.getReturnType();
     if (isa<ScalableVectorType>(RetTy))
       return BaseT::getIntrinsicInstrCost(ICA, CostKind);
 
     unsigned VF = ICA.getVectorFactor();
     unsigned RetVF =
         (RetTy->isVectorTy() ? cast<FixedVectorType>(RetTy)->getNumElements()
                              : 1);
     assert((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
     const IntrinsicInst *I = ICA.getInst();
     const SmallVectorImpl<const Value *> &Args = ICA.getArgs();
     FastMathFlags FMF = ICA.getFlags();
-    Intrinsic::ID IID = ICA.getID();
     switch (IID) {
     default:
       // FIXME: all cost kinds should default to the same thing?
       if (CostKind != TTI::TCK_RecipThroughput)
         return BaseT::getIntrinsicInstrCost(ICA, CostKind);
       break;
 
     case Intrinsic::cttz:

llvm/include/llvm/IR/Function.h

public:
   /// defined in llvm/Intrinsics.h.
   Intrinsic::ID getIntrinsicID() const LLVM_READONLY { return IntID; }
 
   /// isIntrinsic - Returns true if the function's name starts with "llvm.".
   /// It's possible for this function to return true while getIntrinsicID()
   /// returns Intrinsic::not_intrinsic!
   bool isIntrinsic() const { return HasLLVMReservedName; }
 
+  /// isTargetIntrinsic - Returns true if this IID is an intrinsic specific to a
+  /// certain target. If this is a generic intrinsic, false is returned.
+  static bool isTargetIntrinsic(Intrinsic::ID IID);
+
   /// isTargetIntrinsic - Returns true if this function is an intrinsic and the
   /// intrinsic is specific to a certain target. If this is not an intrinsic
   /// or a generic intrinsic, false is returned.
   bool isTargetIntrinsic() const;
 
   /// Returns true if the function is one of the "Constrained Floating-Point
   /// Intrinsics". Returns false if not, and returns false when
   /// getIntrinsicID() returns Intrinsic::not_intrinsic.

llvm/lib/IR/Function.cpp

 #undef GET_INTRINSIC_NAME_TABLE
 };
 
 /// Table of per-target intrinsic name tables.
 #define GET_INTRINSIC_TARGET_DATA
 #include "llvm/IR/IntrinsicImpl.inc"
 #undef GET_INTRINSIC_TARGET_DATA
 
+bool Function::isTargetIntrinsic(Intrinsic::ID IID) {
+  return IID > TargetInfos[0].Count;
+}
+
 bool Function::isTargetIntrinsic() const {
-  return IntID > TargetInfos[0].Count;
+  return isTargetIntrinsic(IntID);
 }
 
 /// Find the segment of \c IntrinsicNameTable for intrinsics with the same
 /// target as \c Name, or the generic table if \c Name is not target specific.
 ///
 /// Returns the relevant slice of \c IntrinsicNameTable
 static ArrayRef<const char *> findTargetSubtable(StringRef Name) {
   assert(Name.startswith("llvm."));

llvm/test/Analysis/CostModel/ARM/target-intrinsics.ll

 ; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py
 ; RUN: opt < %s -cost-model -analyze -mtriple=thumbv8.1m.main -mattr=+mve | FileCheck %s --check-prefix=CHECK-THUMB2-RECIP
 ; RUN: opt < %s -cost-model -analyze -cost-kind=throughput -mtriple=thumbv8.1m.main -mattr=+mve | FileCheck %s --check-prefix=CHECK-THUMB2-RECIP
 ; RUN: opt < %s -cost-model -analyze -cost-kind=latency -mtriple=thumbv8.1m.main -mattr=+mve | FileCheck %s --check-prefix=CHECK-THUMB2-LAT
 ; RUN: opt < %s -cost-model -analyze -cost-kind=code-size -mtriple=thumbv8.1m.main -mattr=+mve | FileCheck %s --check-prefix=CHECK-THUMB2-SIZE
 
 define void @intrinsics() {
 ; CHECK-THUMB2-RECIP-LABEL: 'intrinsics'
 ; CHECK-THUMB2-RECIP-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %t1 = call i32 @llvm.arm.ssat(i32 undef, i32 undef)
 ; CHECK-THUMB2-RECIP-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %t2 = tail call { <8 x half>, <8 x half> } @llvm.arm.mve.vld2q.v8f16.p0f16(half* undef)
 ; CHECK-THUMB2-RECIP-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %t3 = call { i32, i32 } @llvm.arm.mve.sqrshrl(i32 undef, i32 undef, i32 undef, i32 48)
-; CHECK-THUMB2-RECIP-NEXT:  Cost Model: Found an estimated cost of 135 for instruction: %t4 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 0, i32 0, i32 0, i32 0, <8 x i16> undef, <8 x i16> undef)
+; CHECK-THUMB2-RECIP-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %t4 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 0, i32 0, i32 0, i32 0, <8 x i16> undef, <8 x i16> undef)
 ; CHECK-THUMB2-RECIP-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
 ; CHECK-THUMB2-LAT-LABEL: 'intrinsics'
 ; CHECK-THUMB2-LAT-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %t1 = call i32 @llvm.arm.ssat(i32 undef, i32 undef)
 ; CHECK-THUMB2-LAT-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %t2 = tail call { <8 x half>, <8 x half> } @llvm.arm.mve.vld2q.v8f16.p0f16(half* undef)
 ; CHECK-THUMB2-LAT-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %t3 = call { i32, i32 } @llvm.arm.mve.sqrshrl(i32 undef, i32 undef, i32 undef, i32 48)
 ; CHECK-THUMB2-LAT-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %t4 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 0, i32 0, i32 0, i32 0, <8 x i16> undef, <8 x i16> undef)
 ; CHECK-THUMB2-LAT-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void