[PATCH] D90597: [CostModel] Make target intrinsics cheap by default

[Dave Green via Phabricator via llvm-commits]

dmgreen created this revision.
dmgreen added reviewers: spatel, RKSimon, samparker.
Herald added subscribers: dexonsmith, hiraditya.
Herald added a project: LLVM.
dmgreen requested review of this revision.

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 <https://reviews.llvm.org/D90554>, which would also end up changing the size costs which then has an effect on loop unrolling and unswitch.


https://reviews.llvm.org/D90597

Files:
  llvm/include/llvm/CodeGen/BasicTTIImpl.h
  llvm/include/llvm/IR/Function.h
  llvm/lib/IR/Function.cpp
  llvm/test/Analysis/CostModel/ARM/target-intrinsics.ll


Index: llvm/test/Analysis/CostModel/ARM/target-intrinsics.ll
===================================================================
--- llvm/test/Analysis/CostModel/ARM/target-intrinsics.ll
+++ llvm/test/Analysis/CostModel/ARM/target-intrinsics.ll
@@ -9,7 +9,7 @@
 ; 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'
Index: llvm/lib/IR/Function.cpp
===================================================================
--- llvm/lib/IR/Function.cpp
+++ llvm/lib/IR/Function.cpp
@@ -640,8 +640,12 @@
 #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
Index: llvm/include/llvm/IR/Function.h
===================================================================
--- llvm/include/llvm/IR/Function.h
+++ llvm/include/llvm/IR/Function.h
@@ -199,6 +199,10 @@
   /// 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.
Index: llvm/include/llvm/CodeGen/BasicTTIImpl.h
===================================================================
--- llvm/include/llvm/CodeGen/BasicTTIImpl.h
+++ llvm/include/llvm/CodeGen/BasicTTIImpl.h
@@ -1290,6 +1290,10 @@
     unsigned ScalarizationCostPassed = ICA.getScalarizationCost();
     bool SkipScalarizationCost = ICA.skipScalarizationCost();
 
+    // Assume that target intrinsics are cheap
+    if (Function::isTargetIntrinsic(IID))
+      return TargetTransformInfo::TCC_Basic;
+
     VectorType *VecOpTy = nullptr;
     if (!Tys.empty()) {
       // The vector reduction operand is operand 0 except for fadd/fmul.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D90597.302247.patch
Type: text/x-patch
Size: 3229 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201102/cc8efdef/attachment.bin>