Cost model support for lowered math builtins.

We make the cost for calling libm functions extremely high as emitting the
calls is expensive and causes spills (on x86) so performance suffers. We still
vectorize important calls like ceilf and friends on SSE4.1. and fabs.

Differential Revision: http://llvm-reviews.chandlerc.com/D466

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@176287 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Analysis/CostModel.cpp b/lib/Analysis/CostModel.cpp
index 44684a9685bba119dc2088f0c67eac2009e0336f..df70d157b41dec451c7b4f74e87a2c1396487df9 100644 (file)
--- a/lib/Analysis/CostModel.cpp
+++ b/lib/Analysis/CostModel.cpp
@@ -23,6 +23,7 @@
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/Value.h"
 #include "llvm/Pass.h"
 #include "llvm/Support/Debug.h"
@@ -189,6 +190,16 @@ unsigned CostModelAnalysis::getInstructionCost(const Instruction *I) const {
                                  0);
     return -1;
   }
+  case Instruction::Call:
+    if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
+      SmallVector<Type*, 4> Tys;
+      for (unsigned J = 0, JE = II->getNumArgOperands(); J != JE; ++J)
+        Tys.push_back(II->getArgOperand(J)->getType());
+
+      return TTI->getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(),
+                                        Tys);
+    }
+    return -1;
   default:
     // We don't have any information on this instruction.
     return -1;

diff --git a/lib/CodeGen/BasicTargetTransformInfo.cpp b/lib/CodeGen/BasicTargetTransformInfo.cpp
index e8b5b4fe8d1b9e709befe8835870a8001b87befa..4cd1b80dc0b85833c2f40335c1e12f26d29e446c 100644 (file)
--- a/lib/CodeGen/BasicTargetTransformInfo.cpp
+++ b/lib/CodeGen/BasicTargetTransformInfo.cpp
@@ -379,22 +379,77 @@ unsigned BasicTTI::getMemoryOpCost(unsigned Opcode, Type *Src,
   return LT.first;
 }
 
-unsigned BasicTTI::getIntrinsicInstrCost(Intrinsic::ID, Type *RetTy,
+unsigned BasicTTI::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
                                          ArrayRef<Type *> Tys) const {
-  // assume that we need to scalarize this intrinsic.
-  unsigned ScalarizationCost = 0;
-  unsigned ScalarCalls = 1;
-  if (RetTy->isVectorTy()) {
-    ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
-    ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
-  }
-  for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
-    if (Tys[i]->isVectorTy()) {
-      ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
+  unsigned ISD = 0;
+  switch (IID) {
+  default: {
+    // Assume that we need to scalarize this intrinsic.
+    unsigned ScalarizationCost = 0;
+    unsigned ScalarCalls = 1;
+    if (RetTy->isVectorTy()) {
+      ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
       ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
     }
+    for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
+      if (Tys[i]->isVectorTy()) {
+        ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
+        ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
+      }
+    }
+
+    return ScalarCalls + ScalarizationCost;
+  }
+  // Look for intrinsics that can be lowered directly or turned into a scalar
+  // intrinsic call.
+  case Intrinsic::sqrt:    ISD = ISD::FSQRT;  break;
+  case Intrinsic::sin:     ISD = ISD::FSIN;   break;
+  case Intrinsic::cos:     ISD = ISD::FCOS;   break;
+  case Intrinsic::exp:     ISD = ISD::FEXP;   break;
+  case Intrinsic::exp2:    ISD = ISD::FEXP2;  break;
+  case Intrinsic::log:     ISD = ISD::FLOG;   break;
+  case Intrinsic::log10:   ISD = ISD::FLOG10; break;
+  case Intrinsic::log2:    ISD = ISD::FLOG2;  break;
+  case Intrinsic::fabs:    ISD = ISD::FABS;   break;
+  case Intrinsic::floor:   ISD = ISD::FFLOOR; break;
+  case Intrinsic::ceil:    ISD = ISD::FCEIL;  break;
+  case Intrinsic::trunc:   ISD = ISD::FTRUNC; break;
+  case Intrinsic::rint:    ISD = ISD::FRINT;  break;
+  case Intrinsic::pow:     ISD = ISD::FPOW;   break;
+  case Intrinsic::fma:     ISD = ISD::FMA;    break;
+  case Intrinsic::fmuladd: ISD = ISD::FMA;    break; // FIXME: mul + add?
+  }
+
+  std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(RetTy);
+
+  if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
+    // The operation is legal. Assume it costs 1.
+    // If the type is split to multiple registers, assume that thre is some
+    // overhead to this.
+    // TODO: Once we have extract/insert subvector cost we need to use them.
+    if (LT.first > 1)
+      return LT.first * 2;
+    return LT.first * 1;
   }
-  return ScalarCalls + ScalarizationCost;
+
+  if (!TLI->isOperationExpand(ISD, LT.second)) {
+    // If the operation is custom lowered then assume
+    // thare the code is twice as expensive.
+    return LT.first * 2;
+  }
+
+  // Else, assume that we need to scalarize this intrinsic. For math builtins
+  // this will emit a costly libcall, adding call overhead and spills. Make it
+  // very expensive.
+  if (RetTy->isVectorTy()) {
+    unsigned Num = RetTy->getVectorNumElements();
+    unsigned Cost = TopTTI->getIntrinsicInstrCost(IID, RetTy->getScalarType(),
+                                                  Tys);
+    return 10 * Cost * Num;
+  }
+
+  // This is going to be turned into a library call, make it expensive.
+  return 10;
 }
 
 unsigned BasicTTI::getNumberOfParts(Type *Tp) const {

diff --git a/test/Analysis/CostModel/X86/intrinsic-cost.ll b/test/Analysis/CostModel/X86/intrinsic-cost.ll
new file mode 100644 (file)
index 0000000..e235a36
--- /dev/null
+++ b/test/Analysis/CostModel/X86/intrinsic-cost.ll
@@ -0,0 +1,32 @@
+; RUN: opt -S -mtriple=x86_64-apple-darwin -mcpu=core2 -cost-model -analyze < %s | FileCheck %s -check-prefix=CORE2
+; RUN: opt -S -mtriple=x86_64-apple-darwin -mcpu=corei7 -cost-model -analyze < %s | FileCheck %s -check-prefix=COREI7
+
+; If SSE4.1 roundps instruction is available it is cheap to lower, otherwise
+; it'll be scalarized into calls which are expensive.
+define void @test1(float* nocapture %f) nounwind {
+vector.ph:
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %0 = getelementptr inbounds float* %f, i64 %index
+  %1 = bitcast float* %0 to <4 x float>*
+  %wide.load = load <4 x float>* %1, align 4
+  %2 = call <4 x float> @llvm.ceil.v4f32(<4 x float> %wide.load)
+  store <4 x float> %2, <4 x float>* %1, align 4
+  %index.next = add i64 %index, 4
+  %3 = icmp eq i64 %index.next, 1024
+  br i1 %3, label %for.end, label %vector.body
+
+for.end:                                          ; preds = %vector.body
+  ret void
+
+; CORE2: Printing analysis 'Cost Model Analysis' for function 'test1':
+; CORE2: Cost Model: Found an estimated cost of 400 for instruction:   %2 = call <4 x float> @llvm.ceil.v4f32(<4 x float> %wide.load)
+
+; COREI7: Printing analysis 'Cost Model Analysis' for function 'test1':
+; COREI7: Cost Model: Found an estimated cost of 1 for instruction:   %2 = call <4 x float> @llvm.ceil.v4f32(<4 x float> %wide.load)
+
+}
+
+declare <4 x float> @llvm.ceil.v4f32(<4 x float>)  nounwind readnone