[TargetTransformInfo] Handle intrinsic call in getInstructionLatency()

[Carrot]

Summary

Usually an intrinsic is a simple target instruction, it should have a small latency. A real function call has much larger latency. So handle the intrinsic call in function getInstructionLatency().

[hfinkel]

include/llvm/Analysis/TargetTransformInfoImpl.h

792 ↗

(On Diff #116096)

I think that you want to do the same thing here that we do in getCallCost (call isLoweredToCall), perhaps like this:

if (!static_cast<T *>(this)->isLoweredToCall(F))
      return 1;

[hfinkel]

include/llvm/Analysis/TargetTransformInfoImpl.h

793 ↗

(On Diff #116263)

Actually, do we want to just return 1 here, or do we want to fall through to the logic below (which returns 3 for FP types)?

[Carrot]

include/llvm/Analysis/TargetTransformInfoImpl.h

793 ↗

(On Diff #116263)

Good catch! It will give more precise result.

[hfinkel]

LGTM

include/llvm/Analysis/TargetTransformInfoImpl.h

791 ↗

(On Diff #116356)

This variable name should start with a capital letter (per our coding conventions).

[Carrot]

Will commit this patch.

llvm/trunk/include/llvm/Analysis/TargetTransformInfoImpl.h

public:
   }
 
   int getInstructionLatency(const Instruction *I) {
     SmallVector<const Value *, 4> Operands(I->value_op_begin(),
                                            I->value_op_end());
     if (getUserCost(I, Operands) == TTI::TCC_Free)
       return 0;
 
-    if (isa<CallInst>(I))
-      return 40;
-
     if (isa<LoadInst>(I))
       return 4;
 
-    Type *dstTy = I->getType();
-    if (VectorType *VectorTy = dyn_cast<VectorType>(dstTy))
-      dstTy = VectorTy->getElementType();
-    if (dstTy->isFloatingPointTy())
+    Type *DstTy = I->getType();
+
+    // Usually an intrinsic is a simple instruction.
+    // A real function call is much slower.
+    if (auto *CI = dyn_cast<CallInst>(I)) {
+      const Function *F = CI->getCalledFunction();
+      if (static_cast<T *>(this)->isLoweredToCall(F))
+        return 40;
+      // Some intrinsics return a value and a flag, we use the value type
+      // to decide its latency.
+      if (StructType* StructTy = dyn_cast<StructType>(DstTy))
+        DstTy = StructTy->getElementType(0);
+      // Fall through to simple instructions.
+    }
+
+    if (VectorType *VectorTy = dyn_cast<VectorType>(DstTy))
+      DstTy = VectorTy->getElementType();
+    if (DstTy->isFloatingPointTy())
       return 3;
 
     return 1;
   }
 };
 }
 
 #endif

llvm/trunk/test/Analysis/CostModel/X86/costmodel.ll

 ; RUN: opt < %s -cost-model -cost-kind=latency -analyze -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 | FileCheck %s --check-prefix=LATENCY
 ; RUN: opt < %s -cost-model -cost-kind=code-size -analyze -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 | FileCheck %s --check-prefix=CODESIZE
 
 ; Tests if the interface TargetTransformInfo::getInstructionCost() works correctly.
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 
+declare { i32, i1 } @llvm.uadd.with.overflow.i32(i32, i32)
+
 define i64 @foo(i64 %arg) {
 
   ; LATENCY:  cost of 0 {{.*}} alloca i32
   ; CODESIZE: cost of 0 {{.*}} alloca i32
   %A1 = alloca i32, align 8
 
   ; LATENCY:  cost of 1 {{.*}} alloca i64, i64 undef
   ; CODESIZE: cost of 1 {{.*}} alloca i64, i64 undef
define i64 @foo(i64 %arg) {
   ; LATENCY:  cost of 0 {{.*}} ptrtoint
   ; CODESIZE: cost of 0 {{.*}} ptrtoint
   %P2I = ptrtoint i8* undef to i64
 
   ; LATENCY:  cost of 0 {{.*}} trunc
   ; CODESIZE: cost of 0 {{.*}} trunc
   %TC = trunc i64 undef to i32
 
+  ; LATENCY:  cost of 1 {{.*}} call
+  ; CODESIZE: cost of 1 {{.*}} call
+  %uadd = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 undef, i32 undef)
+
   ; LATENCY:  cost of 1 {{.*}} ret
   ; CODESIZE: cost of 1 {{.*}} ret
   ret i64 undef
 }