Skip to content

Commit

Permalink
[TargetTransformInfo] Handle intrinsic call in getInstructionLatency()
Browse files Browse the repository at this point in the history
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().

Differential Revision: https://reviews.llvm.org/D38104

llvm-svn: 314003
  • Loading branch information
weiguozhi committed Sep 22, 2017
1 parent 0acf343 commit bce228c
Showing 2 changed files with 24 additions and 7 deletions.
25 changes: 18 additions & 7 deletions llvm/include/llvm/Analysis/TargetTransformInfoImpl.h
Original file line number Diff line number Diff line change
@@ -785,16 +785,27 @@ class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
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;
6 changes: 6 additions & 0 deletions llvm/test/Analysis/CostModel/X86/costmodel.ll
Original file line number Diff line number Diff line change
@@ -5,6 +5,8 @@

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
@@ -39,6 +41,10 @@ define i64 @foo(i64 %arg) {
; 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

0 comments on commit bce228c

Please sign in to comment.