[llvm] r178459 - X86TTI: Add accurate costs for itofp operations, based on the actual instruction counts.
Evan Cheng
evan.cheng at apple.com
Mon Apr 1 20:04:57 PDT 2013
On Apr 1, 2013, at 12:27 PM, Arnold Schwaighofer <aschwaighofer at apple.com> wrote:
> Agreed. But I can't think of an effective way how to do this retroactively/automatically.
>
> The only thing that comes close to this is to co-locate tests that check the cost with tests that check that the lowering (produced assembly) has changed.
>
> We are already doing (should already be doing) this or something that comes close:
>
> When we change lowering of an instruction we (should):
> * update the cost model and add a cost model test
> * add a test case that makes sure the produced assembly is right
>
> The first point is being enforced by an undocumented rule that a change to lowering also requires a change (inspection) of the cost model. Nadav and I are trying to enforce this.
I think that makes sense for now. It would take a lot of infrastructure work in order to generate the cost models programmatically.
Thanks,
Evan
> The second point is part of the normal requirements for a change.
>
> Co-locating them in one file retroactively would mean we would have to do this for every existing test/CodeGen/<ARCH>/<test lowering>.ll test.
>
>
> Best,
> Arnold
>
> On Apr 1, 2013, at 1:29 PM, Evan Cheng <evan.cheng at apple.com> wrote:
>
>> I have a generic question about these tables. If the cost is impacted by how the target / sdisel lowers the node, then the entries can be outdated if lowering changes. It would be nice if we have a mechanism to detect that.
>>
>> Evan
>>
>> On Apr 1, 2013, at 3:23 AM, Benjamin Kramer <benny.kra at googlemail.com> wrote:
>>
>>> Author: d0k
>>> Date: Mon Apr 1 05:23:49 2013
>>> New Revision: 178459
>>>
>>> URL: http://llvm.org/viewvc/llvm-project?rev=178459&view=rev
>>> Log:
>>> X86TTI: Add accurate costs for itofp operations, based on the actual instruction counts.
>>>
>>> Modified:
>>> llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp
>>> llvm/trunk/test/Analysis/CostModel/X86/cast.ll
>>> llvm/trunk/test/Transforms/LoopVectorize/X86/conversion-cost.ll
>>>
>>> Modified: llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp?rev=178459&r1=178458&r2=178459&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp (original)
>>> +++ llvm/trunk/lib/Target/X86/X86TargetTransformInfo.cpp Mon Apr 1 05:23:49 2013
>>> @@ -271,10 +271,33 @@ unsigned X86TTI::getCastInstrCost(unsign
>>> { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
>>> { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 },
>>> { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 1 },
>>> - { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 1 },
>>> - { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 1 },
>>> - { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 1 },
>>> - { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 1 },
>>> +
>>> + { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 },
>>> + { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 8 },
>>> + { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
>>> + { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 },
>>> + { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
>>> + { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
>>> + { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 3 },
>>> + { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
>>> + { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 },
>>> + { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i8, 3 },
>>> + { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i16, 3 },
>>> + { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 },
>>> +
>>> + { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 },
>>> + { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 5 },
>>> + { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
>>> + { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 9 },
>>> + { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 },
>>> + { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 2 },
>>> + { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
>>> + { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 6 },
>>> + { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 },
>>> + { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 },
>>> + { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 },
>>> + { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 },
>>> +
>>> { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 1 },
>>> { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 },
>>> { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 6 },
>>>
>>> Modified: llvm/trunk/test/Analysis/CostModel/X86/cast.ll
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Analysis/CostModel/X86/cast.ll?rev=178459&r1=178458&r2=178459&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/test/Analysis/CostModel/X86/cast.ll (original)
>>> +++ llvm/trunk/test/Analysis/CostModel/X86/cast.ll Mon Apr 1 05:23:49 2013
>>> @@ -77,3 +77,78 @@ define i32 @masks4(<4 x i1> %in) {
>>> ret i32 undef
>>> }
>>>
>>> +define void @sitofp4(<4 x i1> %a, <4 x i8> %b, <4 x i16> %c, <4 x i32> %d) {
>>> + ; CHECK: cost of 3 {{.*}} sitofp
>>> + %A1 = sitofp <4 x i1> %a to <4 x float>
>>> + ; CHECK: cost of 3 {{.*}} sitofp
>>> + %A2 = sitofp <4 x i1> %a to <4 x double>
>>> +
>>> + ; CHECK: cost of 3 {{.*}} sitofp
>>> + %B1 = sitofp <4 x i8> %b to <4 x float>
>>> + ; CHECK: cost of 3 {{.*}} sitofp
>>> + %B2 = sitofp <4 x i8> %b to <4 x double>
>>> +
>>> + ; CHECK: cost of 3 {{.*}} sitofp
>>> + %C1 = sitofp <4 x i16> %c to <4 x float>
>>> + ; CHECK: cost of 3 {{.*}} sitofp
>>> + %C2 = sitofp <4 x i16> %c to <4 x double>
>>> +
>>> + ; CHECK: cost of 1 {{.*}} sitofp
>>> + %D1 = sitofp <4 x i32> %d to <4 x float>
>>> + ; CHECK: cost of 1 {{.*}} sitofp
>>> + %D2 = sitofp <4 x i32> %d to <4 x double>
>>> + ret void
>>> +}
>>> +
>>> +define void @sitofp8(<8 x i1> %a, <8 x i8> %b, <8 x i16> %c, <8 x i32> %d) {
>>> + ; CHECK: cost of 8 {{.*}} sitofp
>>> + %A1 = sitofp <8 x i1> %a to <8 x float>
>>> +
>>> + ; CHECK: cost of 8 {{.*}} sitofp
>>> + %B1 = sitofp <8 x i8> %b to <8 x float>
>>> +
>>> + ; CHECK: cost of 5 {{.*}} sitofp
>>> + %C1 = sitofp <8 x i16> %c to <8 x float>
>>> +
>>> + ; CHECK: cost of 1 {{.*}} sitofp
>>> + %D1 = sitofp <8 x i32> %d to <8 x float>
>>> + ret void
>>> +}
>>> +
>>> +define void @uitofp4(<4 x i1> %a, <4 x i8> %b, <4 x i16> %c, <4 x i32> %d) {
>>> + ; CHECK: cost of 7 {{.*}} uitofp
>>> + %A1 = uitofp <4 x i1> %a to <4 x float>
>>> + ; CHECK: cost of 7 {{.*}} uitofp
>>> + %A2 = uitofp <4 x i1> %a to <4 x double>
>>> +
>>> + ; CHECK: cost of 2 {{.*}} uitofp
>>> + %B1 = uitofp <4 x i8> %b to <4 x float>
>>> + ; CHECK: cost of 2 {{.*}} uitofp
>>> + %B2 = uitofp <4 x i8> %b to <4 x double>
>>> +
>>> + ; CHECK: cost of 2 {{.*}} uitofp
>>> + %C1 = uitofp <4 x i16> %c to <4 x float>
>>> + ; CHECK: cost of 2 {{.*}} uitofp
>>> + %C2 = uitofp <4 x i16> %c to <4 x double>
>>> +
>>> + ; CHECK: cost of 6 {{.*}} uitofp
>>> + %D1 = uitofp <4 x i32> %d to <4 x float>
>>> + ; CHECK: cost of 6 {{.*}} uitofp
>>> + %D2 = uitofp <4 x i32> %d to <4 x double>
>>> + ret void
>>> +}
>>> +
>>> +define void @uitofp8(<8 x i1> %a, <8 x i8> %b, <8 x i16> %c, <8 x i32> %d) {
>>> + ; CHECK: cost of 6 {{.*}} uitofp
>>> + %A1 = uitofp <8 x i1> %a to <8 x float>
>>> +
>>> + ; CHECK: cost of 5 {{.*}} uitofp
>>> + %B1 = uitofp <8 x i8> %b to <8 x float>
>>> +
>>> + ; CHECK: cost of 5 {{.*}} uitofp
>>> + %C1 = uitofp <8 x i16> %c to <8 x float>
>>> +
>>> + ; CHECK: cost of 9 {{.*}} uitofp
>>> + %D1 = uitofp <8 x i32> %d to <8 x float>
>>> + ret void
>>> +}
>>>
>>> Modified: llvm/trunk/test/Transforms/LoopVectorize/X86/conversion-cost.ll
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LoopVectorize/X86/conversion-cost.ll?rev=178459&r1=178458&r2=178459&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/test/Transforms/LoopVectorize/X86/conversion-cost.ll (original)
>>> +++ llvm/trunk/test/Transforms/LoopVectorize/X86/conversion-cost.ll Mon Apr 1 05:23:49 2013
>>> @@ -33,11 +33,10 @@ define i32 @conversion_cost2(i32 %n, i8*
>>>
>>> .lr.ph: ; preds = %0, %.lr.ph
>>> %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 9, %0 ]
>>> - %2 = add nsw i64 %indvars.iv, 3
>>> - %3 = trunc i64 %2 to i32
>>> - %4 = sitofp i32 %3 to float
>>> - %5 = getelementptr inbounds float* %B, i64 %indvars.iv
>>> - store float %4, float* %5, align 4
>>> + %add = add nsw i64 %indvars.iv, 3
>>> + %tofp = sitofp i64 %add to float
>>> + %gep = getelementptr inbounds float* %B, i64 %indvars.iv
>>> + store float %tofp, float* %gep, align 4
>>> %indvars.iv.next = add i64 %indvars.iv, 1
>>> %lftr.wideiv = trunc i64 %indvars.iv.next to i32
>>> %exitcond = icmp eq i32 %lftr.wideiv, %n
>>>
>>>
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
More information about the llvm-commits
mailing list