[llvm] r177105 - ARM cost model: Increase cost of some vector selects we do terrible on
Arnold Schwaighofer
aschwaighofer at apple.com
Fri Mar 15 09:29:13 PDT 2013
Hi Nadav,
On Mar 15, 2013, at 10:59 AM, Nadav Rotem <nrotem at apple.com> wrote:
> Do you know why the target independent logic does not catch these cases ? Is 'select' lowered using custom code in the ARM backend ?
No, vector selects are marked as expand (scalar selects are marked as custom) but the generic logic is:
unsigned BasicTTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
Type *CondTy) const {
…
// Otherwise, assume that the cast is scalarized.
if (ValTy->isVectorTy()) {
unsigned Num = ValTy->getVectorNumElements();
if (CondTy)
CondTy = CondTy->getScalarType();
unsigned Cost = TopTTI->getCmpSelInstrCost(Opcode, ValTy->getScalarType(),
CondTy);
// Return the cost of multiple scalar invocation plus the cost of inserting
// and extracting the values.
return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
}
this does not reflect the cost of going to the stack for the condition parameter (CondTy) conversion.
Thanks
> On Mar 14, 2013, at 12:17 PM, Arnold Schwaighofer <aschwaighofer at apple.com> wrote:
>
>> Author: arnolds
>> Date: Thu Mar 14 14:17:02 2013
>> New Revision: 177105
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=177105&view=rev
>> Log:
>> ARM cost model: Increase cost of some vector selects we do terrible on
>>
>> By terrible I mean we store/load from the stack.
>>
>> This matters on PAQp8 in _Z5trainPsS_ii (which is inlined into Mixer::update)
>> where we decide to vectorize a loop with a VF of 8 resulting in a 25%
>> degradation on a cortex-a8.
>>
>> LV: Found an estimated cost of 2 for VF 8 For instruction: icmp slt i32
>> LV: Found an estimated cost of 2 for VF 8 For instruction: select i1, i32, i32
>>
>> The bug that tracks the CodeGen part is PR14868.
>>
>> radar://13403975
>>
>> Modified:
>> llvm/trunk/lib/Target/ARM/ARMTargetTransformInfo.cpp
>> llvm/trunk/test/Analysis/CostModel/ARM/select.ll
>> llvm/trunk/test/CodeGen/ARM/vselect_imax.ll
>>
>> Modified: llvm/trunk/lib/Target/ARM/ARMTargetTransformInfo.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/ARM/ARMTargetTransformInfo.cpp?rev=177105&r1=177104&r2=177105&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/Target/ARM/ARMTargetTransformInfo.cpp (original)
>> +++ llvm/trunk/lib/Target/ARM/ARMTargetTransformInfo.cpp Thu Mar 14 14:17:02 2013
>> @@ -334,6 +334,30 @@ unsigned ARMTTI::getCmpSelInstrCost(unsi
>> int ISD = TLI->InstructionOpcodeToISD(Opcode);
>> // On NEON a a vector select gets lowered to vbsl.
>> if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT) {
>> + // Lowering of some vector selects is currently far from perfect.
>> + static const TypeConversionCostTblEntry<MVT> NEONVectorSelectTbl[] = {
>> + { ISD::SELECT, MVT::v4i1, MVT::v4i8, 2*4 + 2*1 },
>> + { ISD::SELECT, MVT::v8i1, MVT::v8i8, 2*8 + 1 },
>> + { ISD::SELECT, MVT::v16i1, MVT::v16i8, 2*16 + 1 },
>> + { ISD::SELECT, MVT::v4i1, MVT::v4i16, 2*4 + 1 },
>> + { ISD::SELECT, MVT::v8i1, MVT::v8i16, 2*8 + 1 },
>> + { ISD::SELECT, MVT::v16i1, MVT::v16i16, 2*16 + 1 + 3*1 + 4*1 },
>> + { ISD::SELECT, MVT::v8i1, MVT::v8i32, 4*8 + 1*3 + 1*4 + 1*2 },
>> + { ISD::SELECT, MVT::v16i1, MVT::v16i32, 4*16 + 1*6 + 1*8 + 1*4 },
>> + { ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 },
>> + { ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 },
>> + { ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 }
>> + };
>> +
>> + EVT SelCondTy = TLI->getValueType(CondTy);
>> + EVT SelValTy = TLI->getValueType(ValTy);
>> + int Idx = ConvertCostTableLookup<MVT>(NEONVectorSelectTbl,
>> + array_lengthof(NEONVectorSelectTbl),
>> + ISD, SelCondTy.getSimpleVT(),
>> + SelValTy.getSimpleVT());
>> + if (Idx != -1)
>> + return NEONVectorSelectTbl[Idx].Cost;
>> +
>> std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
>> return LT.first;
>> }
>>
>> Modified: llvm/trunk/test/Analysis/CostModel/ARM/select.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Analysis/CostModel/ARM/select.ll?rev=177105&r1=177104&r2=177105&view=diff
>> ==============================================================================
>> --- llvm/trunk/test/Analysis/CostModel/ARM/select.ll (original)
>> +++ llvm/trunk/test/Analysis/CostModel/ARM/select.ll Thu Mar 14 14:17:02 2013
>> @@ -21,26 +21,39 @@ define void @casts() {
>> ; Vector values
>> ; CHECK: cost of 1 {{.*}} select
>> %v7 = select <2 x i1> undef, <2 x i8> undef, <2 x i8> undef
>> - ; CHECK: cost of 1 {{.*}} select
>> + ; CHECK: cost of 10 {{.*}} select
>> %v8 = select <4 x i1> undef, <4 x i8> undef, <4 x i8> undef
>> - ; CHECK: cost of 1 {{.*}} select
>> + ; CHECK: cost of 17 {{.*}} select
>> %v9 = select <8 x i1> undef, <8 x i8> undef, <8 x i8> undef
>> - ; CHECK: cost of 1 {{.*}} select
>> + ; CHECK: cost of 33 {{.*}} select
>> %v10 = select <16 x i1> undef, <16 x i8> undef, <16 x i8> undef
>>
>> ; CHECK: cost of 1 {{.*}} select
>> %v11 = select <2 x i1> undef, <2 x i16> undef, <2 x i16> undef
>> - ; CHECK: cost of 1 {{.*}} select
>> + ; CHECK: cost of 9 {{.*}} select
>> %v12 = select <4 x i1> undef, <4 x i16> undef, <4 x i16> undef
>> - ; CHECK: cost of 1 {{.*}} select
>> + ; CHECK: cost of 17 {{.*}} select
>> %v13 = select <8 x i1> undef, <8 x i16> undef, <8 x i16> undef
>> + ; CHECK: cost of 40 {{.*}} select
>> + %v13b = select <16 x i1> undef, <16 x i16> undef, <16 x i16> undef
>>
>> ; CHECK: cost of 1 {{.*}} select
>> %v14 = select <2 x i1> undef, <2 x i32> undef, <2 x i32> undef
>> ; CHECK: cost of 1 {{.*}} select
>> %v15 = select <4 x i1> undef, <4 x i32> undef, <4 x i32> undef
>> + ; CHECK: cost of 41 {{.*}} select
>> + %v15b = select <8 x i1> undef, <8 x i32> undef, <8 x i32> undef
>> + ; CHECK: cost of 82 {{.*}} select
>> + %v15c = select <16 x i1> undef, <16 x i32> undef, <16 x i32> undef
>> +
>> ; CHECK: cost of 1 {{.*}} select
>> %v16 = select <2 x i1> undef, <2 x i64> undef, <2 x i64> undef
>> + ; CHECK: cost of 19 {{.*}} select
>> + %v16a = select <4 x i1> undef, <4 x i64> undef, <4 x i64> undef
>> + ; CHECK: cost of 50 {{.*}} select
>> + %v16b = select <8 x i1> undef, <8 x i64> undef, <8 x i64> undef
>> + ; CHECK: cost of 100 {{.*}} select
>> + %v16c = select <16 x i1> undef, <16 x i64> undef, <16 x i64> undef
>>
>> ; CHECK: cost of 1 {{.*}} select
>> %v17 = select <2 x i1> undef, <2 x float> undef, <2 x float> undef
>>
>> Modified: llvm/trunk/test/CodeGen/ARM/vselect_imax.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/ARM/vselect_imax.ll?rev=177105&r1=177104&r2=177105&view=diff
>> ==============================================================================
>> --- llvm/trunk/test/CodeGen/ARM/vselect_imax.ll (original)
>> +++ llvm/trunk/test/CodeGen/ARM/vselect_imax.ll Thu Mar 14 14:17:02 2013
>> @@ -10,3 +10,210 @@ define void @vmax_v4i32(<4 x i32>* %m, <
>> ret void
>> }
>>
>> +; We adjusted the cost model of the following selects. When we improve code
>> +; lowering we also need to adjust the cost.
>> +; RUN: opt < %s -cost-model -analyze -mtriple=thumbv7-apple-ios6.0.0 -march=arm -mcpu=cortex-a8 | FileCheck %s --check-prefix=COST
>> +%T0_3 = type <4 x i8>
>> +%T1_3 = type <4 x i1>
>> +; CHECK: func_blend3:
>> +define void @func_blend3(%T0_3* %loadaddr, %T0_3* %loadaddr2,
>> + %T1_3* %blend, %T0_3* %storeaddr) {
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: vldr
>> + %v0 = load %T0_3* %loadaddr
>> + %v1 = load %T0_3* %loadaddr2
>> + %c = load %T1_3* %blend
>> +; COST: func_blend3
>> +; COST: cost of 10 {{.*}} select
>> + %r = select %T1_3 %c, %T0_3 %v0, %T0_3 %v1
>> + store %T0_3 %r, %T0_3* %storeaddr
>> + ret void
>> +}
>> +%T0_4 = type <8 x i8>
>> +%T1_4 = type <8 x i1>
>> +; CHECK: func_blend4:
>> +define void @func_blend4(%T0_4* %loadaddr, %T0_4* %loadaddr2,
>> + %T1_4* %blend, %T0_4* %storeaddr) {
>> + %v0 = load %T0_4* %loadaddr
>> + %v1 = load %T0_4* %loadaddr2
>> + %c = load %T1_4* %blend
>> +; check: strb
>> +; check: strb
>> +; check: strb
>> +; check: strb
>> +; check: vldr
>> +; COST: func_blend4
>> +; COST: cost of 17 {{.*}} select
>> + %r = select %T1_4 %c, %T0_4 %v0, %T0_4 %v1
>> + store %T0_4 %r, %T0_4* %storeaddr
>> + ret void
>> +}
>> +%T0_5 = type <16 x i8>
>> +%T1_5 = type <16 x i1>
>> +; CHECK: func_blend5:
>> +define void @func_blend5(%T0_5* %loadaddr, %T0_5* %loadaddr2,
>> + %T1_5* %blend, %T0_5* %storeaddr) {
>> + %v0 = load %T0_5* %loadaddr
>> + %v1 = load %T0_5* %loadaddr2
>> + %c = load %T1_5* %blend
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: vld
>> +; COST: func_blend5
>> +; COST: cost of 33 {{.*}} select
>> + %r = select %T1_5 %c, %T0_5 %v0, %T0_5 %v1
>> + store %T0_5 %r, %T0_5* %storeaddr
>> + ret void
>> +}
>> +%T0_8 = type <4 x i16>
>> +%T1_8 = type <4 x i1>
>> +; CHECK: func_blend8:
>> +define void @func_blend8(%T0_8* %loadaddr, %T0_8* %loadaddr2,
>> + %T1_8* %blend, %T0_8* %storeaddr) {
>> + %v0 = load %T0_8* %loadaddr
>> + %v1 = load %T0_8* %loadaddr2
>> + %c = load %T1_8* %blend
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: vld
>> +; COST: func_blend8
>> +; COST: cost of 9 {{.*}} select
>> + %r = select %T1_8 %c, %T0_8 %v0, %T0_8 %v1
>> + store %T0_8 %r, %T0_8* %storeaddr
>> + ret void
>> +}
>> +%T0_9 = type <8 x i16>
>> +%T1_9 = type <8 x i1>
>> +; CHECK: func_blend9:
>> +define void @func_blend9(%T0_9* %loadaddr, %T0_9* %loadaddr2,
>> + %T1_9* %blend, %T0_9* %storeaddr) {
>> + %v0 = load %T0_9* %loadaddr
>> + %v1 = load %T0_9* %loadaddr2
>> + %c = load %T1_9* %blend
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: vld
>> +; COST: func_blend9
>> +; COST: cost of 17 {{.*}} select
>> + %r = select %T1_9 %c, %T0_9 %v0, %T0_9 %v1
>> + store %T0_9 %r, %T0_9* %storeaddr
>> + ret void
>> +}
>> +%T0_10 = type <16 x i16>
>> +%T1_10 = type <16 x i1>
>> +; CHECK: func_blend10:
>> +define void @func_blend10(%T0_10* %loadaddr, %T0_10* %loadaddr2,
>> + %T1_10* %blend, %T0_10* %storeaddr) {
>> + %v0 = load %T0_10* %loadaddr
>> + %v1 = load %T0_10* %loadaddr2
>> + %c = load %T1_10* %blend
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: vld
>> +; COST: func_blend10
>> +; COST: cost of 40 {{.*}} select
>> + %r = select %T1_10 %c, %T0_10 %v0, %T0_10 %v1
>> + store %T0_10 %r, %T0_10* %storeaddr
>> + ret void
>> +}
>> +%T0_14 = type <8 x i32>
>> +%T1_14 = type <8 x i1>
>> +; CHECK: func_blend14:
>> +define void @func_blend14(%T0_14* %loadaddr, %T0_14* %loadaddr2,
>> + %T1_14* %blend, %T0_14* %storeaddr) {
>> + %v0 = load %T0_14* %loadaddr
>> + %v1 = load %T0_14* %loadaddr2
>> + %c = load %T1_14* %blend
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; COST: func_blend14
>> +; COST: cost of 41 {{.*}} select
>> + %r = select %T1_14 %c, %T0_14 %v0, %T0_14 %v1
>> + store %T0_14 %r, %T0_14* %storeaddr
>> + ret void
>> +}
>> +%T0_15 = type <16 x i32>
>> +%T1_15 = type <16 x i1>
>> +; CHECK: func_blend15:
>> +define void @func_blend15(%T0_15* %loadaddr, %T0_15* %loadaddr2,
>> + %T1_15* %blend, %T0_15* %storeaddr) {
>> + %v0 = load %T0_15* %loadaddr
>> + %v1 = load %T0_15* %loadaddr2
>> + %c = load %T1_15* %blend
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; COST: func_blend15
>> +; COST: cost of 82 {{.*}} select
>> + %r = select %T1_15 %c, %T0_15 %v0, %T0_15 %v1
>> + store %T0_15 %r, %T0_15* %storeaddr
>> + ret void
>> +}
>> +%T0_18 = type <4 x i64>
>> +%T1_18 = type <4 x i1>
>> +; CHECK: func_blend18:
>> +define void @func_blend18(%T0_18* %loadaddr, %T0_18* %loadaddr2,
>> + %T1_18* %blend, %T0_18* %storeaddr) {
>> + %v0 = load %T0_18* %loadaddr
>> + %v1 = load %T0_18* %loadaddr2
>> + %c = load %T1_18* %blend
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: strh
>> +; CHECK: strh
>> +; COST: func_blend18
>> +; COST: cost of 19 {{.*}} select
>> + %r = select %T1_18 %c, %T0_18 %v0, %T0_18 %v1
>> + store %T0_18 %r, %T0_18* %storeaddr
>> + ret void
>> +}
>> +%T0_19 = type <8 x i64>
>> +%T1_19 = type <8 x i1>
>> +; CHECK: func_blend19:
>> +define void @func_blend19(%T0_19* %loadaddr, %T0_19* %loadaddr2,
>> + %T1_19* %blend, %T0_19* %storeaddr) {
>> + %v0 = load %T0_19* %loadaddr
>> + %v1 = load %T0_19* %loadaddr2
>> + %c = load %T1_19* %blend
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; COST: func_blend19
>> +; COST: cost of 50 {{.*}} select
>> + %r = select %T1_19 %c, %T0_19 %v0, %T0_19 %v1
>> + store %T0_19 %r, %T0_19* %storeaddr
>> + ret void
>> +}
>> +%T0_20 = type <16 x i64>
>> +%T1_20 = type <16 x i1>
>> +; CHECK: func_blend20:
>> +define void @func_blend20(%T0_20* %loadaddr, %T0_20* %loadaddr2,
>> + %T1_20* %blend, %T0_20* %storeaddr) {
>> + %v0 = load %T0_20* %loadaddr
>> + %v1 = load %T0_20* %loadaddr2
>> + %c = load %T1_20* %blend
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; CHECK: strb
>> +; COST: func_blend20
>> +; COST: cost of 100 {{.*}} select
>> + %r = select %T1_20 %c, %T0_20 %v0, %T0_20 %v1
>> + store %T0_20 %r, %T0_20* %storeaddr
>> + ret void
>> +}
>>
>>
>> _______________________________________________
>> 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