[ARM] Don't pessimize i32 vselect.

[• chatur01]

Summary

The underlying issues surrounding codegen for 32-bit vselects have been resolved. The pessimistic costs for 64-bit vselects remain due to the bad
scalarization that is still happening there.

I tested this on A57 in T32, A32 and A64 modes. I saw no regressions, and some improvements.

From my benchmarks, I saw these improvements in A57 (T32)
spec.cpu2000.ref.177_mesa 5.95%
lnt.SingleSource/Benchmarks/Shootout/strcat 12.93%
lnt.MultiSource/Benchmarks/MiBench/telecomm-CRC32/telecomm-CRC32 11.89%

I also measured A57 A32 and A9 T32 and found no performance regressions. I see much bigger wins in third-party benchmarks with this change.

[rengolin]

Hi Charlie,

The original commit mentions Cortex-A8's 25% improvement performance in PAQp8. It seems you've only tested in OOO cores, can you just double check on either A53 or A7/A8? It's probably a pattern we now recognise better, but would be good to be on the safe side.

If you find no regressions either, and I'm not expecting any, LGTM. Thanks!

cheers,
--renato

[• chatur01]

Thanks Renato. I double-checked A53 T32 and found no regressions in SPEC 2000, 2006 or LNT.

Adobe-C++/loop_unroll was up 6.49%
ASC_Sequoia/AMGmk/AMGmk was up 5.04%

Several other improvements in LNT.

[• chatur01]

I missed some mechanical changes in test/Analysis/CostModel/ARM/select.ll in this review. I committed the changes along with what's in this revision, which I felt were appropriate for post-commit review if necessary. Sorry about that!

lib/Target/ARM/ARMTargetTransformInfo.cpp

 
 int ARMTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) {
 
   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 NEONVectorSelectTbl[] = {
-      { 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(DL, CondTy);
     EVT SelValTy = TLI->getValueType(DL, ValTy);
     if (SelCondTy.isSimple() && SelValTy.isSimple()) {

test/CodeGen/ARM/vselect_imax.ll

 ; RUN: opt < %s  -cost-model -analyze -mtriple=thumbv7-apple-ios6.0.0 -march=arm -mcpu=cortex-a8 | FileCheck %s --check-prefix=COST
 ; RUN: llc -mtriple=arm-eabi -mattr=+neon %s -o - | FileCheck %s
 ; Make sure that ARM backend with NEON handles vselect.
 
 define void @vmax_v4i32(<4 x i32>* %m, <4 x i32> %a, <4 x i32> %b) {
 ; CHECK: vmax.s32 {{q[0-9]+}}, {{q[0-9]+}}, {{q[0-9]+}}
     %cmpres = icmp sgt <4 x i32> %a, %b
     %maxres = select <4 x i1> %cmpres, <4 x i32> %a,  <4 x i32> %b
     store <4 x i32> %maxres, <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.
 %T0_10 = type <16 x i16>
 %T1_10 = type <16 x i1>
 ; CHECK-LABEL: func_blend10:
 define void @func_blend10(%T0_10* %loadaddr, %T0_10* %loadaddr2,
                            %T1_10* %blend, %T0_10* %storeaddr) {
   %v0 = load %T0_10, %T0_10* %loadaddr
   %v1 = load %T0_10, %T0_10* %loadaddr2
   %c = icmp slt %T0_10 %v0, %v1
 ; CHECK: vmin.s16
 ; CHECK: vmin.s16
 ; COST: func_blend10
-; COST: cost of 40 {{.*}} select
+; COST: cost of 2 {{.*}} 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-LABEL: func_blend14:
 define void @func_blend14(%T0_14* %loadaddr, %T0_14* %loadaddr2,
                            %T1_14* %blend, %T0_14* %storeaddr) {
   %v0 = load %T0_14, %T0_14* %loadaddr
   %v1 = load %T0_14, %T0_14* %loadaddr2
   %c = icmp slt %T0_14 %v0, %v1
 ; CHECK: vmin.s32
 ; CHECK: vmin.s32
 ; COST: func_blend14
-; COST: cost of 41 {{.*}} select
+; COST: cost of 2 {{.*}} 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-LABEL: func_blend15:
 define void @func_blend15(%T0_15* %loadaddr, %T0_15* %loadaddr2,
                            %T1_15* %blend, %T0_15* %storeaddr) {
 ; CHECK: vmin.s32
 ; CHECK: vmin.s32
   %v0 = load %T0_15, %T0_15* %loadaddr
   %v1 = load %T0_15, %T0_15* %loadaddr2
   %c = icmp slt %T0_15 %v0, %v1
 ; COST: func_blend15
-; COST: cost of 82 {{.*}} select
+; COST: cost of 4 {{.*}} select
   %r = select %T1_15 %c, %T0_15 %v0, %T0_15 %v1
   store %T0_15 %r, %T0_15* %storeaddr
   ret void
 }
+
+; We adjusted the cost model of the following selects. When we improve code
+; lowering we also need to adjust the cost.
 %T0_18 = type <4 x i64>
 %T1_18 = type <4 x i1>
 ; CHECK-LABEL: func_blend18:
 define void @func_blend18(%T0_18* %loadaddr, %T0_18* %loadaddr2,
                            %T1_18* %blend, %T0_18* %storeaddr) {
 ; CHECK: vbsl
 ; CHECK: vbsl
   %v0 = load %T0_18, %T0_18* %loadaddr