[AArch64][Cost Modlel] Why the cost of vector extract cost so high?

Allen · May 22, 2023, 1:58pm

hi,

Now, the default value of VectorInsertExtractBaseCost is 3, while it usual can be expressed with a mov, such as mov d0, v0.d[1] for double.
llvm-project/insert-extract.ll at f3202b9da663fb618ea4e3dc75a1352bdb2f969a · llvm/llvm-project · GitHub
And even more, a load instruction is estimated cost 1, much smaller than the insertelement/extractelement ?

llvm/llvm-project/blob/f3202b9da663fb618ea4e3dc75a1352bdb2f969a/llvm/test/Analysis/CostModel/AArch64/insert-extract.ll#L115


      
          
          
;; LD1: Load one single-element structure to one lane of one register.
          
          
define <8 x i8> @LD1_B(<8 x i8> %vec, ptr noundef %i) {
          ; KRYO-LABEL: 'LD1_B'
          ; KRYO-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v1 = load i8, ptr %i, align 1
          ; KRYO-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %v2 = insertelement <8 x i8> %vec, i8 %v1, i32 1
          ; KRYO-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret <8 x i8> %v2
          ;
          ; NEO-LABEL: 'LD1_B'
          ; NEO-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %v1 = load i8, ptr %i, align 1
          ; NEO-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2 = insertelement <8 x i8> %vec, i8 %v1, i32 1
          ; NEO-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret <8 x i8> %v2
          ;
          entry:
            %v1 = load i8, ptr %i, align 1
            %v2 = insertelement <8 x i8> %vec, i8 %v1, i32 1
            ret <8x i8> %v2
          }
          
          
define <4 x i16> @LD1_H(<4 x i16> %vec, ptr noundef %i) {

So why the cost of vector extract cost is so high? or is there some cases need many instructions to implement the vector extract ?

fhahn · May 22, 2023, 2:03pm

See ⚙ D142359 [TTI][AArch64] Cost model vector INS instructions for some additional context.

@nilanjana87 is currently looking into adjusting the cost for some CPUs, but as mentioned in ⚙ D142359 [TTI][AArch64] Cost model vector INS instructions I think this requires careful testing as it can have unintended knock-on effects triggering more SLP vectorization for example, which in turn can expose other weaknesses of the cost model.

fhahn · May 22, 2023, 2:04pm

Note that this is on AArch64 only, might be good to adjust the title to make this clear.

Allen · May 23, 2023, 1:05am

Thanks for your comment, added AArch64 for this title