[SLPVectorizer] Add reduction of integer not vectorized

[nikic]

target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128" 
target triple = "x86_64-unknown-linux-gnu"
  
define i16 @test(ptr %x) { 
start:
  %_51 = load i64, ptr %x, align 8
  %_4.i.i.i.sroa.4.16.extract.trunc = trunc i64 %_51 to i16
  %0 = and i16 %_4.i.i.i.sroa.4.16.extract.trunc, 255
  %1 = trunc i64 %_51 to i16
  %2 = lshr i16 %1, 8
  %_6.0.i.i.i.i.i.i.i.i.i.1 = add nuw nsw i16 %0, %2
  %_4.i.i.i.sroa.4.18.extract.shift = lshr i64 %_51, 16
  %_4.i.i.i.sroa.4.18.extract.trunc = trunc i64 %_4.i.i.i.sroa.4.18.extract.shift to i16
  %3 = and i16 %_4.i.i.i.sroa.4.18.extract.trunc, 255
  %_6.0.i.i.i.i.i.i.i.i.i.2 = add nuw nsw i16 %_6.0.i.i.i.i.i.i.i.i.i.1, %3
  %_4.i.i.i.sroa.4.19.extract.shift = lshr i64 %_51, 24
  %_4.i.i.i.sroa.4.19.extract.trunc = trunc i64 %_4.i.i.i.sroa.4.19.extract.shift to i16
  %4 = and i16 %_4.i.i.i.sroa.4.19.extract.trunc, 255
  %_6.0.i.i.i.i.i.i.i.i.i.3 = add nuw nsw i16 %_6.0.i.i.i.i.i.i.i.i.i.2, %4
  %_4.i.i.i.sroa.4.20.extract.shift = lshr i64 %_51, 32
  %_4.i.i.i.sroa.4.20.extract.trunc = trunc i64 %_4.i.i.i.sroa.4.20.extract.shift to i16
  %5 = and i16 %_4.i.i.i.sroa.4.20.extract.trunc, 255
  %_6.0.i.i.i.i.i.i.i.i.i.4 = add nuw nsw i16 %_6.0.i.i.i.i.i.i.i.i.i.3, %5
  %_4.i.i.i.sroa.4.21.extract.shift = lshr i64 %_51, 40
  %_4.i.i.i.sroa.4.21.extract.trunc = trunc i64 %_4.i.i.i.sroa.4.21.extract.shift to i16
  %6 = and i16 %_4.i.i.i.sroa.4.21.extract.trunc, 255
  %_6.0.i.i.i.i.i.i.i.i.i.5 = add nuw nsw i16 %_6.0.i.i.i.i.i.i.i.i.i.4, %6
  %_4.i.i.i.sroa.4.22.extract.shift = lshr i64 %_51, 48
  %_4.i.i.i.sroa.4.22.extract.trunc = trunc i64 %_4.i.i.i.sroa.4.22.extract.shift to i16 
  %7 = and i16 %_4.i.i.i.sroa.4.22.extract.trunc, 255
  %_6.0.i.i.i.i.i.i.i.i.i.6 = add nuw nsw i16 %_6.0.i.i.i.i.i.i.i.i.i.5, %7
  %_4.i.i.i.sroa.4.23.extract.shift = lshr i64 %_51, 56
  %_4.i.i.i.sroa.4.23.extract.trunc = trunc i64 %_4.i.i.i.sroa.4.23.extract.shift to i16
  %_6.0.i.i.i.i.i.i.i.i.i.7 = add nuw nsw i16 %_6.0.i.i.i.i.i.i.i.i.i.6, %_4.i.i.i.sroa.4.23.extract.trunc
  ret i16 %_6.0.i.i.i.i.i.i.i.i.i.7
} 

This could converted into a zext <8 x i8> to <8 x i16> followed by vector.reduce.add, but currently isn't. (I expect this to be profitable based on psadbw, but I'm not particularly familiar with this.)

[nikic]

@alexey-bataev Just to double check, SLPVectorizer currently doesn't handle "vectors" that are represented as integers (with "extracts" done via lshr+trunc) at all, right?

[alexey-bataev]

@alexey-bataev Just to double check, SLPVectorizer currently doesn't handle "vectors" that are represented as integers (with "extracts" done via lshr+trunc) at all, right?

Yes, it is not aware of this pattern yet. Plus, looks like there is still the problem with the copyable elements, because of this we have very small non-profitable tree

[RKSimon]

We have a similar problem on #48223 where SROA leaves us with a lot of zext/shift/or/trunc ops to pack elements into i64 elements.

[dianqk]

This is another similar IR.

; ModuleID = 'simd_wide_sum.e3e91bc971ec0fa5-cgu.0'
source_filename = "simd_wide_sum.e3e91bc971ec0fa5-cgu.0"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

; Function Attrs: mustprogress nofree norecurse nosync nounwind nonlazybind willreturn memory(argmem: read) uwtable
define noundef i16 @wider_reduce_into_iter(ptr noalias nocapture noundef readonly align 8 dereferenceable(8) %x) unnamed_addr #0 personality ptr @rust_eh_personality {
start:
  %0 = load i64, ptr %x, align 8
  %self.sroa.4.16.extract.trunc = trunc i64 %0 to i16
  %_0.i.i.i.i.i.i.i.i.i = and i16 %self.sroa.4.16.extract.trunc, 255
  %1 = trunc i64 %0 to i16
  %2 = lshr i16 %1, 8
  %_4.0.i.i.i.i.i.i.i.i.1 = add nuw nsw i16 %_0.i.i.i.i.i.i.i.i.i, %2
  %self.sroa.4.18.extract.shift = lshr i64 %0, 16
  %self.sroa.4.18.extract.trunc = trunc i64 %self.sroa.4.18.extract.shift to i16
  %_0.i.i.i.i.i.i.i.i.i.2 = and i16 %self.sroa.4.18.extract.trunc, 255
  %_4.0.i.i.i.i.i.i.i.i.2 = add nuw nsw i16 %_4.0.i.i.i.i.i.i.i.i.1, %_0.i.i.i.i.i.i.i.i.i.2
  %self.sroa.4.19.extract.shift = lshr i64 %0, 24
  %self.sroa.4.19.extract.trunc = trunc i64 %self.sroa.4.19.extract.shift to i16
  %_0.i.i.i.i.i.i.i.i.i.3 = and i16 %self.sroa.4.19.extract.trunc, 255
  %_4.0.i.i.i.i.i.i.i.i.3 = add nuw nsw i16 %_4.0.i.i.i.i.i.i.i.i.2, %_0.i.i.i.i.i.i.i.i.i.3
  %self.sroa.4.20.extract.shift = lshr i64 %0, 32
  %self.sroa.4.20.extract.trunc = trunc i64 %self.sroa.4.20.extract.shift to i16
  %_0.i.i.i.i.i.i.i.i.i.4 = and i16 %self.sroa.4.20.extract.trunc, 255
  %_4.0.i.i.i.i.i.i.i.i.4 = add nuw nsw i16 %_4.0.i.i.i.i.i.i.i.i.3, %_0.i.i.i.i.i.i.i.i.i.4
  %self.sroa.4.21.extract.shift = lshr i64 %0, 40
  %self.sroa.4.21.extract.trunc = trunc i64 %self.sroa.4.21.extract.shift to i16
  %_0.i.i.i.i.i.i.i.i.i.5 = and i16 %self.sroa.4.21.extract.trunc, 255
  %_4.0.i.i.i.i.i.i.i.i.5 = add nuw nsw i16 %_4.0.i.i.i.i.i.i.i.i.4, %_0.i.i.i.i.i.i.i.i.i.5
  %self.sroa.4.22.extract.shift = lshr i64 %0, 48
  %self.sroa.4.22.extract.trunc = trunc i64 %self.sroa.4.22.extract.shift to i16
  %_0.i.i.i.i.i.i.i.i.i.6 = and i16 %self.sroa.4.22.extract.trunc, 255
  %_4.0.i.i.i.i.i.i.i.i.6 = add nuw nsw i16 %_4.0.i.i.i.i.i.i.i.i.5, %_0.i.i.i.i.i.i.i.i.i.6
  %self.sroa.4.23.extract.shift = lshr i64 %0, 56
  %self.sroa.4.23.extract.trunc = trunc i64 %self.sroa.4.23.extract.shift to i16
  %_4.0.i.i.i.i.i.i.i.i.7 = add nuw nsw i16 %_4.0.i.i.i.i.i.i.i.i.6, %self.sroa.4.23.extract.trunc
  ret i16 %_4.0.i.i.i.i.i.i.i.i.7
}

; Function Attrs: nonlazybind uwtable
declare noundef i32 @rust_eh_personality(i32 noundef, i32 noundef, i64 noundef, ptr noundef, ptr noundef) unnamed_addr #1

attributes #0 = { mustprogress nofree norecurse nosync nounwind nonlazybind willreturn memory(argmem: read) uwtable "probe-stack"="inline-asm" "target-cpu"="x86-64" }
attributes #1 = { nonlazybind uwtable "probe-stack"="inline-asm" "target-cpu"="x86-64" }

!llvm.module.flags = !{!0, !1}
!llvm.ident = !{!2}

!0 = !{i32 8, !"PIC Level", i32 2}
!1 = !{i32 2, !"RtLibUseGOT", i32 1}
!2 = !{!"rustc version 1.74.0-dev"}

https://llvm.godbolt.org/z/Y4b3x5x9z is a similar IR but gets vectorized in LoopVectorizePass.