[X86] cttz.elts +avx512f always returns 16

[yrouban]

The problem is that with target +avx512f the following intrinsic call returns 16 (the full target vector length) instead of 4 (the number of elements in the input vector):
call i32 @llvm.experimental.cttz.elts.i32.v4i1(<4 x i1> zeroinitializer, i1 false)

Here is the full repro main.ll:

target triple = "x86_64-unknown-linux"

define i32 @main() #0 {
entry:
  %first.inactive.lane = call i32 @llvm.experimental.cttz.elts.i32.v4i1(<4 x i1> zeroinitializer, i1 false)
  ret i32 %first.inactive.lane
}

; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.experimental.cttz.elts.i32.v4i1(<4 x i1>, i1 immarg) #1

attributes #0 = \{ "target-features"="+avx512f" }
attributes #1 = \{ nocallback nofree nosync nounwind speculatable willreturn memory(none) }

To reproduce make an x86_64 executable and run:

$ llc main.ll --filetype=obj -o main.o && g++ main.o -o main && ./main; echo "Returned $?. Expected 4"
Returned 16. Expected 4

Compare the outputs of trunk and llc 22.1.0 (https://godbolt.org/z/Pxb54K4M9 ):

.LCPI0_0:
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
main: # @main
  vphminposuw xmm0, xmmword ptr [rip + .LCPI0_0]
  vmovd eax, xmm0
  add al, 17
  movzx eax, al
  ret

and

main: # @main
  mov eax, 4
  ret

[llvmorg-github-actions]

@llvm/issue-subscribers-backend-x86

Author: Yevgeny Rouban (yrouban)

Details

The problem is that with target +avx512f the following intrinsic call returns 16 (the full target vector length) instead of 4 (the number of elements in the input vector): `call i32 @llvm.experimental.cttz.elts.i32.v4i1(<4 x i1> zeroinitializer, i1 false)`

Here is the full repro main.ll:

target triple = "x86_64-unknown-linux"

define i32 @<!-- -->main() #<!-- -->0 {
entry:
  %first.inactive.lane = call i32 @<!-- -->llvm.experimental.cttz.elts.i32.v4i1(&lt;4 x i1&gt; zeroinitializer, i1 false)
  ret i32 %first.inactive.lane
}

; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @<!-- -->llvm.experimental.cttz.elts.i32.v4i1(&lt;4 x i1&gt;, i1 immarg) #<!-- -->1

attributes #<!-- -->0 = \{ "target-features"="+avx512f" }
attributes #<!-- -->1 = \{ nocallback nofree nosync nounwind speculatable willreturn memory(none) }

To reproduce make an x86_64 executable and run:

$ llc main.ll --filetype=obj -o main.o && g++ main.o -o main && ./main; echo "Returned $?. Expected 4"
Returned 16. Expected 4

Compare the outputs of trunc and llc 22.1.0 (https://godbolt.org/z/Pxb54K4M9 ):

.LCPI0_0:
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
  .byte 255 # 0xff
  .byte 0 # 0x0
main: # @<!-- -->main
  vphminposuw xmm0, xmmword ptr [rip + .LCPI0_0]
  vmovd eax, xmm0
  add al, 17
  movzx eax, al
  ret

and

main: # @<!-- -->main
  mov eax, 4
  ret

[RKSimon]

Seems like a problem in generic legalization:

Optimized type-legalized selection DAG: %bb.0 'main:entry'
SelectionDAG has 8 nodes:
    t0: ch,glue = EntryToken
      t2: v4i1 = BUILD_VECTOR Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>
    t3: i32 = cttz_elts t2
  t6: ch,glue = CopyToReg t0, Register:i32 $eax, t3
  t7: ch = X86ISD::RET_GLUE t6, TargetConstant:i32<0>, Register:i32 $eax, t6:1



Legalizing vector op: t3: i32 = cttz_elts t2
Expanding
Creating constant: t9: i8 = Constant<1>
Creating constant: t10: i8 = Constant<2>
Creating constant: t11: i8 = Constant<3>
Creating new node: t12: v4i8 = BUILD_VECTOR Constant:i8<0>, Constant:i8<1>, Constant:i8<2>, Constant:i8<3>
Creating new node: t13: v16i8 = poison
Creating constant: t14: i64 = Constant<0>
Creating new node: t15: v16i8 = insert_subvector poison:v16i8, t12, Constant:i64<0>
Creating new node: t16: v16i1 = BUILD_VECTOR Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>
Creating new node: t17: v16i1 = insert_subvector t16, t2, Constant:i64<0>
Creating constant: t18: i8 = Constant<16>
Creating new node: t19: v16i8 = BUILD_VECTOR Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>, Constant:i8<16>
Creating new node: t20: v16i8 = sub t19, t15
Creating new node: t21: v16i8 = BUILD_VECTOR Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>, Constant:i8<0>
Creating new node: t22: v16i8 = vselect t17, t20, t21
Creating new node: t23: i8 = vecreduce_umax t22
Creating new node: t24: i8 = sub Constant:i8<16>, t23
Creating new node: t25: i32 = zero_extend t24