[DAG] Regression due to creation of avgfloors nodes

[RKSimon]

; llc -mtriple=riscv64 test.ll -o -
define signext i64 @func000000000000002b(i32 signext %0) #0 {
entry:
  %1 = zext nneg i32 %0 to i64
  %2 = add nsw i64 %1, -1
  %3 = lshr i64 %2, 1
  %4 = add nuw nsw i64 %3, 1
  %5 = and i64 %4, 9223372036854775806
  ret i64 %5
}

Before:

func000000000000002b:
        addi    a0, a0, -1
        srli    a0, a0, 1
        addi    a0, a0, 1
        andi    a0, a0, -2
        ret

After:

func000000000000002b:
        addi    a0, a0, -1
        srli    a0, a0, 1
        addi    a0, a0, 1
        li      a1, -3
        srli    a1, a1, 1
        and     a0, a0, a1
        ret

Reported by @dtcxzyw on #92096

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Simon Pilgrim (RKSimon)

```ll ; llc -mtriple=riscv64 test.ll -o - define signext i64 @func000000000000002b(i32 signext %0) #0 { entry: %1 = zext nneg i32 %0 to i64 %2 = add nsw i64 %1, -1 %3 = lshr i64 %2, 1 %4 = add nuw nsw i64 %3, 1 %5 = and i64 %4, 9223372036854775806 ret i64 %5 } ``` Before: ```asm func000000000000002b: addi a0, a0, -1 srli a0, a0, 1 addi a0, a0, 1 andi a0, a0, -2 ret ``` After: ```asm func000000000000002b: addi a0, a0, -1 srli a0, a0, 1 addi a0, a0, 1 li a1, -3 srli a1, a1, 1 and a0, a0, a1 ret ```

Reported by @dtcxzyw on #92096

[topperc]

Is this also a regression X86?

clang-18 gives

func000000000000002b:                   # @func000000000000002b
        mov     eax, edi
        dec     rax
        shr     rax
        inc     rax
        and     rax, -2
        ret

by local build gives

func000000000000002b:                   # @func000000000000002b
        .cfi_startproc
# %bb.0:                                # %entry
        movl    %edi, %ecx
        decq    %rcx
        shrq    %rcx
        incq    %rcx
        movabsq $9223372036854775806, %rax      # imm = 0x7FFFFFFFFFFFFFFE
        andq    %rcx, %rax
        retq

[RKSimon]

Yes, the issue isn't target specific - the avgfloors node gets created inside SimplifyDemandedBits (starting from the and), which in turn strips the flags from %4 = add nuw nsw i64 %3, 1 - which a later call to ShrinkDemandedConstant from the and was relying on to simplify the constant mask.

[dtcxzyw]

@RKSimon I confirmed that all regressions were fixed. Thank you!