[ThePo1es]

Fork PR for mentor review before opening upstream. Fixes llvm#206708.

foldSelectBitTest only recognizes an equality bit test when the compared
value is literally and X, Pow2. Known bits from an llvm.assume, a range
attribute, or a dominating branch can establish the same single-bit fact
without the and being spelled out:

%c = icmp ult i32 %a, 2
call void @llvm.assume(i1 %c)
%cmp = icmp eq i32 %a, 0
%res = select i1 %cmp, i32 1, i32 2
  =>
%res = add nuw nsw i32 %a, 1

This falls back to computeKnownBits when the and pattern does not match,
the same way the decomposeBitTestICmp path directly below already refines
its mask with known bits. The query is only reached for selects whose arms
already passed the cheap structural checks hoisted to the function entry in
the preceding NFC commit, which also removes the pre-existing known-bits
query on the decomposeBitTestICmp path for selects that can never fold.

Proofs (generalized, symbolic C / single-bit P)

• X u< 2 => select (icmp eq X, 0), C, C+1 -> add X, C (eq and ne forms, C symbolic): https://alive2.llvm.org/ce/z/8oJ_7N
• X in {0, P}, P = 2^k => select (icmp eq X, 0), C, C op P -> op C, X for add/or/xor/sub (C, P symbolic): https://alive2.llvm.org/ce/z/q4Zeg2
• Concrete shapes and flag boundaries (C == INT_MAX yields add nuw only; sign bit; vector; and with a non-power-of-two mask): https://alive2.llvm.org/ce/z/vwT7Ke
• Negative (expected to fail): range(i32 0, 3) must not fold (counterexample a = 2), and nsw must not be added for the INT_MAX arm: https://alive2.llvm.org/ce/z/m5zjKW

Compile-time and real-world impact

• llvm-opt-benchmark (local run over the 2,784 original inputs, instcombine only): no IR changes, no regressions.
• compile-time-tracker: not run yet (will be requested in the upstream PR)

Notes

• Pointer compares are rejected explicitly (isIntOrIntVectorTy): the consumers zext/trunc the compared value.
• No one-use check is needed on the compare: CreateAnd is false on this path, so the transform is instruction-count neutral even when the compare has other users (@select_icmp_bittest_range_multiuse_cmp).
• Canonicalizing icmp eq X, 0 to trunc nuw in foldICmpUsingKnownBits would only cover the bit-0 case; @select_icmp_bittest_assume_high_bit (the single known bit is the sign bit) is not expressible that way.
• Pre-commit tests are in the first commit, including negative tests for a non-power-of-two known range, a fully unknown operand, non-adjacent select arms, and pointer compares.

Closes llvm#206708 (upstream).

Thanks to @Kristianerik, who described the same idea in the issue.

AI tool usage

Per the LLVM AI Tool Use Policy: AI assistance was used while writing parts of this patch and its tests. All transformations were verified with Alive2 (links above), llvm-lit passes over llvm/test/Transforms/InstCombine with an assertions-enabled build, and I reviewed the final implementation and every test expectation myself.

[ParkHanbum]

Style-wise, I think this looks better, but the current version also seems perfectly fine to submit as-is.

const APInt *TC, *FC;
bool ConstArmsCanFold =
    match(TrueVal, m_APInt(TC)) && match(FalseVal, m_APInt(FC)) &&
    ((*TC ^ *FC).isPowerOf2() || (*TC - *FC).isPowerOf2() ||
     (*FC - *TC).isPowerOf2());

bool BinOpArmsCanFold =
    match(FalseVal, m_BinOp(m_Specific(TrueVal), m_Power2())) ||
    match(TrueVal, m_BinOp(m_Specific(FalseVal), m_Power2()));

if (!ConstArmsCanFold && !BinOpArmsCanFold)
  return nullptr;