[llvm-dev] should we do this time-consuming transform in InstCombine?

[Roman Lebedev via llvm-dev]

On Tue, Dec 18, 2018 at 10:18 AM Zheng CZ Chen via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
>
> Hi,
Hi.

> There is an opportunity in instCombine for following instruction pattern:
>
> %mul = mul nsw i32 %b, %a
> %cmp = icmp sgt i32 %mul, -1
> %sub = sub i32 0, %a
> %mul2 = mul nsw i32 %sub, %b
> %cond = select i1 %cmp, i32 %mul, i32 %mul2
>
> Source code for above pattern:
> return (a*b) >=0 ? (a*b) : -a*b;
>
> Currently, llvm(-O3) can not recognize this as abs(a*b).
>
> I initially think we could do this in instCombine phase in opt. Below is what I think:
>
> %res = OP i32 %b, %a
> %sub = sub i32 0, %b
> %res2 = OP i32 %sub, %a
>
> We could do the transform:
> %res2 = OP i32 %sub, %a ==> %res2 = sub i32 0, %res
>
> Then we can get the advantage:
> 1: if %res2 is the only user of %sub, %sub can be eliminated;
> 2: if %res2 is not the only user of %sub, we could change some heavy instruction like div to sub;
> 3: expose more abs opportunity for later pass.
>
> But my concern is finding %res is a little compiling time-consuming.
> At least we need MIN(user_count(%b), user_count(%a)) times to check if instruction with same opcode and same operands exists.
In instcombine, no user checking is performed/allowed.
This should match that *specific* pattern (other than verifying the
correct equal binop types), although i have not tested it:

  ICmpInst::Predicate Pred;
  Value *A, *B, *Mul, *Sub, *Mul2;
  if (match(&SI,
            m_Select(m_ICmp(Pred,
                            m_CombineAnd(m_BinOp(m_Value(A), m_Value(B)),
                                         m_Value(Mul)),
                            m_AllOnes()),
                     m_Deferred(Mul),
                     m_CombineAnd(
                         m_c_BinOp(m_CombineAnd(m_Sub(m_Zero(), m_Deferred(A)),
                                                m_Value(Sub)),
                                   m_Deferred(B)),
                         m_Value(Mul2)))) &&
      Pred == ICmpInst::Predicate::ICMP_SGT) {
  }

> Could you guys give some comment? Is there any better idea for this transform?
>
> Thanks.
>
> BRS//
> Chen Zheng
> Power Compiler Backend Developer
Roman.

> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev