[LLVMdev] Instruction combiner multiplication canonicalization

[Nema, Ashutosh]

Thanks Sanjoy for detailed explanation. 
I got your point, there is a possibility of overflow. 
Hence we can't do explained transformation.

Regards,
Ashutosh

-----Original Message-----
From: Sanjoy Das [mailto:sanjoy at playingwithpointers.com] 
Sent: Wednesday, April 15, 2015 10:43 PM
To: Nema, Ashutosh
Cc: llvmdev at cs.uiuc.edu; Nick Lewycky
Subject: Re: Instruction combiner multiplication canonicalization

> Consider following scenarios:
> 1) If input is ((X+C1)*C2)<nsw>
> Then post canonicalization output should be (X*C2+C1*C2) <nsw>

Say X = INT_SMAX, C1 = C2 = 1.

Then (X + C1) = INT_SMIN and INT_SMIN * 1 does not overflow.

But X*C2 = INT_SMAX and C1*C2 = 1 and INT_SMAX + 1 does sign overflow.

> 2) If input is (((X+C1)<nsw>)*C2)<nsw>
> Then post canonicalization output should be ((X*C2) <nsw>+(C1*C2) <nsw>)
> <nsw>

This is more interesting, let X = INT_SMIN, C1 = 1, C2 -1.

Then the first set of operations don't sign overflow.  But X * C2 =
INT_SMIN * -1 will sign overflow

-- Sanjoy