[PATCH] D50465: [InstCombine] Optimize redundant 'signed truncation check pattern'.

[Roman Lebedev via Phabricator via llvm-commits]

lebedev.ri added inline comments.


================
Comment at: lib/Transforms/InstCombine/InstCombineAndOrXor.cpp:1311-1316
+  assert(I.getOpcode() == Instruction::And);
+  // We are looking for an 'and' of two icmp's
+  auto *ICmp0 = dyn_cast<ICmpInst>(I.getOperand(0));
+  auto *ICmp1 = dyn_cast<ICmpInst>(I.getOperand(1));
+  if (!(ICmp0 && ICmp1))
+    return nullptr;
----------------
spatel wrote:
> Can this be moved within InstCombiner::foldAndOfICmps() to eliminate these preliminary checks?
> 
> Any chance for shared functionality with foldAndOrOfICmpsOfAndWithPow2() or simplifyRangeCheck()?
I don't think i should squeeze it into `InstCombiner::foldAndOrOfICmpsOfAndWithPow2()`,
as that looks for a *very* specific, different pattern.

`InstCombiner::simplifyRangeCheck()` also does not look too promising,
we will then only handle the extreme cases (which is, admittedly, probably
sufficient for my needs right now), but not the general case :/

None of those handle this case because they mostly require for the both icmp's
to operate on the same value, and don't understand what
```
%t = add        i32 %arg,    128
%r = icmp   ult i32 %t,      256
```
(and similar patterns) mean.

So i have sunk it down into `InstCombiner::foldAndOfICmps()` for now..


Repository:
  rL LLVM

https://reviews.llvm.org/D50465