[PATCH] D29774: [InstCombine] fold icmp sgt/slt (add nsw X, C2), C --> icmp sgt/slt X, (C - C2)

[Sanjay Patel via Phabricator via llvm-commits]

spatel created this revision.
Herald added a subscriber: mcrosier.

I found one special case of this transform for 'slt 0', so I removed that and added the general transform.

Here are Alive programs that try to verify the correctness of the logic:
http://rise4fun.com/Alive/GH

(is there a better way to write the no-overflow constraints?)

Name: slt_no_overflow
Pre: ((C1 < 0) && (C2 < 0)) || ((C1 >= 0) && (C2 >= 0)) || (((C2 - C1) < 0) && (C2 < 0)) || (((C2 - C1) >= 0) && (C2 >= 0))
%a = add nsw i8 %x, C1
%b = icmp slt %a, C2

  =>

%b = icmp slt %x, C2 - C1

Name: sgt_no_overflow
Pre: ((C1 < 0) && (C2 < 0)) || ((C1 >= 0) && (C2 >= 0)) || (((C2 - C1) < 0) && (C2 < 0)) || (((C2 - C1) >= 0) && (C2 >= 0))
%a = add nsw i8 %x, C1
%b = icmp sgt %a, C2

  =>

%b = icmp sgt %x, C2 - C1

And here are the cases where the subtraction does overflow:
http://rise4fun.com/Alive/ao

Name: slt_overflow_false
Pre: C2 < 0 && C1 >= 0 && (C2 - C1) >= 0
%a = add nsw i8 %x, C1
%b = icmp slt %a, C2

  =>

%b = false

Name: slt_overflow_true
Pre: C2 >= 0 && C1 < 0 && (C2 - C1) < 0
%a = add nsw i8 %x, C1
%b = icmp slt %a, C2

  =>

%b = true

Name: sgt_overflow_true
Pre: C2 < 0 && C1 >= 0 && (C2 - C1) >= 0
%a = add nsw i8 %x, C1
%b = icmp sgt %a, C2

  =>

%b = true

Name: sgt_overflow_false
Pre: C2 >= 0 && C1 < 0 && (C2 - C1) < 0
%a = add nsw i8 %x, C1
%b = icmp sgt %a, C2

  =>

%b = false


https://reviews.llvm.org/D29774

Files:
  lib/Transforms/InstCombine/InstCombineCompares.cpp
  test/Transforms/InstCombine/icmp-add.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D29774.87837.patch
Type: text/x-patch
Size: 4534 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170209/202f2ba9/attachment.bin>