[llvm-bugs] [Bug 51968] New: Regression with GEPs after ""[InstCombine] Negator: -(X << C) --> X * (-1 << C)""
via llvm-bugs
llvm-bugs at lists.llvm.org
Sat Sep 25 08:36:36 PDT 2021
https://bugs.llvm.org/show_bug.cgi?id=51968
Bug ID: 51968
Summary: Regression with GEPs after ""[InstCombine] Negator:
-(X << C) --> X * (-1 << C)""
Product: libraries
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: enhancement
Priority: P
Component: Scalar Optimizations
Assignee: unassignedbugs at nondot.org
Reporter: david.bolvansky at gmail.com
CC: llvm-bugs at lists.llvm.org
llvm/test/Transforms/InstCombine/icmp.llm see test24
define i1 @test24_neg_offs(i32* %p, i64 %offs) {
%p1 = getelementptr inbounds i32, i32* %p, i64 %offs
%conv1 = ptrtoint i32* %p to i64
%conv2 = ptrtoint i32* %p1 to i64
%delta = sub i64 %conv1, %conv2
%cmp = icmp eq i64 %delta, 8
ret i1 %cmp
}
After that commit, LLVM produces:
define i1 @test24_neg_offs(i32* nocapture readnone %p, i64 %offs)
local_unnamed_addr #0 {
%p1.idx.neg = mul i64 %offs, -4
%cmp = icmp eq i64 %p1.idx.neg, 8
ret i1 %cmp
}
instead of:
define i1 @mulnsw(i64 %offs) local_unnamed_addr #0 {
%cmp = icmp eq i64 %offs, -2
ret i1 %cmp
}
And there is nothing we can do that form to simplify it (missing nsw).
But.. we can add nsw on mul and everything will work fine.
define i1 @src(* %p, i64 %offs) {
%0:
%p1 = gep inbounds * %p, 4 x i64 %offs
%conv1 = ptrtoint * %p to i64
%conv2 = ptrtoint * %p1 to i64
%delta = sub i64 %conv1, %conv2
%cmp = icmp eq i64 %delta, 8
ret i1 %cmp
}
=>
define i1 @tgt(* nocapture noread nowrite %p, i64 %offs) {
%0:
%p1.idx.neg = mul nsw i64 %offs, -4
%cmp = icmp eq i64 %p1.idx.neg, 8
ret i1 %cmp
}
Transformation seems to be correct!
And then LLVM can optimize:
define i1 @mulnsw(i64 %offs) local_unnamed_addr #0 {
%p1.idx.neg = mul nsw i64 %offs, -4
%cmp = icmp eq i64 %p1.idx.neg, 8
ret i1 %cmp
}
To:
define i1 @mulnsw(i64 %offs) local_unnamed_addr #0 {
%cmp = icmp eq i64 %offs, -2
ret i1 %cmp
}
--
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210925/c6ac65ca/attachment-0001.html>
More information about the llvm-bugs
mailing list