[llvm-bugs] [Bug 47241] New: ValueTracking: Compute known bits doesn't leverage range information on gep

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=47241

            Bug ID: 47241
           Summary: ValueTracking: Compute known bits doesn't leverage
                    range information on gep
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Global Analyses
          Assignee: unassignedbugs at nondot.org
          Reporter: quentin.colombet at gmail.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 23863
  --> https://bugs.llvm.org/attachment.cgi?id=23863&action=edit
Case that gets simplified

The ValueTracking analysis only tries to fill up trailing zeros when going
through GEPs.
In particular, the analysis doesn't take advantage of the range information
that may be available, resulting and suboptimal code.

* Example *

Consider the following snippet:
```
define i1 @test4_gep(i64* %ptr) {
entry:
  %val = load volatile i64, i64* %ptr, !range !{i64 64, i64 65536}
  %valptr = inttoptr i64 %val to float*
  %gep = getelementptr float, float* %valptr, i64 128 
  %res = icmp ugt float* %gep, inttoptr (i64 523 to float*)
  ret i1 %res
}
```

Using the range information this example can be simplified into the following
code, since 64 + 4*128 is guaranteed to be bigger than 523:
```
define i1 @test4_gep(i64* %ptr) {
entry:
  %val = load volatile i64, i64* %ptr, !range !{i64 64, i64 65536}
  ret i1 true
}
```

But given ValueTracking doesn't leverage the range information, no
simplification happens.

Now, rewrite this code with an add like so:
```
define i1 @test3_add(i64* %ptr) {
entry:
  %val = load volatile i64, i64* %ptr, !range !{i64 64, i64 65536}
  %valPlus512 = add i64 %val, 512
  %res = icmp ugt i64 %valPlus512, 523
  ret i1 %res
}
```

And ValueTracking provides the proper known bits and the simplification can
happen.

* To Reproduce *
# Not simplified.
opt -S -instcombine value_tracking_gep.ll  -o -
# Simplified.
opt -S -instcombine value_tracking_add.ll  -o -

Note: I have to use a volatile load to prevent instcombine for dropping the
range information. I'll file a separate PR for that.

-- 
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/20200819/6fe5cd82/attachment.html>