Unnecessarily lost inbounds attribute in GEP

[amehsan]

We lose inbounds flag of GEPs for different reasons during compilation. My primary intention of opening the issue is to gather information and record them here., since I am not aware of any pressing need to fix this. Keeping inbounds could help loop opts in general, but first I want to work with some examples and see how things work in different passes (as opposed to developing a quick fix for a given case).

[amehsan]

for this loop

void foo (char * restrict A,   int64_t   n) {
       for (int64_t j = 10;  j < n ; j ++) {
         A[j+6] = A[j-3] + 4;
       }
}

before CSE:

for.body:                                         ; preds = %for.cond
  %0 = getelementptr i8, ptr %A, i64 %j.0   <-------------
  %arrayidx = getelementptr i8, ptr %0, i64 -3
  %1 = load i8, ptr %arrayidx, align 1, !tbaa !10
  %add = add i8 %1, 4
  %2 = getelementptr inbounds nuw i8, ptr %A, i64 %j.0   <------------ 
  %arrayidx3 = getelementptr inbounds nuw i8, ptr %2, i64 6
  store i8 %add, ptr %arrayidx3, align 1, !tbaa !10
  %inc = add nuw nsw i64 %j.0, 1
  br label %for.cond, !llvm.loop !11

After CSE

for.body:                                         ; preds = %for.cond
  %0 = getelementptr i8, ptr %A, i64 %j.0   <---------- CSEed but lost the flag
  %arrayidx = getelementptr i8, ptr %0, i64 -3
  %1 = load i8, ptr %arrayidx, align 1, !tbaa !10
  %add = add i8 %1, 4
  %arrayidx3 = getelementptr inbounds nuw i8, ptr %0, i64 6
  store i8 %add, ptr %arrayidx3, align 1, !tbaa !10
  %inc = add nuw nsw i64 %j.0, 1
  br label %for.cond, !llvm.loop !11

CSE finds two GEPs with the same offset and base and it even calls combineIRFlags but it fails to keep the inbounds.

The first two GEPs are created by InstCombine as discussed here
#159846 (comment)

[amehsan]

Also as discussed here sometimes we fail to prove a GEP is inbounds in the first place.

#162280 (comment)

also see the subsequent comment:

#162280 (comment)