pmatos added inline comments.
================ Comment at: llvm/test/Transforms/InstCombine/fsh.ll:664 +; CHECK-NEXT: [[T1:%.*]] = and i32 [[A:%.*]], -65536 +; CHECK-NEXT: [[T2:%.*]] = call i32 @llvm.fshl.i32(i32 [[T1]], i32 [[T1]], i32 16) ; CHECK-NEXT: ret i32 [[T2]] ---------------- pmatos wrote: > nikic wrote: > > We still want to simplify this case. Could possibly be done by checking > > whether all demanded bits are zero for one of the operands in the rotate > > case. > Ah, yes, right. That should be just a simple shift right. Will see how to > still allow that change. Thanks. I am still looking into the best way to handle this case. The issue is that we only know if the demanded bits are zero when analyzing the uses of the value. This is done in SimplifyDemandedBits which in turn calls SimplifyDemandedUseBits, but we cannot call these functions to obtain demanded bits because they'll change the instruction straightaway. I was seeing if there was a way to do the checks inside the block of code already changed, but I don't think that'll be possible. I might have to add a check to SimplifyDemandedUseBits to only simplify in this specific case we want. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D150670/new/ https://reviews.llvm.org/D150670 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
