search for: machinebranchprob

Hello LLVM-Dev, The other day as I was digging through lldb’s test suite I noticed they support something kinda neat. In their python test harness, the attribute they use to denote expected failures supports a parameter for specifying the bug number. This got me thinking. I believe that any test that is marked XFAIL is a bug, and we can use LIT to enforce that. So I wrote a patch

...at 10:12 AM, Krzysztof Parzyszek via llvm-dev <llvm-dev at lists.llvm.org> wrote: > > On 9/28/2016 11:39 AM, Chris Bieneman via llvm-dev wrote: >> I believe that any test that is marked XFAIL is a bug > > I don't know if that's true. > > test/CodeGen/Generic/MachineBranchProb.ll: > > ; ARM & AArch64 run an extra SimplifyCFG which disrupts this test. > ; XFAIL: arm,aarch64 > > ; Hexagon runs passes that renumber the basic blocks, causing this test > ; to fail. > ; XFAIL: hexagon > > -Krzysztof > > -- > Qualcomm Innovation Cent...

...lt MBB, one edge for the jumptable, and another one for the jump handling the case when the index is not in the jumptable. For some reason, our backend assigns those edges different weights (let's say 24 for the default edge and 16 for the others). In LLVM 3.1, getSumForBlock in lib/CodeGen/MachineBranchProbability.cpp used getEdgeWeight(MBB*, MBB*), which always return the weight of the first edge to the default block, i.e., it calculated the sum of the weights as 24+16+24+16+16 = 96. In LLVM 3.2 getSumForBlock was changed to use getEdgeWeight(MBB*, MBB::succ_iterator), causing it to calculate the...