Seek for rebase

[gitoleg]

Hi all!
I have the next question: when do you plan the next rebase?
The reason I'm asking is that the previous one was done in the moment when a bug existed in the trunk and it kind of affects ClangIR as well.

Reprocase:
clang tmp.c -O2 --target=aarch64-none-linux-gnu -c (one may enable -fclangir as well, it doesn't change anything)

typedef struct {
    char d;
} A;

typedef struct {
    A a;
} B;

#define MACRO(a, b) a <= b ? a : 0

B* ptr;

char i, j, k, l, m, n;

void foo() {
    char ag = i ? ptr->a.d : 0;
    char af = MACRO(ag, j);
    l = af / k;
    m = 1 ? af % k : n;
}

This code fails with the weird error in the llvm's DAGCombiner

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp:16156: llvm::SDValue {anonymous}::DAGCombiner::visitFREEZE(llvm::SDNode*): Assertion `DAG.isGuaranteedNotToBeUndefOrPoison(R, false) && "Can't create node that may be undef/poison!"' failed.

I did not try to bisect to identify the exact commit in trunk, but the point is - the bug is already fixed, I checked it with the current main branch of llvm-project and it would be great if we do the rebasing one day

cc @bcardosolopes

[smeenai]

I'll look into it.

[smeenai]

I pushed a new rebase.

[seven-mile]

@smeenai It seems this too recent commit is not included. Do we have a standard / policy of choosing a base commit to synchronize with?

[smeenai]

I picked a base commit from Tuesday, so we missed out by one day :) I can just run another rebase though, it should hopefully be pretty quick and painless given that I just performed one.

Edited to add: we generally pick a base commit on the day we start the rebase, make sure it's in a good state (passed check-llvm check-clang check-mlir etc.), and then stick with that base through the rebase. For the rebase itself, we make sure check-clang-cir passes after every single rebased commit, which is why it takes a few days to perform.

[bcardosolopes]

@smeenai thanks for all rebase work and getting our community happy to be living on the edge :D

[smeenai]

Pushed a new rebase. @seven-mile your work enabled us to get rid of a ClangIR-specific change we had to Targets.h, which is very nice :)