[BFI][CGP] Assertion failed: (Match && "BFI mismatch"), function verifyMatch, file BlockFrequencyInfoImpl.h, line 1800.

[PavelKopyl]

Hello,

I faced with a failed assertion in BlockFrequencyInfoImpl on LLVM TOT.

Repro case: bfi-bug.ll

; Function Attrs: noreturn
define private fastcc void @__test() unnamed_addr #0 {
entry:
  br i1 undef, label %join, label %division
 
division:
  br label %join
 
join:
  tail call fastcc void @encode(i32 1024)
  unreachable
}
 
define private fastcc void @encode(i32 %0) unnamed_addr {
entry:
  unreachable
}
 
attributes #0 = { noreturn }

LLVM should be built in the debug mode.

opt --codegenprepare --mtriple=riscv32 --cgp-verify-bfi-updates bfi-bug3.ll -S
Freq mismatch: entry 8 vs 16
This
block-frequency-info: __test
 - entry: float = 1.0, int = 8

Other
block-frequency-info: __test
 - entry: float = 1.0, int = 16

Assertion failed: (Match && "BFI mismatch"), function verifyMatch, file BlockFrequencyInfoImpl.h, line 1800.

Please, note that the actual triplet value doesn't seem to matter, as the failed assertion is in target-independent part of the code.

As I understand, the issues happens because the BFI updates are not synced with actual code transformations performed by CodeGenPrepare pass. Namely, the BBs frequency info, expressed as FrequencyData::Integer, is not updated.

Does it make sense to compare Scaled values instead?

--- a/llvm/include/llvm/Analysis/BlockFrequencyInfoImpl.h
+++ b/llvm/include/llvm/Analysis/BlockFrequencyInfoImpl.h
@@ -1777,7 +1777,7 @@ void BlockFrequencyInfoImpl<BT>::verifyMatch(
         BlockNode OtherNode = OtherValidNodes[BB];
         const auto &Freq = Freqs[Node.Index];
         const auto &OtherFreq = Other.Freqs[OtherNode.Index];
-        if (Freq.Integer != OtherFreq.Integer) {
+        if (Freq.Scaled != OtherFreq.Scaled) {
           Match = false;
           dbgs() << "Freq mismatch: " << bfi_detail::getBlockName(BB) << " "
                  << Freq.Integer << " vs " << OtherFreq.Integer << "\n";

[hjyamauchi]

You may be right. I don't remember exactly but it seems that both the float and int values can be slightly off due to the float point precision and the min/max values that exist in the same CFG. I don't work in this area any more so you're welcome to improve the code if you'd like, not worry about it or revert the commit.

…

On Fri, Jul 28, 2023 at 5:22 AM PavelKopyl ***@***.***> wrote: Assigned #64197 <#64197> to @hjyamauchi <https://github.com/hjyamauchi>. — Reply to this email directly, view it on GitHub <#64197 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANQ3RQDHTCEQ7OQXV3VHAYTXSOVG7ANCNFSM6AAAAAA23NTN5M> . You are receiving this because you were assigned.Message ID: ***@***.***>

[PavelKopyl]

@hjyamauchi Thank you for the reply.