Bolt "BOLT-WARNING: sizes differ for function"

Subprojects BOLT
,
1

Currently encountering some problems in the use of bolt.
In --enable-bat mode,

  1. In no-lbr mode, when bolt2perf is performed for the second time, BOLT-WARNING: sizes differ for function, I collected the logs, and warnings will appear for all “*.org.0” symbols.
  2. If use lbr mode, although no alarm will appear, the if (FDEI->first != Address) branch will be entered.
  1094      auto FDEI = CFIRdWrt->getFDEs().lower_bound(Address);
  1095      if (FDEI != CFIRdWrt->getFDEs().end()) {
  1096        const dwarf::FDE &FDE = *FDEI->second;
  1097        if (FDEI->first != Address) {
....
  1114        } else if (FDE.getAddressRange() != SymbolSize) {
  1115          if (SymbolSize) {
  1116            // Function addresses match but sizes differ.
  1117            errs() << "BOLT-WARNING: sizes differ for function " << UniqueName
....

Since this problem has not been reproduced in a simple demo, I cannot provide a reproduction example yet. I would like to ask if there are any troubleshooting directions or possible reasons?

2

You can safely ignore these warnings. In case you are curious, the reason you see them is that we don’t reset FDEs for “*.org.*” functions. Those functions will normally contain just a jump to an optimized function without “.org.*” suffix.

3

This doesn’t seem to be consistent with what I’ve observed. In the no-lbr mode, the size of the FDE is different from the size in the symbol table, so I can see this warning, and in the lbr mode, the FDE of the *.ori symbol address cannot be found.

“we don’t reset FDEs for “.org.” functions.”

Does it mean in lbr mode?

4

What is the non-lbr mode? Instrumentation or basic cycle sampling?

5

basic cycle sampling.
Use the --nl option when using perf2bolt

6

Most likely you are getting different profile coverage with different methods of profile collection. Hard to tell exactly without looking at the binaries.

7

Sorry, my previous information was wrong. It has nothing to do with whether it is lbr.
The reason for the difference is:

  1. Using bolt-14, use the option --enable-bat to optimize the binary bolted_binary1
  2. Used bolt-17 perf2bolt to convert perf data from bolted_binary1
    No warning will appear for such steps, but if bolt-17 is used in both steps 1 and 2, a warning will appear.
    Is that because bolt-14 does not set the dwarf table for *.org, and the dwarf table of bolted_binary1 does not have the FDE of *.org.0, so the branch entered when step 2 is executed is 1097 if (FDEI->first != Address ) {.
    For bolt-17, the dwarf table is set and the original size is retained, and the size set in the symbol table is 5, leaving only one jump statement.
8

Now it seems that such logs only appear in bolt-17, so another question is, if it can be ignored, why is it still displayed in the form of a warning by default?

9

Passing BOLTed binaries back to BOLT is something we didn’t expect to be a common practice. Of course with the BAT mode it is expected to happen. It’s a bug.