LGTM, thanks.

LGTM.

Thanks for addressing all the comments! Looks good to me now except for a couple minor things.

A general question about the design as I read more into the code. I was trying to figure out if a FunctionSample, its SampleContext and ContextTrieNode stick together as a whole, as also commented in another place. It seems to me multiple FunctionSamples can share the ContextTrieNode. Am I correct? What is the reason for this?

Updating D121651: [llvm-profdata] Convert nested profile to CS flat profile.

LGTM.

LGTM, thanks for addressing the feedbacks!

Thanks for sharing more about the context and the background, @kazu

Updating D121651: [llvm-profdata] Convert nested profile to CS flat profile.

Updating D121651: [llvm-profdata] Convert nested profile to CS flat profile.

Updating D126434: [PseudoProbe] Use callee name as callsite identfier for MCDecodedPseudoProbeInlineTree.

Would be good to have some numbers for build time and memory usage in the summary.

Rebasing.

Performance testing came back neutral, with a 99.962% profile similarity.

Updating D125266: [CSSPGO][CSProfileConverter] Remove call target samples when including callee samples into caller.

Updating D122609: [llvm-profgen] Update callsite body samples by summing up all call target samples.

Updating D122609: [llvm-profgen] Update callsite body samples by summing up all call target samples.

Updating D122609: [llvm-profgen] An option to update callsite body samples by summing up all call target samples.

Updating D125266: [CSSPGO][CSProfileConverter] Remove call target samples when including callee samples into caller.

Replaced by D125448: [llvm-profgen] Filter out oversized LBR ranges..

Updating D123361: [llvm-profgen] Filter out oversized LBR ranges.

In D125246#3504823, @hoy wrote:

In D125246#3502570, @wlei wrote:

In D125246#3501359, @hoy wrote:

Our experiment on one internal service showed that the context key's memory can be reduced from 80GB to 300MB.

This is awesome! Thanks for the work!

How about running time?

Just tested the trie based is faster than the old one, E2E running time : 103mins vs 122mins.

The reason for the speed-up should be due to the time spent on getting caller's FunctionSamples. For the new version, it's just call getParentContext() which is O(1). The old version it generate the caller's context frames then call getFunctionProfileForContext(...), which is O(n), n is the context length.

Very nice. BTW, looks like this is done to line number based profile generation only. How about pseudo probes? Am I missing anything?

In D125246#3502570, @wlei wrote:

In D125246#3501359, @hoy wrote:

Our experiment on one internal service showed that the context key's memory can be reduced from 80GB to 300MB.

This is awesome! Thanks for the work!

How about running time?

Just tested the trie based is faster than the old one, E2E running time : 103mins vs 122mins.

The reason for the speed-up should be due to the time spent on getting caller's FunctionSamples. For the new version, it's just call getParentContext() which is O(1). The old version it generate the caller's context frames then call getFunctionProfileForContext(...), which is O(n), n is the context length.

LGTM with a minor suggestion.

In D125255#3502157, @wenlei wrote:

I'm not sure if a universal threshold for filtering warning is a good way to deal with this. Maybe look at this from a different angle -- for some warnings, any instance is bad, for others, we can tolerate them. Hence, an alternative: for critical ones, we should always print out any non-zero value; for others, print them only under a "verbose" mode?

Updating D125266: [CSSPGO][CSProfileConverter] Remove call target samples when including callee samples into caller.

Extending test to check scaling.

In D125249#3501454, @davidxl wrote:

Should we have a test case to cover the post-inlining scaling. Otherwise looks good.

In D125249#3501442, @davidxl wrote:

Does the data need to be scaled first?

Our experiment on one internal service showed that the context key's memory can be reduced from 80GB to 300MB.