fix: sanitize pprof references

[kolesnikovae]

Resolves #3203

[kolesnikovae]

I had to remove the test as it's no longer valid: invalid elements are discarded at normalization/merge

[kolesnikovae]

Now we're doing this at reference sanitisation

[kolesnikovae]

The only reason we do it here is that Merge might be called without Normalize. It is a fairly expensive operation, but it's better than a failure. Ideally, normalization and sanity checks should be performed at parsing

[simonswine]

Generally LGTM.

A few things that a simple fuzzing session #3222 uncovered (not originating from this PR):

--- a/pkg/pprof/merge.go
+++ b/pkg/pprof/merge.go
@@ -196,6 +196,9 @@ func combineHeaders(a, b *profilev1.Profile) error {
 // returns nil if the profiles are compatible; otherwise an error with
 // details on the incompatibility.
 func compatible(a, b *profilev1.Profile) error {
+       if a == nil || b == nil {
+               return fmt.Errorf("a nil profile was specified")
+       }
        if !equalValueType(a.PeriodType, b.PeriodType) {
                return fmt.Errorf("incompatible period types %v and %v", a.PeriodType, b.PeriodType)
        }
@@ -213,6 +216,12 @@ func compatible(a, b *profilev1.Profile) error {
 // equalValueType returns true if the two value types are semantically
 // equal. It ignores the internal fields used during encode/decode.
 func equalValueType(st1, st2 *profilev1.ValueType) bool {
+       if st1 == nil && st2 == nil {
+               return true
+       }
+       if st1 == nil || st2 == nil {
+               return false
+       }
        return st1.Type == st2.Type && st1.Unit == st2.Unit
 }

@@ -242,11 +251,13 @@ func RewriteStrings(p *profilev1.Profile, n []uint32) {
        }
        p.DropFrames = int64(n[p.DropFrames])
        p.KeepFrames = int64(n[p.KeepFrames])
-       if p.PeriodType.Type != 0 {
-               p.PeriodType.Type = int64(n[p.PeriodType.Type])
-       }
-       if p.PeriodType.Unit != 0 {
-               p.PeriodType.Unit = int64(n[p.PeriodType.Unit])
+       if p.PeriodType != nil {
+               if p.PeriodType.Type != 0 {
+                       p.PeriodType.Type = int64(n[p.PeriodType.Type])
+               }
+               if p.PeriodType.Unit != 0 {
+                       p.PeriodType.Unit = int64(n[p.PeriodType.Unit])
+               }
        }
        for i, x := range p.Comment {
                p.Comment[i] = int64(n[x])

[kolesnikovae]

I see that we actually do not check if an element is nil (e.g. nil sample) at all – to be added in this PR