zstd: Fix corrupted output in "best" (#876)

* zstd: Fix corrupted output in "best"

Regression from #784 and followup #793

Fixes #875

A 0 offset backreference was possible when "improve" was successful twice in a row in the "skipBeginning" part, only finding 2 (previously unmatches) length 4 matches, but where start offset decreased by 2 in both cases.

This would result in output where the end offset would equal to the next 's', thereby doing a self-reference.

Add a general check in "improve" and just reject these. Will also guard against similar issues in the future.

This also hints at some potentially suboptimal hash indexing - but I will take that improvement separately.

Fuzz test set updated.

Author: klauspost

zstd/enc_best.go

@@ -197,12 +197,13 @@ encodeLoop:
 
 		// Set m to a match at offset if it looks like that will improve compression.
 		improve := func(m *match, offset int32, s int32, first uint32, rep int32) {
-			if s-offset >= e.maxMatchOff || load3232(src, offset) != first {
+			delta := s - offset
+			if delta >= e.maxMatchOff || delta <= 0 || load3232(src, offset) != first {
 				return
 			}
 			if debugAsserts {
-				if offset <= 0 {
-					panic(offset)
+				if offset >= s {
+					panic(fmt.Sprintf("offset: %d - s:%d - rep: %d - cur :%d - max: %d", offset, s, rep, e.cur, e.maxMatchOff))
 				}
 				if !bytes.Equal(src[s:s+4], src[offset:offset+4]) {
 					panic(fmt.Sprintf("first match mismatch: %v != %v, first: %08x", src[s:s+4], src[offset:offset+4], first))
@@ -343,8 +344,8 @@ encodeLoop:
 		if best.rep > 0 {
 			var seq seq
 			seq.matchLen = uint32(best.length - zstdMinMatch)
-			if debugAsserts && s <= nextEmit {
-				panic("s <= nextEmit")
+			if debugAsserts && s < nextEmit {
+				panic("s < nextEmit")
 			}
 			addLiterals(&seq, best.s)