[flang] Fix UTF-8 minimality checks

[klausler]

UTF-8 encodings are required to be minimal, but the checks for minimality of 3-byte and 4-byte sequences were incorrect. Fix.

[llvmbot]

@llvm/pr-subscribers-flang-parser

Author: Peter Klausler (klausler)

Changes

UTF-8 encodings are required to be minimal, but the checks for minimality of 3-byte and 4-byte sequences were incorrect. Fix.

---

Full diff: https://github.com/llvm/llvm-project/pull/159142.diff

2 Files Affected:

• (modified) flang/lib/Parser/characters.cpp (+9-6)
• (added) flang/test/Parser/utf8-01.f90 (+15)

diff --git a/flang/lib/Parser/characters.cpp b/flang/lib/Parser/characters.cpp
index 1a00b16eefe9d..69b6d2ed5fafb 100644
--- a/flang/lib/Parser/characters.cpp
+++ b/flang/lib/Parser/characters.cpp
@@ -158,21 +158,24 @@ DecodedCharacter DecodeRawCharacter<Encoding::UTF_8>(
     const char *cp, std::size_t bytes) {
   auto p{reinterpret_cast<const std::uint8_t *>(cp)};
   char32_t ch{*p};
-  if (ch <= 0x7f) {
+  // Valid UTF-8 encodings must be minimal.
+  if (ch <= 0x7f) { // 1 byte: 7 bits of payload
     return {ch, 1};
-  } else if ((ch & 0xf8) == 0xf0 && bytes >= 4 && ch > 0xf0 &&
-      ((p[1] | p[2] | p[3]) & 0xc0) == 0x80) {
+  } else if ((ch & 0xf8) == 0xf0 && bytes >= 4 &&
+      ((p[1] | p[2] | p[3]) & 0xc0) == 0x80 && (ch > 0xf0 || p[1] > 0x8f)) {
+    // 4 bytes: 3+6+6+6=21 bits of payload
     ch = ((ch & 7) << 6) | (p[1] & 0x3f);
     ch = (ch << 6) | (p[2] & 0x3f);
     ch = (ch << 6) | (p[3] & 0x3f);
     return {ch, 4};
-  } else if ((ch & 0xf0) == 0xe0 && bytes >= 3 && ch > 0xe0 &&
-      ((p[1] | p[2]) & 0xc0) == 0x80) {
+  } else if ((ch & 0xf0) == 0xe0 && bytes >= 3 &&
+      ((p[1] | p[2]) & 0xc0) == 0x80 && (ch > 0xe0 || p[1] > 0x9f)) {
+    // 3 bytes: 4+6+6=16 bits of payload
     ch = ((ch & 0xf) << 6) | (p[1] & 0x3f);
     ch = (ch << 6) | (p[2] & 0x3f);
     return {ch, 3};
   } else if ((ch & 0xe0) == 0xc0 && bytes >= 2 && ch > 0xc0 &&
-      (p[1] & 0xc0) == 0x80) {
+      (p[1] & 0xc0) == 0x80) { // 2 bytes: 5+6=11 bits of payload
     ch = ((ch & 0x1f) << 6) | (p[1] & 0x3f);
     return {ch, 2};
   } else {
diff --git a/flang/test/Parser/utf8-01.f90 b/flang/test/Parser/utf8-01.f90
new file mode 100644
index 0000000000000..3a3745524d807
--- /dev/null
+++ b/flang/test/Parser/utf8-01.f90
@@ -0,0 +1,15 @@
+!RUN: %flang_fc1 -fdebug-unparse %s 2>&1 | FileCheck %s
+
+character(kind=4), parameter :: c(2) = [character(kind=4) :: &
+4_'🍌', 4_'水' ]
+print *, '🍌'
+print *, 4_'🍌'
+print *, '水'
+print *, 4_'水'
+end
+
+!CHECK: CHARACTER(KIND=4_4), PARAMETER :: c(2_4) = [CHARACTER(KIND=4,LEN=1)::4_"\360\237\215\214",4_"\346\260\264"]
+!CHECK: PRINT *, "\360\237\215\214"
+!CHECK: PRINT *, 4_"\360\237\215\214"
+!CHECK: PRINT *, "\346\260\264"
+!CHECK: PRINT *, 4_"\346\260\264"

[eugeneepshteyn]

Where do 0x8f and 0x9f come from?

[klausler]

The second and later bytes of a UTF-8 encoding sequence are always values in the range 0x80 to 0xbf. Their least significant six bits carry their 6-bit payloads. A four-byte UTF-8 sequence has 21 bits of payload: three bits in the first byte, and six bits in each of the next three bytes. Since a 3-byte UTF-8 sequence has 16 bits of payload, a minimal four-byte UTF-8 sequence must have at least one of its most significant five (21-16) payload bits set. These five most significant payload bits are encoded as the three least significant bits of the first byte and as the 0x20 and 0x10 bits of the second byte. So either the first byte (in the range 0xf0-0xf7) has to be greater than 0xf0, or the second byte (in the range 0x80-0xbf) has to be greater than 0x8f.