resolve_btfids: Match the arm64 NOP opcode for LLVM 24 in text_mc
PATCH 1/1 resolve_btfids: Match the arm64 NOP opcode for LLVM 24 in text_mc
PR #170968, “[AArch64] Treat NOP as a separate instruction” [1] gave NOP its own instruction definition; LLVM 24 decodes "1f 20 03 d5" as the opcode NOP. Earlier releases decode it as HINT #0, which is what the arm64 table looks for. So the test of the nop fails, text_arch_init() gives up and a kernel with CONFIG_BPF_FMODRET_SCAN=y cannot be built:
text_mc: LLVM 24.0.0git: 1f 20 03 d5 is NOP, length 4, nop 0 ...
expected: nop 1 ...
text_mc: LLVM 24.0.0git: the table of the arch does not fit it. ...
FAILED: no decoder for the code of ELF machine 183
Add a pattern for NOP. An LLVM before this has no opcode of that name, and a name which is missing is an error. So a pattern can now be
OPTIONAL: it is left out if this LLVM lacks the opcode. The pattern of HINT #0 stays for the releases before, and the test shows that one of the two matches.
With LLVM 24.0.0git the scan of an arm64 vmlinux finds no instruction it cannot decode: 3385 of 65138 functions are tagged (the smaller number of matches is due to CONFIG_KCSAN causing functions to be marked as unsafe for fmod_ret).
[1] https://github.com/llvm/llvm-project/pull/170968
--- tools/bpf/resolve_btfids/text_mc.cpp | 2 ++ tools/bpf/resolve_btfids/text_mc.h | 6 ++++++ tools/bpf/resolve_btfids/text_mc_arm64.c | 2 ++ 3 files changed, 10 insertions(+)
tools/bpf/resolve_btfids/text_mc.cpp +2 -0
@@ -195,6 +195,8 @@ bool read_opcodes() auto it = by_name.find(pattern->opcode); if (it == by_name.end()) { + if (pattern->optional) + continue; error() << "no opcode " << pattern->opcode << "\n"; ok = false; } else if (pattern->kind == TEXT_RETURN) {
tools/bpf/resolve_btfids/text_mc.h +6 -0
@@ -50,6 +50,11 @@ struct mc_predicate {* has, and DEFS. * With NOP the instruction is a nop if the first predicate holds. * +* An opcode which this LLVM does not have is an error, unless the pattern is +* OPTIONAL: then it is left out. That is for an opcode which LLVM split off +* from another one in some release, while the pattern of the old one stays +* for the releases before. The tests show that one of the two matches. +* * A pattern of kind TEXT_RETURN names a return to the caller. isReturn() holds * for a return from an exception as well, which does not hand the return value * register to a caller. Without such a pattern it is an opcode like any other,
@@ -71,6 +76,7 @@ struct mc_pattern {s64 value;u32 defs;/* of TEXT_OTHER */ bool nop;/* of TEXT_OTHER */ +bool optional;/* the opcode is missing in some LLVM releases */ struct mc_predicate predicates[2 ]; };
tools/bpf/resolve_btfids/text_mc_arm64.c +2 -0
@@ -25,6 +25,8 @@ static const struct mc_pattern arm64_patterns[] = {/* HINT: nop, bti, yield, wfe ... */ { "HINT", TEXT_OTHER }, { "AUTIASP", TEXT_OTHER }, { "PRFMui", TEXT_OTHER }, { "HINT", TEXT_OTHER, .nop = true, .predicates = { IMM(0, 0) } }, /* nop */ +/* LLVM 24 decodes nop as NOP, not as HINT #0 */ + { "NOP", TEXT_OTHER, .nop = true, .optional = true }, { "DMB", TEXT_OTHER }, { "MRS", TEXT_OTHER }, { "MSR", TEXT_OTHER }, { "STLRB", TEXT_OTHER }, { "STLRW", TEXT_OTHER }, WX("LDXR", "", TEXT_OTHER), WX("STXR", "", TEXT_OTHER), WX("STLXR", "", TEXT_OTHER),
Signature · 1 line
2.56.0
[email protected]
AI reviewTL;DR: Commit message and code comment may name the wrong LLVM version: LLVM 22 and 23 seem to decode arm64 nop as NOP too, so "LLVM 24" should likely read "LLVM 22 and later".
Quoted 1 line
resolve_btfids: Match the arm64 NOP opcode for LLVM 24 intext_mc
TL;DR: Commit message and code comment may name the wrong LLVM version: LLVM 22 and 23 seem to decode arm64 nop as NOP too, so "LLVM 24" should likely read "LLVM 22 and later".
Quoted 6 lines
resolve_btfids: Match the arm64 NOP opcode for LLVM 24 intext_mcPR #170968, "[AArch64] Treat NOP as a separate instruction" [1] gave NOP its own instruction definition; LLVM 24 decodes "1f 20 03 d5" as the opcode NOP. Earlier releases decode it as HINT #0, which is what the arm64 table looks for.
Is this boundary accurate? That LLVM change was merged to trunk when it was 22.0.0git, and the release/22.x AArch64InstrInfo.td already has:
def NOP : SystemNoOperands<0b000, "hint\t#0">;
def : InstAlias<"nop", (NOP)>;So it looks like the released LLVM 22 and 23 decode nop as NOP as well, and only LLVM 17 to 21 decode it as HINT #0. An arm64 build with CONFIG_BPF_FMODRET_SCAN=y would then fail with LLVM 22 and 23 in the same way as with 24.0.0git.
Could the subject and body say "LLVM 22 and later" instead of LLVM 24?
Quoted 3 lines
tools/bpf/resolve_btfids/text_mc_arm64.c +0
-0
[ ... ]
tools/bpf/resolve_btfids/text_mc_arm64.c static const struct mc_pattern arm64_patterns[] = { · line 28
Quoted 5 lines
tools/bpf/resolve_btfids/text_mc_arm64.c +1
-0@@ -25,6 +25,8 @@ static const struct mc_pattern arm64_patterns[] = {/* HINT: nop, bti, yield, wfe ... */ { "HINT", TEXT_OTHER }, { "AUTIASP", TEXT_OTHER }, { "PRFMui", TEXT_OTHER }, { "HINT", TEXT_OTHER, .nop = true, .predicates = { IMM(0, 0) } }, /* nop */ + /* LLVM 24 decodes nop as NOP, not as HINT #0 */
This isn't a bug, but the same version boundary applies to this comment. Should it say LLVM 22 and later rather than LLVM 24?
tools/bpf/resolve_btfids/text_mc_arm64.c static const struct mc_pattern arm64_patterns[] = { · line 30
Quoted 2 lines
tools/bpf/resolve_btfids/text_mc_arm64.c +1
-0+ { "NOP", TEXT_OTHER, .nop = true, .optional = true }, { "DMB", TEXT_OTHER }, { "MRS", TEXT_OTHER }, { "MSR", TEXT_OTHER },
---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md
CI run summary: https://github.com/kernel-patches/bpf/actions/runs/44386217738
Review
Quoted 5 lines · Alan Maguire
(quoted diff) +2
-0if (it == by_name.end()) { + if (pattern->optional) + continue; error() << "no opcode " << pattern->opcode << "\n"; ok = false;
Thanks for the quick fix. Applied.
But should we go further and skip unknown names? Like:
tools/bpf/resolve_btfids/text_mc.cpp +9 -8
@@ -165,10 +165,9 @@ bool read_regs() return ok; } -bool read_opcodes() +void read_opcodes() { std::unordered_map<std::string_view, Opcode *> by_name; - bool ok = true; d.opcodes.resize(d.MII->getNumOpcodes()); by_name.reserve(d.opcodes.size());
@@ -194,10 +193,9 @@ bool read_opcodes() for (const mc_pattern *pattern = mc_arch.patterns; pattern->opcode; pattern++) { auto it = by_name.find(pattern->opcode); - if (it == by_name.end()) { - error() << "no opcode " << pattern->opcode << "\n"; - ok = false; - } else if (pattern->kind == TEXT_RETURN) { + if (it == by_name.end()) + continue; + if (pattern->kind == TEXT_RETURN) { it->second->is_return = true; } else if (pattern->kind != TEXT_OTHER) { it->second->patterns.push_back(pattern);
@@ -208,7 +206,6 @@ bool read_opcodes() it->second->nop = pattern; } } - return ok; } const Reg ®_of(const MCInst &inst, const Opcode &o, u8 nr, bool def)
@@ -422,12 +419,16 @@ bool run_tests() extern "C" const struct text_arch *text_arch_init(unsigned int machine) { static text_arch arch = mc_arch.arch; + bool ok; if (machine != mc_arch.machine) { errs() << "text_mc: built for " << mc_arch.triple << ", not for ELF machine " << machine << "\n"; return nullptr; } - if (!create() || !read_regs() || !read_opcodes() || !run_tests()) { + ok = create() && read_regs(); + if (ok) + read_opcodes(); + if (!ok || !run_tests()) { error() << "the table of the arch does not fit it. Set LLVM_CONFIG to the " "llvm-config of another LLVM, or switch CONFIG_BPF_FMODRET_SCAN off\n"; return nullptr;
tools/bpf/resolve_btfids/text_mc.h +3 -2
@@ -13,8 +13,9 @@* It does not say what an instruction computes: the selection patterns are not * a part of it. And the headers with the enums of a target are not installed. * So opcodes and registers are given by name, as MCInstrInfo::getName() and -* MCRegisterInfo::getName() return them, and looked up once. A name which this -* LLVM does not have is an error. +* MCRegisterInfo::getName() return them, and looked up once. A register which +* this LLVM does not have is an error. A pattern for an opcode which it does +* not have is left out: each pattern has a test, which fails if it is missed. */ /* A condition on the operands of an MCInst, like an MCInstPredicate of TableGen */
tools/bpf/resolve_btfids/text_mc_arm64.c +1 -0
@@ -25,6 +25,7 @@ static const struct mc_pattern arm64_patterns[] = {/* HINT: nop, bti, yield, wfe ... */ { "HINT", TEXT_OTHER }, { "AUTIASP", TEXT_OTHER }, { "PRFMui", TEXT_OTHER }, { "HINT", TEXT_OTHER, .nop = true, .predicates = { IMM(0, 0) } }, /* nop */ + { "NOP", TEXT_OTHER, .nop = true }, /* nop since llvm-project PR 170968 */ { "DMB", TEXT_OTHER }, { "MRS", TEXT_OTHER }, { "MSR", TEXT_OTHER }, { "STLRB", TEXT_OTHER }, { "STLRW", TEXT_OTHER }, WX("LDXR", "", TEXT_OTHER), WX("STXR", "", TEXT_OTHER), WX("STLXR", "", TEXT_OTHER),
This patch was applied to bpf/bpf-next.git (master) by Alexei Starovoitov <[email protected]>:
Hello:
This patch was applied to bpf/bpf-next.git (master) by Alexei Starovoitov <[email protected]>:
Quoted 8 lines · On Wed, 18 Nov 2026 19:40:30 +0000 you wrote:
PR #170968, “[AArch64] Treat NOP as a separate instruction” [1] gave NOP its own instruction definition; LLVM 24 decodes "1f 20 03 d5" as the opcode NOP. Earlier releases decode it as HINT #0, which is what the arm64 table looks for. So the test of the nop fails,
text_arch_init()gives up and a kernel withCONFIG_BPF_FMODRET_SCAN=ycannot be built:[...]
Here is the summary with links:
- [bpf-next] resolve_btfids: Match the arm64 NOP opcode for LLVM 24 in text_mc
https://git.kernel.org/bpf/bpf-next/c/f8c00958c5b9
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html