[MC][Mips] Fix wrong assumption about Immediate operand.

[romainthomas]

While trying to evaluate the branch of this MIPS (EL) non-branch instruction: ldxc1 $f2, $4($7) the function fails on the assert: assert(isImm() && ...).

This commit ensures that the operand is an immediate before accessing the value.

• Triple: mipsel-unknown-linux-gnu-elf with +mips32r2
• Instruction: ldxc1 $f2, $4($7)
• Raw instruction: 0x81, 0x0, 0xe4, 0x4c

[github-actions]

⚠️ C/C++ code formatter, clang-format found issues in your code. ⚠️

You can test this locally with the following command:

git-clang-format --diff a9eb8f0e3dbaf16b6bd83eecb960b6ea8ecaa8c3 d0b9d784e2e1257eeab6c8a52f4845a10ae151f6 --extensions cpp -- llvm/lib/Target/Mips/MCTargetDesc/MipsMCTargetDesc.cpp

View the diff from clang-format here.

diff --git a/llvm/lib/Target/Mips/MCTargetDesc/MipsMCTargetDesc.cpp b/llvm/lib/Target/Mips/MCTargetDesc/MipsMCTargetDesc.cpp
index 4b38ce70c3..eaf4de9d37 100644
--- a/llvm/lib/Target/Mips/MCTargetDesc/MipsMCTargetDesc.cpp
+++ b/llvm/lib/Target/Mips/MCTargetDesc/MipsMCTargetDesc.cpp
@@ -143,7 +143,7 @@ public:
     switch (Info->get(Inst.getOpcode()).operands()[NumOps - 1].OperandType) {
     case MCOI::OPERAND_UNKNOWN:
     case MCOI::OPERAND_IMMEDIATE: {
-      const MCOperand& Op = Inst.getOperand(NumOps - 1);
+      const MCOperand &Op = Inst.getOperand(NumOps - 1);
       if (!Op.isImm())
         return false;
       // j, jal, jalx, jals

[brad0]

cc @yingopq @wzssyqa

[nikic]

Test?

[romainthomas]

From what I understand, most of the logic of Mips disassembler testing is done in llvm/test/MC/Disassembler/Mips/. There is also some disassembler unittests in llvm/unittests/MC but not for the Mips Target (only AMDGPU, X86, SystemZ).

I don't think we could exercise this code path with a lit test¹ using llvm-mc so a unittest would be more appropriated. But since there is no unittest for Mips it would require bootstrapping them.

Who should own this bootstrapping? Me in this PR? Me in a different PR? The Mips code owners?

Footnotes

1. I can be completely wrong on this point ↩

[s-barannikov]

I don't think we could exercise this code path with a lit test using llvm-mc

I think you can do that using llvm-objdump --symbolize-operands.

[wzssyqa]

Can you give out a test case for it?

[romainthomas]

Feel free to re-open this PR if you think it is relevant and when there is a clear status about Mips testing in LLVM (c.f. #119056 (comment))