Fix MIPS PIC level and work around an LLVM bug for mips(el)-linux-gnueabi(hf)

[alexrp]

• Set the correct PIC level for MIPS.
• Work around Re-enable FastISel for MIPS O32 with LLVM 20 #21215 until [llvm][Mips] Bail on underaligned loads/stores in FastISel. llvm/llvm-project#106231 trickles down.
• Revert the workaround for struct field load through underaligned pointer has incorrect alignment #14904.

This makes zig build test-modules -fqemu --glibc-runtimes <...> -Dtest-slow-targets -Dtest-target-filter=mips-linux.4.19...6.10.3-gnueabi -Dtest-target-filter=mipsel-linux.4.19...6.10.3-gnueabi mostly work, save for an fchmodat test failure that I still need to investigate.

[The-King-of-Toasters]

Can you share the fchmodat failure here (or in another issue)? I wrote that code so I can give my 2¢.

[alexrp]

I don't recall the exact failure, but it was one of the regfile asserts here:

zig/lib/std/posix/test.zig

Lines 1273 to 1289 in 1a178d4

	try posix.fchmodat(tmp.dir.fd, "regfile", 0o640, 0);
	try expectMode(tmp.dir.fd, "regfile", 0o640);
	try posix.fchmodat(tmp.dir.fd, "regfile", 0o600, posix.AT.SYMLINK_NOFOLLOW);
	try expectMode(tmp.dir.fd, "regfile", 0o600);
	try posix.fchmodat(tmp.dir.fd, "symlink", 0o640, 0);
	try expectMode(tmp.dir.fd, "regfile", 0o640);
	try expectMode(tmp.dir.fd, "symlink", sym_mode);
	var test_link = true;
	posix.fchmodat(tmp.dir.fd, "symlink", 0o600, posix.AT.SYMLINK_NOFOLLOW) catch |err| switch (err) {
	error.OperationNotSupported => test_link = false,
	else => |e| return e,
	};
	if (test_link)
	try expectMode(tmp.dir.fd, "symlink", 0o600);
	try expectMode(tmp.dir.fd, "regfile", 0o640);

[jacobly0]

This API usage should not exist, we already have the ability to modify the module freely.

[alexrp]

As I understand it, similar to float ABI, you can't set a code model on a bitcode module. So for these, we're stuck with the API (or llc later).

PIC/PIE are just module flags though, so I can look into emitting the module flags on our side in a follow-up PR if that's ok.

[jacobly0]

As I understand it, similar to float ABI, you can't set a code model on a bitcode module.

Correct for float ABI, but not for code model:

!7 = !{i32 1, !"Code Model", i32 1}

Generally anything being set on llvm::Module is part of the module, anything just being passed to codegen is not.

[alexrp]

Filed #21238 so this isn't forgotten.

[alexrp]

I don't recall the exact failure, but it was one of the regfile asserts here:

@The-King-of-Toasters

test-modules
└─ test-std
   └─ run test std-mips-linux.4.19...6.10.3-gnueabihf.2.28-mips32-Debug-libc 2706/2792 passed, 1 failed, 85 skipped
error: 'posix.test.test.fchmodat smoke test' failed: expected 416, found 488
/home/alexrp/Source/zig/lib/std/testing.zig:97:17: 0xcac51b in expectEqualInner__anon_40880 (test)
                return error.TestExpectedEqual;
                ^
/home/alexrp/Source/zig/lib/std/posix/test.zig:1250:5: 0x25e90e7 in expectMode (test)
    try expectEqual(mode, st.mode & 0b111_111_111);
    ^
/home/alexrp/Source/zig/lib/std/posix/test.zig:1274:5: 0x25e973b in test.fchmodat smoke test (test)
    try expectMode(tmp.dir.fd, "regfile", 0o640);
    ^
error: while executing test 'zig.system.darwin.macos.test.detect', the following test command failed:
qemu-mips -L /opt/glibc/mips-linux-gnueabihf /home/alexrp/Source/zig/.zig-cache/o/a5712f20e8c3b4c16897214a501633b6/test --seed=0x55598d5a --cache-dir=/home/alexrp/Source/zig/.zig-cache --listen=-