Upgrade to LLVM 14

[gwenzek]

Relevant LLVM doc: https://llvm.org/docs/OpaquePointers.html

Traditionally, LLVM IR pointer types have contained a pointee type. For example, i32* is a pointer that points to an i32 somewhere in memory. However, due to a lack of pointee type semantics and various issues with having pointee types, there is a desire to remove pointee types from pointers.
The opaque pointer type project aims to replace all pointer types containing pointee types in LLVM with an opaque pointer type. The new pointer type is tentatively represented textually as ptr.

This lead to the deprecation of a number of C API function that now required to be given both the pointer LLVMValueRef and the pointee LLVMTypeRef:

LLVMBuildLoad -> LLVMBuildLoad2
LLVMBuildCall -> LLVMBuildCall2
LLVMBuildInvoke -> LLVMBuildInvoke2
LLVMBuildGEP -> LLVMBuildGEP2
LLVMBuildInBoundsGEP -> LLVMBuildInBoundsGEP2
LLVMBuildStructGEP -> LLVMBuildStructGEP2
LLVMBuildPtrDiff -> LLVMBuildPtrDiff2
LLVMConstGEP -> LLVMConstGEP2
LLVMConstInBoundsGEP -> LLVMConstInBoundsGEP2
LLVMAddAlias -> LLVMAddAlias2
Additionally, it will no longer be possible to call LLVMGetElementType() on a pointer type.

codegen.cpp make extensive use of all the above functions so a number of changes need to happen there.
Trying to compile Zig with LLVM14 currentlly yield 200 warnings and 20 errors which look all related to the above changes.

I've started working on some changes, but I'm not really familiar with codegen.cpp and I'm afraid of making errors and passing the wrong types. But hopefully doing the most of the refactoring myself will allow @andrewrk to focus on the semantic.

[gwenzek]

Preview of ongoing changes: #11374

[gwenzek]

I've hit an issue, which I don't quite understand

zig0 fails to compile stage1:

(gdb) r
Starting program: /home/guw/github/zig-bootstrap/out/build-zig-host/zig0 src/stage1.zig -target native -mcpu=baseline --name zig1 --zig-lib-dir /home/guw/github/zig-bootstrap/zig/lib -femit-bin=/home/guw/github/zig-bootstrap/out/build-zig-host/zig1.o -fcompiler-rt -OReleaseFast --strip -lc --pkg-begin build_options /home/guw/github/zig-bootstrap/out/build-zig-host/config.zig --pkg-end --pkg-begin compiler_rt /home/guw/github/zig-bootstrap/zig/lib/std/special/compiler_rt.zig --pkg-end

zig0: /home/guw/github/zig-bootstrap/out/host/include/llvm/IR/InstrTypes.h:1535: void llvm::CallBase::addParamAttr(unsigned int, llvm::Attribute): Assertion `ArgNo < arg_size() && "Out of bounds"' failed.
Program received signal SIGABRT, Aborted.

(gdb) bt
#0  __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:49
#1  0x00007ffff7a098a4 in __GI_abort () at abort.c:79
#2  0x00007ffff7a09789 in __assert_fail_base (fmt=<optimized out>, assertion=<optimized out>, 
    file=<optimized out>, line=<optimized out>, function=<optimized out>) at assert.c:92
#3  0x00007ffff7a18a16 in __GI___assert_fail (assertion=0x43ac8a8 "ArgNo < arg_size() && \"Out of bounds\"", 
    file=0x43ac808 "/home/guw/github/zig-bootstrap/out/host/include/llvm/IR/InstrTypes.h", line=1535, 
    function=0x43ac860 "void llvm::CallBase::addParamAttr(unsigned int, llvm::Attribute)") at assert.c:101
#4  0x0000000000931ca2 in llvm::CallBase::addParamAttr (this=0x170630450, ArgNo=1, Attr=...)
    at /home/guw/github/zig-bootstrap/out/host/include/llvm/IR/InstrTypes.h:1535
#5  0x000000000092d1e0 in ZigLLVMSetCallSret (Call=0x170630450, return_type=0x16e98e400)
    at /home/guw/github/zig-bootstrap/zig/src/zig_llvm.cpp:1091
#6  0x00000000007ee428 in ir_render_call (g=0x7626690, executable=0x897cdd8, instruction=0x27249e50)
    at /home/guw/github/zig-bootstrap/zig/src/stage1/codegen.cpp:5068
#7  0x00000000007f8022 in ir_render_instruction (g=0x7626690, executable=0x897cdd8, instruction=0x27249e50)
    at /home/guw/github/zig-bootstrap/zig/src/stage1/codegen.cpp:7492
#8  0x00000000007f89f6 in ir_render (g=0x7626690, fn_entry=0x897cd50)
    at /home/guw/github/zig-bootstrap/zig/src/stage1/codegen.cpp:7666
#9  0x00000000007fef96 in do_code_gen (g=0x7626690)
    at /home/guw/github/zig-bootstrap/zig/src/stage1/codegen.cpp:9116
#10 0x00000000008040f8 in codegen_build_object (g=0x7626690)
    at /home/guw/github/zig-bootstrap/zig/src/stage1/codegen.cpp:10354
#11 0x00000000007d5e25 in zig_stage1_build_object (stage1=0x7626690)
    at /home/guw/github/zig-bootstrap/zig/src/stage1/stage1.cpp:132
#12 0x00000000007d22af in main (argc=22, argv=0x7fffffffdb68)
    at /home/guw/github/zig-bootstrap/zig/src/stage1/zig0.cpp:477

This happen when trying to compile the empty_workaround function:

zig/lib/std/target.zig

Lines 648 to 650 in 5466e87

	pub fn empty_workaround() Set {
	return Set{ .ints = [1]usize{0} ** usize_count };
	}

IIUC we are transforming this function of 0 arguments to a function that takes a pointer to return the value, then we are trying to add an attribute to this new parameter, but we crash because we are writing to the attribute array at a non exisiting offset.
To "fix" this I've changed the index in ZigLLVMSetCallSret to 0 instead of 1. Looking at downstream code it look like they are also offsetting the offset, so maybe that could explain the off by one error ? I need to check that the attribute ends up on the correct argument, but this allows me to go further.

void ZigLLVMSetCallSret(LLVMValueRef Call, LLVMTypeRef return_type) {
    CallInst *call_inst = unwrap<CallInst>(Call);
    Type *llvm_type = unwrap<Type>(return_type);
    // fixed to 0 ???
    call_inst->addParamAttr(1, Attribute::getWithStructRetType(call_inst->getContext(), llvm_type));
}

LLVM_NODISCARD AttributeList addParamAttribute(
      LLVMContext &C, unsigned ArgNo, Attribute::AttrKind Kind) const {
    return addAttributeAtIndex(C, ArgNo + FirstArgIndex, Kind);
  }

[gwenzek]

I think my change of attribute argument was indeed correct.
The rest of the code seems to consider the struct return value as being the first argument (see want_first_arg_sret)

I also had to change the corresponding AddSretArg to also modify the first argument.
WIth this changes I can know compile stage1 with zig0. But the resulting zig binary segfaults when trying to build stage2.

I need to run the tests for the stage1 and see what's going on, but enough for today.

[gwenzek]

Relevant video from Nikita Popov Opaque Pointers Are Coming

My own take on the video:

• Typed pointer provide no semantic information (or redundant)
• Typed pointer uses memory
• Typed pointer often requires bitcasting between different typed pointers
• Those bitcasting need to be ignored by the optimizations passes and most of them don't
• Self refrencing struct types become simpler to instantiate for LLVM: struct T { *T } become simply a struct T { ptr }
• Typed pointers are deprecated and "best-effort" starting from LLVM15 and will be removed in LLVM16 (fixed based on nikic comment)

Zig LLVM-backend isn't too much impacted AFAICT because we already store the pointee type in ZIR, we just need to pass the type down to LLVM. The "hard" part is just to pass the correct type in the correct place. And we should also be able to remove some pointer bitcast instructions.

The transition should be smooth:

1. replace all untyped Load/GEP by their corresponding typed version by passing the correct pointee type.
   LLVM will assert that the receive type is the same than the pointer type (and therefore that we didn't change the semantic)
2. replace pointer type creation to opaque pointers (at which point there won't be LLVM asserts)
3. remove pointer to pointer bitcasts

[gwenzek]

Status update, after replacing most LLVMBuildLoad by LLVMBuildLoad2 I was able to run 971 behavior tests.

#11587 prevents me from running the vector related tests: (vector.zig, select.zig, shuffle.zig and math.zig)

There also seems to be some changes to Assembly in LLVM14 preventing me from running asm.zig and forcing me to rewrite x86BitRmw to a simple implementation not using inline assembly.

I also had to replace the call to zig ld.lld by a direct call to ld.lld.

Next steps for me:

• understand the changes to assembly
• finish replacing the calls to the deprecated LLVM C API
• remove unnecessary bitcasts between pointers of different types

Then I'll dig deeper into the ld.lld regression.