[llvm-bugs] [Bug 51820] New: Improve TLV performance in ORC Runtime

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=51820

            Bug ID: 51820
           Summary: Improve TLV performance in ORC Runtime
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: OrcJIT
          Assignee: unassignedbugs at nondot.org
          Reporter: lhames at gmail.com
                CC: 1101.debian at gmail.com, llvm-bugs at lists.llvm.org

For convenience, the initial TLV support in the ORC runtime (for both
macho_platform and elfnix_platform) just saves all registers, then jumps into a
function to do table lookup for the requested TLV instance address. I think the
following scheme should provide better performance, while still allowing us to
add new TLVs at runtime:

1. Instead of pointing directly at a TLV manager object the thread data pointer
for the current thread should point to a table structure containing a N + 2
pointer sized objects. The first two entries will contain a pointer to the TLV
manager object and a table size, and all subsequent entries will contain
addresses of TLV instances for the current thread. As new TLVs are added to the
JIT they are assigned indexes within this table (these indexes may extend past
the end of the currently allocated table).

```
tlv_table:
  <pointer to TLV manager>
  <number of elements N>
  <element 0>
  ...
  <element N - 1>
```

2. TLV descriptors (in the current MachO parlance) should become a triple of
(get_addr_fn, key, index)  (instead of the current (get_addr_fn, key, addr)),
and...

3. The new tlv_get_addr function should look something like this:

```
if (desc->idx > table->size)
  return expandTableAndGet(desc);
else if (table->elements[desc->idx] == 0)
  return allocateElementAndGet(desc);
return table->elements[desc->idx];
```
Or, in x86-64 assembly, something like:

```
_tlv_get_addr:
        movq    8(%rdi),%rax                    // get key from desc            
        movq    %gs:0x0(,%rax,8),%rsi           // get table for thread         
        movq    16(%rdi), %rax                  // get index from desc
        cmpq    8(%rsi), %rax                   // compare index and table size
        jae     LexpandTableAndGet              // if index > size then expand
table
        movq    16(%rsi,%rax,8), %rax.          // otherwise get table entry
        testq   %rax, %rax                      // check for a null entry
        je      LallocateElementAndGet          // if null then allocate entry 
        ret                                     // Non-null entry: we have our
result.
LexpandTableAndGet
        jmp __resize_table
LallocateElementAndGet
        jmp __allocate_element
```

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210911/f49fead2/attachment.html>