[Pseudo-ObjC] Render @autoreleasepool blocks

[AngeloD2022]

Translate autorelease pool runtime function calls to @autoreleasepool {...} blocks.

Original Objective-C snippet:

for (NSUInteger itemNo = 1; itemNo <= 3; itemNo++) {
    @autoreleasepool {
        NSString *item = [NSString stringWithFormat:@"batch %lu, item %lu",
                            (unsigned long)batchNo,
                            (unsigned long)itemNo];
        NSString *displayItem = [item uppercaseString];
        [items addObject:displayItem];
    }
}

Pseudo-C snippet:

for (int64_t i = 1; i <= 3; i += 1)
{
    void* context = _objc_autoreleasePoolPush();
    int64_t x0;
    int64_t var_80_1 = x0;
    i_1 = i;
    id location_2 = _objc_retainAutoreleasedReturnValue(_objc_msgSend(
        _OBJC_CLASS_$_NSString, "stringWithFormat:", &cfstr_batch_%lu,_item_%lu));
    int64_t location_1 = _objc_retainAutoreleasedReturnValue(_objc_msgSend(
        location_2, "uppercaseString"));
    _objc_msgSend(var_20, "addObject:", location_1);
    _objc_storeStrong(&location_1, nullptr);
    _objc_storeStrong(&location_2, nullptr);
    _objc_autoreleasePoolPop(context);
}

Equivalent Pseudo-Objective-C:

for (int64_t i = 1; i <= 3; i += 1)
{
    @autoreleasepool
    {
        int64_t x0;
        int64_t var_80_1 = x0;
        i_1 = i;
        id location_2 =
            [[NSString stringWithFormat:@"batch %lu, item %lu"] retain];
        int64_t location_1 = [[location_2 uppercaseString] retain];
        [var_20 addObject:location_1];
        _objc_storeStrong(&location_1, nullptr);
        _objc_storeStrong(&location_2, nullptr);
    }
}

[bdash]

Thanks for the PR! The @autoreleasepool rendering is a nice addition. Factoring the handling of HLIL_BLOCK so it can be overridden by Obj-C rendering seems like generally the right shape.

The autorelease region matching needs some work, though.

The fundamental issue is that TryEmitBlockRegion only ever looks at the immediate children of a single block. Nothing requires that a pool's pops are all siblings of the push. An autorelease pool containing a conditional return, or a conditional break or continue when the pool is inside a loop, will result in multiple pops for a single push, with at least one at a deeper level. Those nested pops are missed by the current approach, so they'll survive into the output despite referencing a variable that is no longer visible:

while (true)
{
    @autoreleasepool
    {
        ...
        if (cond)
        {
            // What is `context` that this refers to?
            _objc_autoreleasePoolPop(context);
            break;
        }
    }
    ...
}

I think the shape you want is to pair pops with pushes using the pool handle rather than by position, and then elide any pop of a handle whose region is currently open, wherever it appears inside that region.

[bdash]

The code within EmitStandardBlockStatement assumes that isLast means the statement is the final statement in the block. This pairs it with the block.exprIndex == root check to detect and suppress a trailing void return. After this refactoring, isLast now only means it's the final statement in the span that was passed to EmitBlockStatements as exprs. PseudoObjCFunction::TryEmitBlockRegion passes a subspan of the block's instructions, which violates that assumption.

I'm not entirely sure how this would manifest in practice, but it makes the code harder to reason about.

[bdash]

This being an output parameter is easy to miss.

[bdash]

This snippet, compiled with cc -fobjc-arc -Os -o test test.m, demonstrates the issue I described with pops that are not siblings of the push:

#import <Foundation/Foundation.h>

int main(int argc, char** argv)
{
    for (id obj in NSProcessInfo.processInfo.environment)
    {
        @autoreleasepool {
            if (!obj)
                break;
            NSLog(@"%@", obj);
        }
    }
}

[AngeloD2022]

Appreciate the review, Mark! I'll investigate these tomorrow as soon as I can.

[AngeloD2022]

I made EmitBlockStatements focus solely on the span of statements provided to it, and moved the void return check up from EmitStandardBlockStatement to the loop inside of EmitBlockStatements. Part of the test happens in GetExpr_BLOCK, where the block instruction is in scope, and is passed through the EmitBlockStatements isBlockRoot parameter.