llvm-project
e38f5645 - [CIR] Implement PromotableRegionOpInterface for `cir.if`,`cir.scope` and `cir.ternary` (#215780)

Commit
9 hours ago
[CIR] Implement PromotableRegionOpInterface for `cir.if`,`cir.scope` and `cir.ternary` (#215780) mem2reg gives up on a slot if any use lives in a nested region whose parent op doesn't implement PromotableRegionOpInterface. No CIR op implements it. So this doesn't promote: ```mlir %a = cir.alloca "a" align(4) : !cir.ptr<!s32i> %c = cir.const #cir.int<42> : !s32i cir.store %c, %a : !s32i, !cir.ptr<!s32i> cir.if %cond { %v = cir.load %a : !cir.ptr<!s32i>, !s32i cir.call @use(%v) : (!s32i) -> () } ``` You have to run -cir-flatten-cfg first, which is why mem2reg.cir does. That's a lot of destruction for a value that never leaves the region. This implements the interface for cir.if and cir.scope. Both regions are entered straight from before the op, so they see the reaching definition unchanged. Nothing to merge, nothing to yield. Stores inside a region are refused. A new definition has to exit through a result, cir.if has none, cir.scope has one it may already be using. Supporting that means giving those ops results. Separate patch. My use case specifically: Downstream, I'm propagating constants into CUDA kernel arguments.[ Every launch site sits inside a cir.if, the __cudaPushCallConfiguration guard](https://github.com/llvm/llvm-project/blob/5bf59e2c4b54a85e4e7f0e188b99061beb2708f6/clang/lib/CIR/CodeGen/CIRGenCUDARuntime.cpp#L33), so the argument reaching the device stub call site is always a cir.load. In particular this allows other MLIR passes like sccp, determine if such value is in constant form through the call graph. So the argument flow would be like this: mem2reg turns it into the block argument, sccp turns that into a cir.const, and you get the constant at the call site. Claude assisted with this patch. --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Author
Parents
Loading