Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@00c8e61
patch: llvm/llvm-project#170474
sha256: e1e63bcd3a22680ec5e583dc5a2234f6b77d8c5b28444809d0841633c127399e
commit: cf96440

3 files changed, 144 insertions(+), 164 deletions(-)

Improvements:
  indvars.NumReplaced 69106 -> 69110 +0.01%
  scalar-evolution.NumExitCountsComputed 4044447 -> 4044451 +0.00%
  lcssa.NumLCSSA 15128894 -> 15128902 +0.00%
  gvn.NumGVNSimpl 4415377 -> 4415379 +0.00%
  memdep.NumCacheCompleteNonLocalPtr 5348618 -> 5348620 +0.00%
  scalar-evolution.NumExitCountsNotComputed 11931820 -> 11931824 +0.00%
  memdep.NumCacheNonLocalPtr 266903707 -> 266903739 +0.00%
Regressions:
  licm.NumSunk 268195 -> 268187 -0.00%
  local.NumPHICSEs 186115 -> 186111 -0.00%
  gvn.NumGVNEqProp 427034 -> 427030 -0.00%
  simplifycfg.NumHoistCommonInstrs 2431944 -> 2431936 -0.00%
  instsimplify.NumSimplified 2467336 -> 2467332 -0.00%
  early-cse.NumCSE 5560451 -> 5560447 -0.00%
  instcombine.NumSunkInst 3381567 -> 3381565 -0.00%
  instcombine.NumCombined 120701405 -> 120701377 -0.00%

-10 pola-rs/3j0wo23154dycf1fpnzd17ptx.ll

This LLVM IR diff reflects optimizations targeting loop trip count computation and control flow simplification in the polars_arrow::array::Array::null_count function (and related code). Here are the 5 major changes:

1. Elimination of redundant and i64 %x, 4294967295 masking:
   Three instances of %wide.trip.count*.i.* = and i64 %val, 4294967295 (which masks the lower 32 bits) were removed — at lines 41906, 42085, 42892, and 43058. This indicates the trip count is already known to fit in 32 bits, making the mask unnecessary.

2. Replacement of 64-bit loop exit comparisons with 32-bit equivalents:
   Loop exit conditions (%exitcond*.not.*) were changed from icmp eq i64 %indvars.iv.next*, %wide.trip.count* to icmp eq i32 %trip_limit, %trunc_next, where %trip_limit (e.g., %152, %218, %526, %578) is a pre-existing 32-bit value. This avoids 64-bit arithmetic and enables better optimization (e.g., induction variable elimination, smaller registers).

3. Simplification of loop phi nodes and critical edge handling:
   Several phi nodes in loop exit blocks (e.g., %195, %.lcssa.i.us.pre-phi.i, %.sroa.0.3.lcssa.i) were updated to directly use truncated loop counters (%181, %557, %623) instead of redundant truncations computed in separate basic blocks (e.g., %199, %565). This removes dead code and merges control flow.

4. Removal of dead basic blocks and redundant truncations:
   Blocks like .lr.ph.split.us.i.._crit_edge.loopexit.split.loop.exit35.i.loopexit_crit_edge.us.i and its trunc instruction %565 = trunc ... were deleted, as their logic was folded into predecessor blocks. Similarly, %262 = trunc ... was hoisted and reused, eliminating duplication.

5. Control flow cleanup in loop exit branches:
   A branch target was updated from %default.unreachable269 → %default.unreachable277, and another branch was redirected from %.lr.ph.split.us.i.._crit_edge... to %._crit_edge... (line 42920), reflecting refined loop exit structure after merging paths. Also, an invoke continuation label was renamed from .cont292 → .cont302, suggesting minor CFG restructuring (likely due to inlining or exception-handling optimization).

Overall, these changes reflect aggressive loop canonicalization, induction variable narrowing, and dead code elimination, likely driven by Loop Vectorize or Loop Simplify passes — improving both code size and potential vectorization readiness.

model: qwen-plus-latest
CompletionUsage(completion_tokens=668, prompt_tokens=4941, total_tokens=5609, completion_tokens_details=None, prompt_tokens_details=None)

/close