pre-commit: PR143208

[zyw-bot]

Link: llvm/llvm-project#143208
Requested by: @artagnon

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@01d648a
patch: llvm/llvm-project#143208
sha256: 787218a49f4457bcdf159165c1301adfa260f6997e0296e7ddd65d8f7b5c4026
commit: 16c678d

7 files changed, 2521 insertions(+), 2752 deletions(-)

Improvements:
  indvars.NumElimIdentity 1885 -> 1888 +0.16%
  indvars.NumFoldedUser 2202 -> 2204 +0.09%
  constmerge.NumIdenticalMerged 15663 -> 15671 +0.05%
  loop-delete.NumBackedgesBroken 46391 -> 46405 +0.03%
  indvars.NumElimCmp 57987 -> 58001 +0.02%
  loop-idiom.NumMemCpy 9706 -> 9708 +0.02%
  gvn.NumGVNBlocks 200845 -> 200875 +0.01%
  globalsmodref-aa.NumNonAddrTakenGlobalVars 428951 -> 428965 +0.00%
  loop-delete.NumDeleted 120051 -> 120053 +0.00%
  indvars.NumElimExt 310098 -> 310101 +0.00%
Regressions:
  indvars.NumLFTR 337255 -> 337251 -0.00%
  instsimplify.NumSimplified 2622166 -> 2622148 -0.00%
  gvn.IsValueFullyAvailableInBlockNumSpeculationsMax 610372 -> 610368 -0.00%
  lcssa.NumLCSSA 16136379 -> 16136275 -0.00%
  gvn.NumGVNSimpl 4741634 -> 4741611 -0.00%
  scalar-evolution.NumExitCountsComputed 4284860 -> 4284843 -0.00%
  loop-vectorize.LoopsAnalyzed 2062807 -> 2062805 -0.00%
  memdep.NumUncacheNonLocalPtr 269995956 -> 269995823 -0.00%
  sccp.NumInstRemoved 2106735 -> 2106734 -0.00%
  memdep.NumCacheNonLocalPtr 284061609 -> 284061513 -0.00%

34 56 bench/clamav/optimized/crc.ll
21 32 bench/hdf5/optimized/H5checksum.ll
383 467 bench/linux/optimized/i2c-core-smbus.ll
46 106 bench/openmpi/optimized/crc.ll
53 77 bench/wireshark/optimized/packet-mstp.ll

[github-actions]

The provided LLVM IR diff introduces several changes across multiple benchmark files related to CRC (Cyclic Redundancy Check) table initialization and computation. Below is a high-level summary of the up to 5 major changes, ignoring formatting, comments, and reordering:

---

1. Introduction of Precomputed CRC Tables as Private Constants

Across all modified files (crc.ll, H5checksum.ll, i2c-core-smbus.ll, packet-mstp.ll), new private global constants like @.crctable or similar are introduced that contain precomputed CRC lookup tables.

• These tables replace dynamic loop-based CRC computation previously used during initialization.
• This change likely improves performance by eliminating runtime computation of CRC values in favor of direct memory loads from precomputed data.

Example:

@.crctable = private unnamed_addr constant [256 x i32] [...]

---

2. Simplification of CRC Initialization Loops

Functions like _Z9InitCRC32Pj, prte_initialize_crc_table, and others were previously computing CRC values using complex bit manipulation loops with shifts and XORs over 8 iterations per byte.

• These loops have been replaced with simple memcpy operations from the precomputed table.
• The use of memcpy to copy the static CRC table into the target array significantly simplifies code and reduces runtime overhead.

Before:
Manual CRC polynomial calculation via bitwise operations in a loop.

After:

tail call void @llvm.memcpy.p0.p0.i64(...)

---

3. Inlining of CRC Table Access into Computation Functions

In several cases (e.g., @i2c_smbus_pec, @dissect_mstp), the original code computed CRC values using iterative bit shifting and conditional XORs in a loop.

• These computations are now replaced with:
  • A byte-to-index conversion,
  • A load from the precomputed @.crctable,
  • A shift and xor to update the running CRC value.

This transforms an O(8) per-byte loop into a single lookup plus shift/xor, which is more efficient and better suited for modern CPUs.

Key benefit: Eliminates small inner loops in CRC logic, reducing branch mispredictions and improving throughput.

---

4. Elimination of Polynomial-Based Bitwise Loops

Several functions previously used a loop structure that iterated 8 times per byte to compute CRC manually using the generator polynomial.

• These loops are removed and replaced with indexed loads from the precomputed CRC table.
• For example, the function i2c_smbus_pec no longer uses a loop to manipulate bits; instead, it computes the index using zext and getelementptr.

Result: More compact and faster code due to reduced control flow complexity.

---

5. Refactoring of Loop Structures and PHI Nodes

Loop structures in many functions are rewritten with simplified induction variables and fewer PHIs.

• Phi nodes are updated to reflect simpler iteration patterns.
• Some labels and branches associated with inner loops are eliminated.
• Loop metadata is updated accordingly.

This indicates that the optimizer has recognized and merged or vectorized some of the previous complex control flows, especially where CRC calculations were involved.

Impact: Cleaner control flow graphs, potentially enabling further optimizations like unrolling or vectorization.

---

Summary

These changes collectively represent a shift from runtime CRC computation to lookup-based CRC using precomputed tables. Key benefits include:

• Reduced CPU usage due to elimination of per-byte bit-twiddling loops.
• Faster initialization via memcpy.
• Better cache utilization and predictability from memory-bound lookups.

The patch appears to be optimizing performance-critical paths involving CRC calculations in various subsystems (filesystems, I2C bus, network protocols, etc.).

model: qwen-plus-latest
CompletionUsage(completion_tokens=807, prompt_tokens=72547, total_tokens=73354, completion_tokens_details=None, prompt_tokens_details=None)

[artagnon]

/close