metadata

language:
  - en
tags:
  - Compiler
  - LLVM
  - Intermediate Representation
  - IR
  - Path
  - Hot Path
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: path
      dtype: string
    - name: count
      dtype: int64
    - name: source_file
      dtype: string
    - name: label
      dtype: int64
  splits:
    - name: train
      num_bytes: 3468576
      num_examples: 1190
    - name: validation
      num_bytes: 647074
      num_examples: 211
    - name: test
      num_bytes: 194998
      num_examples: 160
  download_size: 798471
  dataset_size: 4310648

Dataset Card for Compiler Hot Paths

Dataset Description

This dataset consists of 1561 compiler paths generated from 26 C programs in the Polybench Benchmark Suite using the Ball-Larus Algorithm. Each path, a sequence of LLVM IR instructions, is has three associated values:

count, an integer indicating the number of times this path is executed in the original program.
source_file, a string indicating which program was this path from.
label, an integer of 0 or 1 indicating whether this path is "cold" or "hot" respectively.

Note: 4 programs (deriche, cholesky, gramschmidt, correlation) were excluded because we encountered errors when running them.

Uses

This dataset was used to train/fine-tune machine learning models to perform hot path predictions: Given a path, predict whether it is "hot" or "cold". A path is considered "hot" if it is executed more than a threshold of n times, where we defined n = 1, otherwise it is considered "cold".

Dataset Structure

The dataset is split into train (1190, 75%), validation (211, 15%), and test (160, 10%) sets. The test set consists of paths from 4 programs (in PolyBench), namely, jacobi-2d, syr2k, durbin, 2mm. These 4 programs were randomly selected to be the test set before generating the paths. This guarantees that the models have never seen the test set's programs. The train and validation sets consist of the remaining 22 programs, which were randomly split after generating the paths (while maintaining the hot-to-cold-paths ratio), meaning that some paths in the validation set and training set may come from the same C program. However, this likely won't be an issue since the paths themselves are distinct.