Document PTX register width issue analysis and investigation findings

[Copilot]

This PR documents the investigation findings for the PTX register width issue that affects GPGPU-Sim's ability to correctly simulate 64-bit integer operations.

Issue Summary

The investigation identified a critical limitation in GPGPU-Sim's PTX parser that causes incorrect simulation results when handling 64-bit integer literals. This issue is tracked in Issue gpgpu-sim#325: "64-bit integer operand support".

Problem Description

When CUDA programs use 64-bit integer constants, the PTX compiler generates instructions like:

mov.u64 %rd3, 123456789012345;

However, GPGPU-Sim's PTX parser incorrectly handles these large constants because:

• The lexer (src/cuda-sim/ptx.l) uses atoi() to parse integer literals
• atoi() only supports 32-bit integers, causing truncation of the high 32 bits
• This results in incorrect values being loaded into 64-bit registers during simulation

Impact

This bug affects simulation accuracy for any CUDA application that:

• Uses 64-bit integer constants larger than 2^31-1
• Performs computations with large memory addresses on GPUs with >4GB memory
• Implements algorithms requiring precise 64-bit arithmetic

Investigation Results

The root cause was traced to three key locations:

1. src/cuda-sim/ptx.l:273 - Lexical analysis using atoi()
2. src/cuda-sim/ptx.y:644-646 - Grammar rules for integer operands
3. src/cuda-sim/ptx_parser.cc:869-872 - Parser implementation assuming 32-bit integers

Related Issues

This investigation also identified:

• Issue PTXAS Summary Parsing Error on CUDA 12.8 gpgpu-sim/gpgpu-sim_distribution#324: PTXAS parsing errors with CUDA 12.8 due to new "barriers" output format
• PR Adapt ptx grammar rules for CUDA 12.8 gpgpu-sim/gpgpu-sim_distribution#328: Attempted fix for CUDA 12.8 compatibility

Next Steps

The documented analysis provides a foundation for implementing a proper fix that would:

• Replace atoi() with strtoll() for 64-bit integer parsing
• Extend the INT_OPERAND type or add a new INT64_OPERAND token type
• Update the ptx_reg_t union and related data structures
• Ensure backward compatibility with existing 32-bit integer handling

This documentation will help future contributors understand the scope and complexity of implementing 64-bit integer support in GPGPU-Sim's PTX simulation.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.