Sorted in order of personal preference.

Note that picking your tool of choice by exact numbers is silly. A 1% gain here or there will not matter in a real-world scenario - stability and featuresets will matter much more. However, these are good in establishing a general ballpark of performance.

• Oric-1 compiler benchmark (Source)
  • Tested compilers: cc65, gcc-6502, KickC, lcc65, llvm-mos, sdcc, vbcc.
  • This is the most comprehensive one. Focuses on typical benchmark samples.
  • Miscompilations should not be taken at face value; they tend to modify runtime/standard library code to fit their build system.

• sgadrat's Super Tilt Bro benchmark (Source)
  • Tested compilers: cc65, gcc-6502, KickC, llvm-mos (via pull request), vbcc.
  • Seems to mostly be based on real-world/tutorial cc65 code snippets. Some have optimized ASM to compare against.

• GGLabs benchmark (Source on website)
  • Tested compilers: cc65, gcc-6502, KickC, sdcc, vbcc.
  • Based on typical benchmark samples.

• NESFab benchmark (Could not locate source.)
  • Tested compilers: cc65, gcc-6502, KickC, llvm-mos, vbcc.
  • No methodology; in addition, NESFab is more akin to a high-level assembler than a C compiler.