AI Coding — Compiler Optimization

A new April-2026 AI Coding prompt. Implement `extract_time_and_mem_cost(instruction_file)` for a toy compiler IR. Catch: instruction costs are not given — they must be inferred from the unit-test assertions. Requirements Starter project: a few folders each with a test.py, an instructionN.txt, and a src/ skeleton. Each instruction file is straight-line three-address code: res1 = var1 + var2 res2 = var3 - var4 res3 = res2 + var5 res = res1 + res3 Implement extract_time_and_mem_cost(instruction) returning (time, mem). Unit tests look like assert extract_time_and_mem_cost('test/instruction1.txt') == (14, ...). The per-operator costs (+, -, *, /, =) are deliberately omitted from the prompt. You must back them out from the numeric assertions in the test files. Examples Reported pattern matched the assertions when +, -, = cost 1 and *, / cost 5; multiple test files cross-validate the constants. One report notes the interviewer hints toward inferring constants only in the last 5 minutes. Notes Underspecified by design — the interviewer expects you to ask clarifying questions and then notice that the cost numbers are encoded in the tests, not in the prompt. One candidate's session had Opus 4.6 crash mid-round; interviewer asked them to switch to GPT-5.x. Be ready for model swaps without losing context. Reports note that AI-generated cost values are fabricated unless you give it the assertions; treat the model's first guesses as suspect and re-prompt with the unit-test snippets. Preparation Practice parsing simple three-address-code files in Python (split on =, tokenize RHS, recognize operators and operands). Rehearse the meta-skill of "infer constants from a few labeled examples": given two (file, total_cost) pairs and one operator-count delta, solve a tiny linear system by hand. Re-check the forum 1-2 weeks before interview — this prompt is new and likely to mutate.