Lecture Slides on Compiler Organization | CMSC 212, Study notes of Computer Science

Download Lecture Slides on Compiler Organization | CMSC 212 and more Study notes Computer Science in PDF only on Docsity! 1 1 1CMSC 212 – S06 (lect 25) Announcements
Program #6 – is due in a week from Thursday
Reading – Notes 2CMSC 212 – S06 (lect 25) Improving Code Performance
This lecture is not about algorithms, but rather – how to convert algorithms into efficient code – how to refine code to make it run faster
Helpful to know a bit about – what a compiler can and can't do – what takes time on this specific hardware
Key Ideas – Be systematic and data driven in what you do • Easy to get into making random code changes – Programs spend their time in • loops (or recursion) – a sequence of code executed once is fast • doing I/O 2 2 3CMSC 212 – S06 (lect 25) Know Your Compiler
Compilers can be asked to "optimize" your code – -O flag (and -On) enables the optimizer – they are supposed to never break your code • only safe changes to the program are permitted
Limits on compiler optimizations – should never alter correct programs • may discover latent bugs though – limited understanding of the program • you are much smarter than the best compiler – need to compile program quickly 4CMSC 212 – S06 (lect 25) Understanding Modern Processors
Pipelined – parts of multiple instructions running at once – decode instruction, load from memory
Superscalar processors – Can execute 2 or more instructions at once
Branch prediction – Computer guess which way a branch will go – Allows pipeline to stay full
Cache memory – keep copies of recently accessed memory in fast storage • recently accessed memory is much faster
Floating point takes much longer than integer – typically 20-30x for the same operation 5 5 9CMSC 212 – S06 (lect 25) Common Code Changes - Use Pointers
Recall int limit = strlen(data); for (i=0; i < limit; i++) { if (data[i] > 'A' && data[i] < 'Z') data[i] -= ('A' - 'a'); }
With a function call int limit = strlen(data); for (i=0; i < limit; i++) { makelower(&data[i]); }
Can eliminate strlen call for (ptr = data; *ptr; ++ptr) { if (*ptr > 'A' && *ptr < 'Z') *ptr -= ('A' - 'a'); }
Can move one subtraction outside int diff = ‘A’ –’a’; for (ptr = data; *ptr; ++ptr) { if (*ptr > 'A' && *ptr < 'Z') *ptr -= diff; } 10CMSC 212 – S06 (lect 25) Common Code Transformation - Don't write small functions
Functions calls take time – Recall last week's code to setup calls stacks
1-2 line functions – take more time for the call than work – provide relatively little isolation of ideas
Solution: – Define abstractions so work is larger than a few lines 6 6 11CMSC 212 – S06 (lect 25) Approach To Applying These Techniques
For a big program – Hard to know where to start to optimize
Start with gprof data – concentrate in parts of the program that take the most time
Ahmdals Law – if tuning a function that takes α share of the time – and make it k times faster – Tnew= ( 1 - α) Told + (αTold)/k – Basically are limited by how big α is