Register Allocation - Compiler Optimisation - Lecture Slides, Slides of Computer Science

Download Register Allocation - Compiler Optimisation - Lecture Slides and more Slides Computer Science in PDF only on Docsity! Register Allocation Christophe Dubach February, 2013 M. O’Boyle Register Allocation February, 2013 1 Course Structure • L1 Introduction and Recap • L2 Course Work • L3+4 Scalar optimisation and dataflow • L5 Code generation • L6 Instruction scheduling • L7 Register allocation • Then high level approaches followed by adaptive compilation M. O’Boyle Register Allocation February, 2013 4 Local allocation • Focuses on basic block and maps virtual registers to physical registers • Top-down allocation computes a priority with most important ones allocated a reg the others are spilled. • Poor as virtual registers allocated a physical reg for the entire scope • Bottom-up - iterates over block allocation on demand. Frees a register if it “knows” that no longer needed. Uses distance to next use as as spill metric. • Spill clean values rather dirty as a way of minimising spill code M. O’Boyle Register Allocation February, 2013 5 Spill code 1 registers: 2 values to manage x = y = = y = x x = Mem[spill] = x y = = y x = Mem[spill] = x R1 = store R1 -> R0 % Mem[R0]=R1 R1 = = R1 load RO -> R1 % R1 = Mem[R0] = R1 Write spilled value to memory Note still need R0 register for storage address. M. O’Boyle Register Allocation February, 2013 6 Local allocation - spill code 2 registers: x1 clean in r1, x2 dirty in r2. Refer x3,x1,x2- must spill one: load x1 -> r1 load x2 -> r2 add r2, 1 -> r2 = x3 = x1 = x2 store r2 -> x2 load x3 -> r2 = r2 (use x3) = r1 (use x1) load x2 -> r2 = r2 (use x2) Spill dirty load x3 -> r1 = r1 (use x3) load x1 -> r1 = r1 (use x1) = r2 (use x2) Spill clean Not always best sequence x3,x1,x3,x1,x2 - better to spill dirty values Taking into account clean/dirty data makes it NP-complete M. O’Boyle Register Allocation February, 2013 9 Interference graph LRa = LRb= =LRb LRd= LRc= ...... LRd=LRc =LRa =LRd LRa LRb LRc LRd Live ranges interfere if one is live at the definition of another and have different values M. O’Boyle Register Allocation February, 2013 10 Graph colouring • Colour graph with k colours/registers • Important observation any node n that has less than k neighbours | n | < k can always be coloured • Pick any node |n| < k and put on stack • Remove that node and its edges - this reduces degree of neighbours • Any remaining nodes - spill one and continue • Pop nodes of stack and colour M. O’Boyle Register Allocation February, 2013 11 Graph colouring 1 2 3 4 5 1 2 3 4 5 b r g b g 3 colours. Remove 1 first as it has a degree less than 3. Colour as we pop M. O’Boyle Register Allocation February, 2013 14 Alternative spilling • Rather than spilling entire live ranges, spill only in high demand area -partial live ranges • Splitting live ranges. Can reduce degree of interference graph. Smart splitting allows spilling to occur in “cheap” regions • Coalesce - if two ranges don’t interfere and are connected by a copy - coalesce into one. Reduces degree of nodes that interfered with both M. O’Boyle Register Allocation February, 2013 15 Coalescing 1:add LRt, LRu -> LRa ... 2:addI LRa, 0 ->LRb 3:xor LRa, 0 -> LRc ... 4:add LRb, LRw -> LRx 5:add LRc, LRy -> LRz 1:add LRt, LRu -> LRab ... ... 3:xor LRab, 0 -> LRc 4:add LRab, LRw -> LRx 5:add LRc, LRy -> LRz Live range of a [1..3], b[2...4], c [3..5] connected by 2 copies in 2,3. Remove one copy here. Can also remove the other M. O’Boyle Register Allocation February, 2013 16 Coalescing Reduces Degree a cb c ab Guaranteed to not increase degree of interference on neighbours. If a node interfered with both both before, coalescing helps As it reduces degree, often applied before colouring takes place M. O’Boyle Register Allocation February, 2013 19 Ongoing work • Register allocation is a well studied topic. Linear scan for JITs • Eisenbeis et al examining optimality of combined reg alloc and scheduling. Difficulty with general control-flow • Partitioned register sets complicate matters. Allocation can require insertion of code which in turn affects allocation. Leupers investigated use of genetic algs for TM series partitioned reg sets. • New work by Fabrice Rastello and others. Chordal graphs reduce complexity • As latency increases see work in combined code generation, instruction scheduling and register allocation M. O’Boyle Register Allocation February, 2013 20 Summary • Local Allocation - spill code • Liveness analysis • Global Allocation based on graph colouring • Bottom-up approaches • Techniques to reduce spill code M. O’Boyle Register Allocation February, 2013