Unoptimized Code Generation

Transcription

1 Unoptimized Code Generation

2 Last time we left off on the procedure abstraction Saman Amarasinghe MIT Fall 1998

3 The Stack Arguments 0 to 6 are in: %b %rbp %rsp %rdi, %rsi, %rdx, %rcx, %r8 and %r9 marks the beginning of the current frame marks the end 8*n+16(%rbp) 16(%rbp) 8(%rbp) 0(%rbp) -8(%rbp) -8*m-8(%rbp) 0(%rsp) argument n argument 7 Return address Previous %rbp local 0 local m Variable size Pre vious Cu urrent Saman Amarasinghe MIT Fall 2001

4 33 Question: Why use a stack? Wh y not use the heap or preallocated in the data segment? Saman Amarasinghe MIT Fall 2001

5 Procedure Linkages Standard procedure linkage procedure p prolog procedure q prolog pre-call post-return epilog epilog Pre-call: Save caller-saved registers Push arguments Prolog: Push old frame pointer Save calle-saved registers Make room for temporaries Epilog: Restore callee-saved Pop old frame pointer Store return value Post-return: Restore R t caller-saved Pop arguments Saman Amarasinghe MIT Fall 2001

6 Calling: Caller Assume %rcx is live and is caller save Call foo(a, B, C, D, E, F, G, H, I) A to I are at -8(%rbp) to -72(%rbp) push push push push %rcx -72(%rbp) -64(%rbp) -56(%rbp) Stack mov -48(%rbp), %r9 mov -40(%rbp), %r8 mov mov mov mov call -32(%rbp), %rcx -24(%rbp), %rdx -16(%rbp), %rsi -8(%rbp), %rdi foo return address previous frame pointer calliee saved registers local variables stack temporaries dynamic area caller saved registers argument 9 argument 8 argument 7 return address rbp rsp Saman Amarasinghe MIT Fall 2001

7 Calling: Calliee Stack return address previous frame pointer calliee saved registers Assume %rbx is used in the function local variables and is calliee save stack temporaries Assume 40 bytes are required for locals dynamic area rbp foo: push %rbp mov sub %rsp, %rbp enter $48, $0 $48, %rsp mov %rbx, -8(%rbp) caller saved registers argument 9 argument 8 argument 7 return address previous frame pointer calliee saved registers local variables stack temporaries dynamic area rsp Saman Amarasinghe MIT Fall 2001

8 Arguments Call foo(a, B, C, D, E, F, G, H, I) CODE Control Flow Procedures Statements Data Access Passed in by pushing before the call push -72(%rbp) push -64(%rbp) push -56(%rbp) mov -48(%rbp), %r9 mov -40(%rbp), %r8 mov -32(%rbp), %rcx mov -24(%rbp), %rdx mov -16(%rbp), %rsi mov -8(%rbp), %rdi call foo Access A to F via registers or put them in local memory Access rest using 16+xx(%rbp) mov 16(%rbp), %rax mov 24(%rbp), %r10 DATA Global Static Variables Global Dynamic Data Local Variables Temporaries Parameter Passing Read-only Data Stack return address previous frame pointer calliee saved registers local variables stack temporaries dynamic area caller saved registers argument 9 argument 8 argument 7 return address previous frame pointer calliee saved registers local variables stack temporaries dynamic area rbp rsp Saman Amarasinghe MIT Fall 2001

9 Locals and Temporaries Stack Calculate the size and allocate space on the stack sub $48, %rsp return address previous frame pointer calliee saved registers local variables stack temporaries dynamic area or enter $48, 0 caller saved registers argument 9 argument 8 argument 7 Access using -8-xx(%rbp) return address mov -28(%rbp), %r10 previous frame pointer mov %r11, -20(%rbp) calliee saved registers CODE Control Flow Procedures Statements Data Access DATA Global Static Variables Global Dynamic Data Local Variables Temporaries Parameter Passing Read-only Data local variables stack temporaries dynamic area rbp rsp Saman Amarasinghe MIT Fall 2001

10 Returning Calliee Stack Assume the return value is the first temporary Restore the caller saved register Put the return value in %rax Tear-down the call stack mov -8(%rbp), %rbx mov -16(%rbp), %rax mov pop ret leave %rbp, %rsp %rbp return address previous frame pointer calliee saved registers local variables stack temporaries dynamic area caller saved registers argument 9 argument 8 argument 7 return address previous frame pointer calliee saved registers local variables stack temporaries dynamic area rbp rsp Saman Amarasinghe MIT Fall 2001

11 Returning Caller Stack return address previous frame pointer calliee saved registers (Assume the return value goes to the first local variables temporary) stack temporaries Restore the stack to reclaim the argument space Restore the caller save registers Save the return value dynamic area caller saved registers argument 9 argument 8 argument 7 rbp rsp call foo CODE DATA add $24, %rsp pop %rcx mov %rax, 8(%rbp) Procedures Global Static Variables Global Dynamic Data Control Flow Local Variables Statements Temporaries Parameter Passing Data Access Read-only Data Saman Amarasinghe MIT Fall 2001

12 47 Question: Do you need the $rbp? What are the advantages and disadvantages of having $rbp? Saman Amarasinghe MIT Fall 2001

13 So far we covered.. CODE Procedures Control Flow Statements Data Access DATA Global Static Variables Global Dynamic Data Local Variables Temporaries Parameter Passing Read-only Data Saman Amarasinghe MIT Fall 1998

14 8 Outline Generation of expressions and statements Generation of control flow x86-64 Processor Guidelines in writing a code generator Saman Amarasinghe MIT Fall 1998

15 Expressions Expressions are represented as trees Expression may produce a value Or, it may set the condition codes (boolean exprs) How do you map expression trees to the machines? How to arrange the evaluation order? Where to keep the intermediate t values? Two approaches Stack Model Flat List Model Saman Amarasinghe MIT Fall 1998

16 Evaluating expression trees Stack model Eval left-sub-tree Put the results on the stack Eval right-sub-tree Put the results on the stack Get top two values from the stack perform the operation OP put the results on the stack Very inefficient! OP Saman Amarasinghe MIT Fall 1998

17 Evaluating expression trees Flat List Model The idea is to linearize the expression tree Left to Right Depth-First Traversal of the expression tree Allocate temporaries for intermediates (all the nodes of the tree) New temporary for each intermediate All the temporaries on the stack (for now) Each expression is a single 3-addr op x = y op z Code generation for the 3-addr expression Load y into register %r10 Load z into register %r11 Perform op %r10, %r11 Store %r11 to x Saman Amarasinghe MIT Fall 1998

18 Issues in Lowering Expressions Map intermediates to registers? registers are limited when the tree is large, registers may be insufficient allocate space in the stack No machine instruction is available May need to expand the intermediate operation into multiple machine ops. Very inefficient too many copies don t worry, we ll take care of them in the optimization passes keep the code generator very simple Saman Amarasinghe MIT Fall 1998

19 What about statements? Assignment statements are simple Generate code for RHS expression Store the resulting value to the LHS address But what about conditionals and loops? Saman Amarasinghe MIT Fall 1998

20 28 Outline Generation of statements Generation of control flow Guidelines in writing a code generator Saman Amarasinghe MIT Fall 1998

21 Two Approaches Template Matching Approach Peephole Optimization Algorithmic Approach Both are based on structural induction Generate a representation for the sub-parts Combine them into a representation for the whole Saman Amarasinghe MIT Fall 1998

22 Generation of control flow: Template Matching Approach Flatten the control structure use a template Put unique labels for control join points Now generate the appropriate code Saman Amarasinghe MIT Fall 1998

23 29 Template for conditionals if (test) true_body else false_body <do the test> joper lab_true <false_body> jmp lab_end lab_true: <true_body> lab _ end: Saman Amarasinghe MIT Fall 1998

24 if(ax > bx) dx = ax - bx; else dx = bx - ax; Example Program <do test>.l0:.l1: joper.l0 <FALSE BODY> jmp.l1 <TRUE BODY> Return address previous frame pointer Local variable px (10) Local variable py (20) Local variable pz (30) Argument 9: cx (30) Argument 8: bx (20) Argument 7: ax (10) Return address previous frame pointer Local variable dx (??) Local variable dy (??) Local variable dz (??) rbp rsp Saman Amarasinghe MIT Fall 1998

25 if(ax > bx) dx = ax - bx; else dx = bx - ax; Example Program.L0:.L1: movq 16(%rbp), %r10 movq 24(%rbp), %r11 cmpq %r10, %r11 jg.l0 <FALSE BODY> jmp.l1 <TRUE BODY> Return address previous frame pointer Local variable px (10) Local variable py (20) Local variable pz (30) Argument 9: cx (30) Argument 8: bx (20) Argument 7: ax (10) Return address previous frame pointer Local variable dx (??) Local variable dy (??) Local variable dz (??) rbp rsp Saman Amarasinghe MIT Fall 1998

26 if(ax > bx) dx = ax - bx; else dx = bx - ax; Example Program.L0:.L1: movq 16(%rbp), %r10 movq 24(%rbp), %r11 cmpq %r10, %r11 jg.l0 movq 24(%rbp), %r10 movq 16(%rbp), %r11 subq %r10, %r11 movq %r11, -8(%rbp) jmp.l1 <TRUE BODY> Return address previous frame pointer Local variable px (10) Local variable py (20) Local variable pz (30) Argument 9: cx (30) Argument 8: bx (20) Argument 7: ax (10) Return address previous frame pointer Local variable dx (??) Local variable dy (??) Local variable dz (??) rbp rsp Saman Amarasinghe MIT Fall 1998

27 if(ax > bx) dx = ax - bx; else dx = bx - ax; Example Program.L0:.L1: movq 16(%rbp), %r10 movq 24(%rbp), %r11 cmpq %r10, %r11 jg.l0 movq 24(%rbp), %r10 movq 16(%rbp), %r11 subq %r10, %r11 movq %r11, -8(%rbp) jmp.l1 movq 16(%rbp), %r10 movq 24(%rbp), %r11 subq %r10, %r11 movq %r11, -8(%rbp) Return address previous frame pointer Local variable px (10) Local variable py (20) Local variable pz (30) Argument 9: cx (30) Argument 8: bx (20) Argument 7: ax (10) Return address previous frame pointer Local variable dx (??) Local variable dy (??) Local variable dz (??) rbp rsp Saman Amarasinghe MIT Fall 1998

28 Template for while loops while (test) body Saman Amarasinghe MIT Fall 1998

29 Template for while loops while (test) body lab_cont: <do the test> joper lab_body body jmp lab_end lab_body: <body> jmp lab_cont lab_end: Saman Amarasinghe MIT Fall 1998

30 31 while (test) body Template for while loops lab_end: An optimized template CODE Control Flow Procedures Statements Data Access DATA Global Static Variables Global Dynamic Data Local Variables Temporaries Parameter Passing Read-only Data lab_cont: <do the test> joper lab_body body jmp lab_end lab_body: <body> jmp lab_cont lab_cont: <do the test> joper lab_end <body> jmp lab_cont lab_end: Saman Amarasinghe MIT Fall 1998

31 33 Question: What is the template for? do body while (test) Saman Amarasinghe MIT Fall 1998

32 33 Question: What is the template for? do body while (test) lab_begin: <body> <do test> joper lab_begin Saman Amarasinghe MIT Fall 1998

33 33 Question: What is a drawback of the template based approach? Saman Amarasinghe MIT Fall 1998

34 Control Flow Graph (CFG) Starting gp point: high level intermediate format, symbol tables Target: CFG CFG Nodes are Instruction Nodes CFG Edges Represent Flow of Control Forks At Conditional Jump Instructions Merges When Flow of Control Can Reach A Point Multiple Ways Entry and Exit Nodes Saman Amarasinghe MIT Fall 1998

35 entry if (x < y) { a = 0; } else { a = 1; } mov $0, a jl xxx < cmp %r10, %r11 mov x, %r10 Mov y, %r11 mov $1, a exit Pattern for if then else Saman Amarasinghe MIT Fall 1998

36 Short-Circuit Conditionals In program, conditionals have a condition written as a boolean expression ((i < n) && (v[i]!= 0)) i > k) Semantics say should execute only as much as required to determine condition Evaluate (v[i]!= 0) only if (i < n) is true Evaluate i > k only if ((i < n) && (v[i]!= 0)) is false Use control-flow graph to represent this short- circuit evaluation Saman Amarasinghe MIT Fall 1998

37 Short-Circuit Conditionals while (i < n && v[i]!= 0) { i=i+1; i+1; } entry jl xxx < cmp %r10, %r11 mov %r11, i add $1, %r11 jl yyy < cmp %r10, %r11 exit mov i, %r11 Saman Amarasinghe MIT Fall 1998

38 More Short-Circuit Conditionals if (a < b c!= 0) { i=i+1; i+1; entry } jl xxx < cmp %r10, %r11 mov %r11, i add $1, %r11 jne yyy < cmp %r10, %r11 mov i, %r11 exit Saman Amarasinghe MIT Fall 1998

39 Routines for Destructuring Program destruct(n) Representation ti generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form shortcircuit(c, t, f) generates short-circuit form of conditional represented by c if c is true, control flows to t node if c is false, control flows to f node returns b - b is begin node for condition evaluation new kind of node - nop node Saman Amarasinghe MIT Fall 1998

40 Destructuring Seq Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form seq x y seq x y Saman Amarasinghe MIT Fall 1998

41 Destructuring Seq Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form seq x y 1: (b x,e x ) = destruct(x); x seq y b x e x Saman Amarasinghe MIT Fall 1998

42 Destructuring Seq Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form seq x y 1: (b x,e x ) = destruct(x); 2: (b y,e y ) = destruct(y); seq x y b x e b e x b y e y Saman Amarasinghe MIT Fall 1998

43 Destructuring Seq Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form seq x y 1: (b x,e x ) = destruct(x); 2: (b y,e y ) = destruct(y); 3: next(e x ) = b y ; seq x y b x e b e x b y e y Saman Amarasinghe MIT Fall 1998

44 Destructuring Seq Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form seq x y 1: (b x,e x ) = destruct(x); 2: (b y,e y ) = destruct(y); 3: next(e x ) = b y ; 4: return (b x, e y ); b x seq x y e x b b y e y Saman Amarasinghe MIT Fall 1998

45 destruct(n) Destructuring If Nodes generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form if c x y c if x y Saman Amarasinghe MIT Fall 1998

46 destruct(n) Destructuring If Nodes generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form if c x y 1: (b x,e x ) = destruct(x); c if x y b x e x Saman Amarasinghe MIT Fall 1998

47 destruct(n) Destructuring If Nodes generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form if c x y 1: (b x,e x ) = destruct(x); 2: (b y,e y ) = destruct(y); if b x e x c x y b y e y Saman Amarasinghe MIT Fall 1998

48 destruct(n) Destructuring If Nodes generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form if c x y 1: (b x,e x ) = destruct(x); 2: (b y,e y ) = destruct(y); 3: e = new nop; c if x y b x b y e x e y e Saman Amarasinghe MIT Fall 1998

49 destruct(n) Destructuring If Nodes generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form if c x y 1: (b x,e x ) = destruct(x); 2: (b y,e y ) = destruct(y); 3: e = new nop; 4: next(e x ) = e; 5: next(e y ) = e; c if x y b x b y e x e y e Saman Amarasinghe MIT Fall 1998

50 destruct(n) Destructuring If Nodes generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form if c x y 1: (b x,e x ) = destruct(x); 2: (b y,e y ) = destruct(y); 3: e = new nop; 4: next(e x ) = e; 5: next(e y ) = e; 6: b c = shortcircuit(c, i b x, b y ); c if x y b c b x b y e x e y e Saman Amarasinghe MIT Fall 1998

51 destruct(n) Destructuring If Nodes generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form if c x y 1: (b x,e x ) = destruct(x); 2: (b y,e y ) = destruct(y); 3: e = new nop; 4: next(e x ) = e; 5: next(e y ) = e; 6: b c = shortcircuit(c, i b x, b y ); 7: return (b c, e); c if x y b c b x b y e x e y e Saman Amarasinghe MIT Fall 1998

52 Destructuring While Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form while c x while c x Saman Amarasinghe MIT Fall 1998

53 Destructuring While Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form while c x 1: e = new nop; while c x e Saman Amarasinghe MIT Fall 1998

54 Destructuring While Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form while c x 1: e = new nop; 2: (b x,e x ) = destruct(x); while c x b x e e x Saman Amarasinghe MIT Fall 1998

55 Destructuring While Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form while c x 1: e = new nop; 2: (b x,e x ) = destruct(x); 3: b c = shortcircuit(c, b x, e); while b c c x b x e e x Saman Amarasinghe MIT Fall 1998

56 Destructuring While Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form while c x 1: e = new nop; 2: (b x,e x ) = destruct(x); 3: b c = shortcircuit(c, b x, e); 4: next(e x ) = b c ; while b c c x b x e e x Saman Amarasinghe MIT Fall 1998

57 Destructuring While Nodes destruct(n) generates lowered form of structured code represented by n returns (b,e) - b is begin node, e is end node in destructed form if n is of the form while c x 1: e = new nop; 2: (b x,e x ) = destruct(x); 3: b c = shortcircuit(c, b x, e); 4: next(e x ) = b c ; 5: return (b c, e); while b c c x b x e e x Saman Amarasinghe MIT Fall 1998

58 Shortcircuiting And Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form c 1 && c 2 c 1 && c 2 Saman Amarasinghe MIT Fall 1998

59 Shortcircuiting And Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form c 1 && c 2 1: b 2 = shortcircuit(c 2, t, f); c 1 && c 2 b 2 f t Saman Amarasinghe MIT Fall 1998

60 Shortcircuiting And Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form c 1 && c 2 1: b 2 = shortcircuit(c 2, t, f); 2: b 1 = shortcircuit(c 1, b 2, f); b 1 c 1 && c 2 b 2 f t Saman Amarasinghe MIT Fall 1998

61 Shortcircuiting And Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form c 1 && c 2 1: b 2 = shortcircuit(c 2, t, f); 2: b 1 = shortcircuit(c 1, b 2, f); 3: return (b 1 ); b b 1 c 1 && c 2 b 2 f t Saman Amarasinghe MIT Fall 1998

62 Shortcircuiting Or Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form c 1 c 2 c 1 c 2 Saman Amarasinghe MIT Fall 1998

63 Shortcircuiting Or Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form c 1 c 2 1: b 2 = shortcircuit(c 2, t, f); c 1 c 2 b 2 t f Saman Amarasinghe MIT Fall 1998

64 Shortcircuiting Or Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form c 1 c 2 1: b 2 = shortcircuit(c 2, t, f); 2: b 1 = shortcircuit(c 1, t, b 2 ); b 1 c 1 c 2 b 2 t f Saman Amarasinghe MIT Fall 1998

65 Shortcircuiting Or Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form c 1 c 2 1: b 2 = shortcircuit(c 2, t, f); 2: b 1 = shortcircuit(c 1, t, b 2 ); 3: return (b 1 ); b b 1 c 1 c 2 b 2 t f Saman Amarasinghe MIT Fall 1998

66 Shortcircuiting Not Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form! c 1 1: b = shortcircuit(c 1, f, t); return(b);! c 1 b f t Saman Amarasinghe MIT Fall 1998

67 Computed Conditions shortcircuit(c, t, f) generates shortcircuit form of conditional represented by c returns b - b is begin node of shortcircuit form if c is of the form e 1 < e 2 1: b = new cbr(e 1 < e 2, t, f); 2: return (b); jl e 1 < e 2 t cmp f e 1 e 2 Saman Amarasinghe MIT Fall 1998

68 Nops In Destructured Representation while (i < n && v[i]!= 0) { i=i+1; i+1; } entry jl xxx < cmp %r10, %r11 mov %r11, i add $1, %r11 nop jl yyy < cmp %r10, %r11 exit mov i, %r11 Saman Amarasinghe MIT Fall 1998

69 Eliminating Nops Via Peephole Optimization i nop Saman Amarasinghe MIT Fall 1998

70 46 Question: What are the pros and cons of template matching vs. algorithmic approach? Saman Amarasinghe MIT Fall 1998

71 49 Outline Generation of statements Generation of control flow x86-64 Processor Guidelines in writing a code generator Saman Amarasinghe MIT Fall 1998

72 Guidelines for the code generator Lower the abstraction level slowly Do many passes, that do few things (or one thing) Easier to break the project down, generate and debug Keep the abstraction level consistent IR should have correct semantics at all time At least you should know the semantics You may want to run some of the optimizations between the passes. Use assertions liberally Use an assertion to check your assumption Saman Amarasinghe MIT Fall 1998

73 Guidelines for the code generator Do the simplest but dumb thing it is ok to generate 0 + 1*x + 0*y Code is painful to look at; let optimizations improve it Make sure you know want can be done at Compile time in the compiler Runtime usinggenerated code Saman Amarasinghe MIT Fall 1998

74 Guidelines for the code generator Remember that optimizations will come later Let the optimizer do the optimizations Think about what optimizer will need and structure your code accordingly Example: Register allocation, algebraic simplification, constant propagation Setup a good testing infrastructure regression tests If a input program creates a bug, use it as a regression test Learn good bug hunting procedures Example: binary search, delta debugging Saman Amarasinghe MIT Fall 1998

75 MIT OpenCourseWare Computer Language Engineering Spring 2010 For information about citing these materials or our Terms of Use, visit: