Intermediate Code Generation

Transcription

1 Intermediate Code Generation Basic Approach and Application to Assignment and xpressions Array xpressions Boolean xpressions Copyright 2015, Pedro C. Diniz, all rights reserved. Students enrolled in the Compilers class at the University of Southern California have explicit permission to make copies of these materials for their personal use.

2 Structure of a Compiler O(n) O(n) O(n log n) Scanner words Parser IR Analysis & Optimization IR Intermediate Code Generation O(n) IR regs Code Generation Very Fast or Very Hard NP Complete asm k regs A compiler is a lot of fast stuff followed by some hard problems The hard stuff is mostly in Code Generation and Optimization For super-scalars, its Allocation & Scheduling that counts

3 Intermediate Code Generation IR Parse tree AST Intermediate Code Generation O(n) IR Three-Address Instructions regs Direct Translation Using SDT scheme Parse Tree to Three-Address Instructions Can be done while Parsing in a Single Pass Needs to be able to deal with Syntactic rrors and Recovery Indirect Translation First valate parsing constructing of AST Uses SDT scheme to build AST Traverse the AST and generate Three Address Instructions This Lecture Same but asier

4 Three-Address Instructions IR High-level Constructs mapped to Three-Address Instructions Register-based IR for xpression valuation Infinite Number of Virtual Registers Still Independent of Target Architecture Parameter Passing Discipline either on Stack or via Registers Addresses and Instructions Symbolic Names are addresses of the corresponding source-level variable. Various constants, such as numeric and offsets (known at compile time) Generic Instruction Format: Label: x = y op z or if exp goto L Statements can have Symbolic Labels Compiler inserts Temporary Variables (any variable with t prefix) Type and Conversions dealt in other Phases of the Code Generation

5 Assignments: Three-Address Instructions x = y op z (binary operator) x = op y (unary) x = y (copy) x = y[i] and x[i] = y (array indexing assignments) x = phi y z (Static Single Assignment instruction) Memory Operations: x = &y; x = *y and *x = y; for assignments via pointer variables.

6 Three-Address Instructions Control Transfer and Function Calls: goto L (unconditional); if (a relop b) goto L (conditional) where relop is a relational operator consistent with the type of the variables a and b; y = call p, n for a function or procedure call instruction to the name or variable p p might be a variable holding a set of possible symbolic names (a function pointer) the value n specifies that before this call there were n putparam instructions to load the values of the arguments. the param x instruction specifies a specific value in reverse order (i.e, the param instruction closest to the call is the first argument value. Later we will talk about parameter passing disciplines (Run-Time nv.)

7 Function Call xample Source Code Three Address Instructions y = p(a, b+1) int p(x,z){ return x+z; } t1 = a t2 = b + 1 putparam t1 putparam t2 y = call p, 2 p: getparam z getparam x t3 = x + z return t3 return

8 Function Call xample Source Code Three Address Instructions y = p(a, b+1) int p(x,z){ return x+z; } argument evaluation passing args on the stack getting values from the stack t1 = a t2 = b + 1 putparam t1 putparam t2 y = call p, 2 p: getparam z getparam x t3 = x + z return t3 return

9 Loop xample Source Code do i = i + 1; while (a[i] < v); Three Address Instructions L: t1 = i + 1 i = t1 t2 = i * 8 t3 = a[t2] if t3 < v goto L

10 Loop xample Source Code do i = i + 1; while (a[i] < v); Three Address Instructions L: t1 = i + 1 i = t1 t2 = i * 8 t3 = a[t2] if t3 < v goto L Where d this come from?

11 SDT for Three Address Code Generation Attributes for the Non-Terminals, and S Location (in terms of temporary variable) of the value of an expression:.place. If.place is t1 it means that the value of is saved in t1. The Code that valuates the xpressions or Statement:.code Markers for beginning and end of sections of the code S.begin, S.end For simplicity these are symbolic labels. Semantic Actions in Productions of the Grammar Functions to create temporaries newtemp, and labels newlabel Auxiliary functions to enter symbols and lookup types corresponding to declarations in a symbol table. To generate the code we use the function gen which creates a list of instructions to be emitted later and can generate symbolic labels corresponding to next instruction of a list. Use of append function on lists of instructions. Make use of Synthesized and Inherited Attributes

12 Assignment Statements S = { p = lookup(.name); if (p!= NULL){ S.code = gen(p =.place); } else } error; S.code = nulllist; } } {.place = newtemp();.code = append( 1.code, 2.code, gen(.place = 1.place + 2.place); } 1 * 2 {.place = newtemp();.code = append( 1.code, 2.code, gen(.place = 1.place * 2.place); } - 1 {.place = newtemp();.code = append( 1.code,gen(.place = - 1.place)); } ( 1 ) {.place = 1.place;.code = 1.code; } { p = lookup(.name); if (p!= NULL).place = p; else error;.code = nulllist; }

13 Assignment: xample x = a * b + c * d - e * f; S = + x * - * * a b c d e f

14 Assignment: xample x = a * b + c * d - e * f; S = + Production: { p = lookup(.name); if (p!= NULL).place = p; else error;.code = null list; } x * - * * a b c d e f

15 Assignment: xample x = a * b + c * d - e * f; S = + Production: { p = lookup(.name); if (p!= NULL).place = p; else error;.code = null list; } x * - * * place = loc(e) code = null place = loc(f) code = null a b c d e f

16 Assignment: xample x = a * b + c * d - e * f; S Production: 1 * 2 {.place = newtemp();.code = gen(.place = 1.place * 2.place);} = + x * - place = loc(t1) code = {t1 = e * f;} * * place = loc(e) code = null place = loc(f) code = null a b c d e f

17 Assignment: xample x = a * b + c * d - e * f; S Production: 1 * 2 {.place = newtemp();.code = gen(.place = 1.place * 2.place);} = + x * place = loc(t2) - code = {t2 = c + d;} place = loc(t1) code = {t1 = e * f;} * place = loc(c) code = null place = loc(d) code = null * place = loc(e) code = null place = loc(f) code = null a b c d e f

18 x = a * b + c * d - e * f; place = loc(x) S = code = null x a Assignment: xample * b + * Production: S = { p = lookup(.name); if (p!= NULL).code = append(.code, gen(p =.place)); code = {t1 = e * f; t2 = c * d; t3 = t2 - t1; t4 = a * b; t5 = t4 + t3; x = t5; } else error; place = loc(t5) } place = loc(a) code = null place = loc(t4) code = {t4 = a * b;} place = loc(b) code = null code = {t1 = e * f; t2 = c * d; t3 = t2 - t1; t4 = a * b; t5 = t4 + t3; } place = loc(c) code = null place = loc(t2) - code = {t2 = c + d;} place = loc(d) code = null place = loc(t3) code = {t1 = e * f; t2 = c * d; t3 = t2 - t1; } * place = loc(e) code = null place = loc(t1) code = {t1 = e + f;} place = loc(f) code = null c d e f

19 Assignment: xample x = a * b + c * d - e * f; S = + x * - t1 = e * f; t2 = c * d; t3 = t2 - t1; t4 = a * b; t5 = t4 + t3; x = t5; * * a b c d e f

20 Reusing Temporary Variables Temporary Variables Short lived Used for valuation of xpressions Clutter the Symbol Table Change the newtemp Function Keep track of when a value created in a temporary is used Use a counter to keep track of the number of active temporaries When a temporary is used in an expression decrement counter When a temporary is generated by newtemp increment counter Initialize counter to zero (0) Alternatively, can be done as a post-processing pass

21 Assignment: xample x = a * b + c * d - e * f; // c = 0 S t1 = e * f; // c = 1 t2 = c * d; // c = 2 = t1 = t2 - t1; // c = 1 x + * - t2 = a * b; // c = 2 t1 = t2 + t1; // c = 1 * * x = t1; // c = 0 a b c d e f Only 2 Temporary Variables and hence only 2 registers are Needed

22 Code Generation for Array Accesses x = A[y,z];; S L = Questions: What is the Base Type of A? What are the Dimensions of A? What is the layout of A? x L list ] Where to Get Answers? Use Symbol Table list, Check Array Layout A [ L L z? t1 = y * 20 t1 = t1 * z t2 = basea - 84 t3 = 4 * t1 t4 = t2[t3] y X = t4

23 How does the Compiler Handle A[i,j]? First, we must agree on a Storage Scheme or Layout Row-major order Lay out as a sequence of consecutive rows Rightmost subscript varies fastest A[1,1], A[1,2], A[1,3], A[2,1], A[2,2], A[2,3] Column-major order Lay out as a sequence of columns Leftmost subscript varies fastest A[1,1], A[2,1], A[1,2], A[2,2], A[1,3], A[2,3] Indirection vectors Vector of pointers to pointers to to values Takes much more space, trades indirection for arithmetic Not amenable to analysis (most languages) (Fortran) (Java)

24 The Concept A Row-major order 1,1 1,2 1,3 1,4 2,1 2,2 2,3 2,4 Laying Out Arrays These have distinct & different cache behavior A 1,1 1,2 1,3 1,4 2,1 2,2 2,3 2,4 Column-major order A 1,1 2,1 1,2 2,2 1,3 2,3 1,4 2,4 Indirection vectors 1,1 1,2 1,3 1,4 A 2,1 2,2 2,3 2,4

25 Accessing an Array lement How to access A[ i ]? Computing the Address of a 1-D Array lement: basea + ( i low ) x sizeof(basetype(a)) In General: base(a) + ( i low ) x sizeof(basetype(a))

26 Computing an Array Address Computing the Address of A[ i ]: Computing the Address of a 1-D Array lement: basea + ( i low ) x sizeof(basetype(a)) In General: base(a) + ( i low ) x sizeof(basetype(a)) Almost always a power of 2, known at compile-time and use a shift for speed Defined by the storage class of the array. int A[1:10] where low is 1; Make low 0 for faster access (saves a ) as in the C language

27 A[ i ] basea + ( i low ) x sizeof(basetype(a)) In general: base(a) + ( i low ) x sizeof(basetype(a)) What about A[i 1,i 2 ]? Computing an Array Address Row-major order, two dimensions basea + (( i 1 low 1 ) x (high 2 low 2 + 1) + i 2 low 2 ) x sizeof(basetype(a)) Column-major order, two dimensions basea + (( i 2 low 2 ) x (high 1 low 1 + 1) + i 1 low 1 ) x sizeof(basetype(a)) Indirection vectors, two dimensions *(A[i 1 ])[i 2 ] where A[i 1 ] is, itself, a 1-d array reference This stuff looks expensive! Lots of implicit +, -, x ops

28 Optimizing Address Calculation for A[i,j] In row-major order where w = sizeof(basetype(a)) basea + (i low 1 )(high 2 low 2 +1) x w + (j low 2 ) x w Which can be factored into basea + i x (high 2 low 2 +1) x w + j x w (low 1 x (high 2 low 2 +1) x w) + (low 2 x w) If low i, high i, and w are known, the last term is a constant Define basea 0 as basea (low 1 x (high 2 low 2 +1)) x w + low 2 x w and len 2 as (high 2 -low 2 +1) Then, the address expression becomes basea 0 + (i x len 2 + j ) x w Compile-time constants

29 Address Calculation for A[i 1,i 2,,i k ] A[i 1,i 2,,i k ] Addressing generalizes to (( ((i 1 n 2 +i 2 )n 3 +i 3 ) )n k +i k ) x w + +basea - (( ((low 1 n 2 +low 2 )n 3 +low 3 ) )m k +low k ) x w where n j = high j - low j + 1 and w = sizeof(basetype(a)) First term can be computed using the recurrence e 1 = i 1 e m = e m-1 x n m + i m at the end multiply by w and add compile-time constant term Compile-time constants

30 SDT for Addressing Arrays lements Three Attributes place: just the name or base address of the array offset: the index value into the array ndim: the number of dimensions Use the Recurrence to Compute Offset offset 1 = i 1 offset m = offset m-1 x n m + i m At the end multiply by w = sizeof(basetype(a)) Add the compile-time constant term Keep track of which dimension at each level Use the auxiliary function n m = numlem(a,m) as the number of elements along the m th dimension of A

31 SDT for Addressing Arrays lements L list ] { L.place = newtemp(); L.offset = newtemp(); code1 = gen(l.place = constterm(list.array)); code2 = gen(l.offset = list.place * sizeof(basetype(list.array))); L.code = append(list.code,code1,code2); } list list 1, { t = newtemp(); m = list 1.ndim + 1; x = A[y,z]; code1 = gen(t = list 1.place * numlem(list 1.array,m)); code2 = gen(t = t +.place); list.array = list 1.array; list.place = t; list.ndim = m; list.code = append(list 1.code,.code,code1,code2); } S L = x list list, A [ L L L z ] list [ { list.array =.place; list.place =.place; list.ndim = 1; List.code =.code; } y

32 SDT for Addressing Arrays lements L { if (L.offset = NULL) then.place = L.place; else.place = newtemp;.code = gen(.place = L.place[L.offset]); } S L = } { if L.offset = NULL then.code = gen(l.place =.place); else S.code = append(.code,gen(l.place[l.offset] =.place); L { L.place =.place; } L.offset = null;

33 SDT for Addressing Arrays lements x = A[y,z]; S A is 10 x 20 array with low 1 = low 2 = 1 sizeof(basetype(a)) = sizeof(int) = 4 bytes L = x list L ] list, A [ L L z y

34 SDT for Addressing Arrays lements x = A[y,z]; S A is 10 x 20 array with low 1 = low 2 = 1 sizeof(basetype(a)) = sizeof(int) = 4 bytes L.place = x L.offset = null =.place = t4 L.place = t2 x list.place = t1 list.ndim = 2 list.array = A L.offset = t3 ] t1 = y * 20 // numlem(a,2) = 20 t1 = t1 + z //.place is z A [ list.place = y list.ndim = 1 list.array = A.place = y,.place = z L.place = z L.offset = null t2 = c // c = base(a) - (20+1)*4 t3 = t1 * 4 // sizeof(int) = 4 t4 = t2[t3] L.place = y L.offset = null z x = t4 y

35 Array References What about Arrays as Actual Parameters? Whole arrays, as call-by-reference parameters Need dimension information build a dope vector Store the values in the calling sequence Pass the address of the dope vector in the parameter slot Generate complete address at each reference Some improvement is possible Save len i and low i rather than low i and high i Pre-compute the fixed terms in prologue sequence basea low 1 high 1 low 2 high 2

36 Array References What about A[12] as an actual parameter? If corresponding parameter is a scalar, it s easy Pass the address or value, as needed Must know about both formal & actual parameter Language definition must force this interpretation What is corresponding parameter is an array? Must know about both formal & actual parameter Meaning must be well-defined and understood Cross-procedural checking of conformability

37 Array References What about Variable-Sized Arrays? Local arrays dimensioned by actual parameters Same set of problems as parameter arrays Requires dope vectors (or equivalent) dope vector at fixed offset in activation record Different access costs for textually similar references This presents a lot of opportunity for a good optimizer Common sub-expressions in the address Contents of dope vector are fixed during each activation Handle them like parameter arrays

38 Array Address Calculations in a Loop DO J = 1, N A[I,J] = A[I,J] + B[I,J] ND DO Naïve: Perform the address calculation twice DO J = 1, N R1 = basea 0 + (J x len 1 + I ) x floatsize R2 = baseb 0 + (J x len 1 + I ) x floatsize MM(R1) = MM(R1) + MM(R2) ND DO

39 Array Address Calculations in a Loop DO J = 1, N A[I,J] = A[I,J] + B[I,J] ND DO Sophisticated: Move common calculations out of loop R1 = I x floatsize c = len 1 x floatsize! Compile-time constant R2 = basea 0 + R1 R3 = baseb 0 + R1 DO J = 1, N a = J x c R4 = R2 + a R5 = R3 + a MM(R4) = MM(R4) + MM(R5) ND DO

40 Array Address Calculations in a Loop DO J = 1, N A[I,J] = A[I,J] + B[I,J] ND DO Very sophisticated: Convert multiply to add (Operator Strength Reduction) R1 = I x floatsize c = len 1 x floatsize! Compile-time constant R2 = basea 0 + R1 ; R3 = baseb 0 + R1 DO J = 1, N R2 = R2 + c R3 = R3 + c MM(R2) = MM(R2) + MM(R3) ND DO

41 SDT Scheme for Boolean xpressions Two Basic Code Generation Flavors Use boolean and, or and not instructions (like arithmetic). Control-flow (or positional code) defines true or false of predicate. Arithmetic valuation Simpler to generate code as just eagerly evaluate the expression. Associate 1 or 0 with outcome of predicates and combine with logic instructions. Use the same SDT scheme explained for arithmetic operations. Control Flow valuation (short circuit evaluation) More efficient in many cases. Complications: Need to Know Address to Jump To in Some Cases Solution: Two Additional Attributes nextstat (Inherited) Indicates the next symbolic location to be generated laststat (Synthesized) Indicates the last location filled As code is generated down and up the tree attributes are filled with the correct values

42 Arithmetic Scheme: Grammar and Actions false.place = newtemp().code = {gen(.place = 0)}.laststat =.nextstat + 1 true.place = newtemp().code = {gen(.place = 1)}.laststat =.nextstat + 1 ( 1 ).place = 1.place;.code = 1.code; 1.nextstat =.nextstat.laststat = 1.laststat not 1.place = newtemp().code = append( 1.code,gen(.place = not 1.place)) 1.nextstat =.nextstat.laststat = 1.laststat + 1

43 Arithmetic Scheme: Grammar and Actions 1 or 2.place = newtemp().code = append( 1.code, 2.code,gen(.place = 1.place or 2.place) 1.nextstat =.nexstat 2.nextstat = 1.laststat.laststat = 2.laststat and 2.place = newtemp() 2.place).code = append( 1.code, 2.code,gen(.place = 1.place and 1.nextstat =.nexstat 2.nextstat = 1.laststat.laststat = 2.laststat relop 2.place = newtemp().code = gen(if 1.place relop 2.place goto.nextstat+3).code = append(.code,gen(.place = 0)).code = append(.code,gen(goto.nextstat+2)).code = append(.code,gen(.place = 1)).laststat =.nextstat + 4

44 Boolean xpressions: xample a < b or c < d and e < f relop a < b or relop and relop c < d e < f 00: if a < b goto 03 01: t1 = 0 02: goto 04 03: t1 = 1 04: if c < d goto 07 05: t2 = 0 06: goto 08 07: t2 = 1 08: if e < f goto 11 09: t3 = 0 10: goto 12 11: t3 = 1 12: t4 = t2 and t3 13: t5 = t1 or t4

45 Summary Intermediate Code Generation Using Syntax-Directed Translation Schemes xpressions and Assignments Array xpressions and Array References Boolean xpressions (Arithmetic Scheme)