An Overview of GIMPLE Control construct explicated into conditional - PowerPoint PPT Presentation

1 July 2012 GIMPLE and RTL: Outline 1/39 Outline Workshop on Essential Abstractions in GCC Manipulating GIMPLE and RTL IRs • An Overview of GIMPLE GCC Resource Center • Using GIMPLE API in GCC-4.6.0 (www.cse.iitb.ac.in/grc) • Adding a GIMPLE Pass to GCC Department of Computer Science and Engineering, • An Internal View of RTL Indian Institute of Technology, Bombay • Manipulating RTL IR 1 July 2012 Essential Abstractions in GCC GCC Resource Center, IIT Bombay 1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 2/39 GIMPLE: A Recap • Language independent three address code representation Part 1 ◮ Computation represented as a sequence of basic operations ◮ Temporaries introduced to hold intermediate values An Overview of GIMPLE • Control construct explicated into conditional and unconditional jumps Essential Abstractions in GCC GCC Resource Center, IIT Bombay

1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 3/39 1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 4/39 Motivation Behind GIMPLE Why Not Abstract Syntax Trees for Optimization? • ASTs contain detailed function information but are not suitable for optimization because • Previously, the only common IR was RTL (Register Transfer Language) ◮ Lack of a common representation across languages • Drawbacks of RTL for performing high-level optimizations ◮ No single AST shared by all front-ends ◮ Low-level IR, more suitable for machine dependent optimizations ◮ So each language would have to have a different implementation of (e.g., peephole optimization) the same optimizations ◮ High level information is difficult to extract from RTL (e.g. array ◮ Difficult to maintain and upgrade so many optimization frameworks ◮ Structural Complexity references, data types etc.) ◮ Introduces stack too soon, even if later optimizations do not require it ◮ Lots of complexity due to the syntactic constructs of each language ◮ Hierarchical structure and not linear structure Control flow explication is required Essential Abstractions in GCC GCC Resource Center, IIT Bombay Essential Abstractions in GCC GCC Resource Center, IIT Bombay 1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 5/39 1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 6/39 Need for a New IR What is GENERIC? What? • Language independent IR for a complete function in the form of trees • Earlier versions of GCC would build up trees for a single statement,and • Obtained by removing language specific constructs from ASTs then lower them to RTL before moving on to the next statement • All tree codes defined in $(SOURCE)/gcc/tree.def • For higher level optimizations, entire function needs to be represented in trees in a language-independent way. Why? • Result of this effort - GENERIC and GIMPLE • Each language frontend can have its own AST • Once parsing is complete they must emit GENERIC Essential Abstractions in GCC GCC Resource Center, IIT Bombay Essential Abstractions in GCC GCC Resource Center, IIT Bombay

1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 7/39 1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 8/39 What is GIMPLE ? GIMPLE Goals • GIMPLE is influenced by SIMPLE IR of McCat compiler The Goals of GIMPLE are • But GIMPLE is not same as SIMPLE (GIMPLE supports GOTO) • Lower control flow Sequenced statements + conditional and unconditional jumps • It is a simplified subset of GENERIC • Simplify expressions ◮ 3 address representation Typically one operator and at most two operands ◮ Control flow lowering • Simplify scope ◮ Cleanups and simplification, restricted grammar Move local scope to block begin, including temporaries • Benefit : Optimizations become easier Essential Abstractions in GCC GCC Resource Center, IIT Bombay Essential Abstractions in GCC GCC Resource Center, IIT Bombay 1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 9/39 1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 10/39 Tuple Based GIMPLE Representation Observing Internal Form of GIMPLE test.c.004t.gimple with compilation option test.c.004t.gimple • Earlier implementation of GIMPLE used trees as internal data structure with compilation option -fdump-tree-all-raw • Tree data structure was much more general than was required for three -fdump-tree-all address statements x = 10; gimple_assign <integer_cst, x, 10, NULL> • Now a three address statement is implemented as a tuple y = 5; gimple_assign <integer_cst, y, 5, NULL> D.1954 = x * y; gimple_assign <mult_expr, D.1954, x, y> • These tuples contain the following information a.0 = a; gimple_assign <var_decl, a.0, a, NULL> ◮ Type of the statement x = D.1954 + a.0; gimple_assign <plus_expr, x, D.1954, a.0> ◮ Result a.1 = a; gimple_assign <var_decl, a.1, a, NULL> ◮ Operator D.1957 = a.1 * x; gimple_assign <mult_expr, D.1957, a.1, x> ◮ Operands y = y - D.1957; gimple_assign <minus_expr, y, y, D.1957> The result and operands are still represented using trees Essential Abstractions in GCC GCC Resource Center, IIT Bombay Essential Abstractions in GCC GCC Resource Center, IIT Bombay

1 July 2012 GIMPLE and RTL: An Overview of GIMPLE 11/39 Observing Internal Form of GIMPLE test.c.004t.gimple test.c.004t.gimple with compilation option with compilation option -fdump-tree-all-raw -fdump-tree-all if (a < c) Part 2 goto <D.1953>; gimple_cond <lt_expr, a,c,<D.1953>, <D.1954>> else gimple_label <<D.1953>> Manipulating GIMPLE goto <D.1954>; gimple_assign <plus_expr, a, b, c> <D.1953>: gimple_goto <<D.1955>> a = b + c; gimple_label <<D.1954>> goto <D.1955>; gimple_assign <minus_expr, a, b, c> <D.1954>: gimple_label <<D.1955>> a = b - c; <D.1955>: Essential Abstractions in GCC GCC Resource Center, IIT Bombay 1 July 2012 GIMPLE and RTL: Manipulating GIMPLE 12/39 1 July 2012 GIMPLE and RTL: Manipulating GIMPLE 12/39 Iterating Over GIMPLE Statements Iterating Over GIMPLE Statements • A basic block contains a doubly linked-list of GIMPLE statements • The statements are represented as GIMPLE tuples, and the operands are represented by tree data structure • Processing of statements can be done through iterators • A basic block contains a doubly linked-list of GIMPLE statements • The statements are represented as GIMPLE tuples, and the operands are basic_block bb; represented by tree data structure gimple_stmt_iterator gsi; • Processing of statements can be done through iterators FOR_EACH_BB (bb) { % for ( gsi=gsi_start_bb (bb); !gsi_end_p (gsi); % gsi_next (&gsi)) find_pointer_assignmentsgsi_stmt (gsi)); } Essential Abstractions in GCC GCC Resource Center, IIT Bombay Essential Abstractions in GCC GCC Resource Center, IIT Bombay

1 July 2012 GIMPLE and RTL: Manipulating GIMPLE 12/39 1 July 2012 GIMPLE and RTL: Manipulating GIMPLE 12/39 Iterating Over GIMPLE Statements Iterating Over GIMPLE Statements • A basic block contains a doubly linked-list of GIMPLE statements • A basic block contains a doubly linked-list of GIMPLE statements • The statements are represented as GIMPLE tuples, and the operands are • The statements are represented as GIMPLE tuples, and the operands are represented by tree data structure represented by tree data structure • Processing of statements can be done through iterators • Processing of statements can be done through iterators basic_block bb; basic_block bb; gimple_stmt_iterator gsi; gimple_stmt_iterator gsi; FOR_EACH_BB (bb) FOR_EACH_BB (bb) { % { % for ( gsi=gsi_start_bb (bb); !gsi_end_p (gsi); % for ( gsi=gsi_start_bb (bb); !gsi_end_p (gsi); % gsi_next (&gsi)) gsi_next (&gsi)) find_pointer_assignmentsgsi_stmt (gsi)); find_pointer_assignmentsgsi_stmt (gsi)); } } Basic block iterator GIMPLE statement iterator Essential Abstractions in GCC GCC Resource Center, IIT Bombay Essential Abstractions in GCC GCC Resource Center, IIT Bombay 1 July 2012 GIMPLE and RTL: Manipulating GIMPLE 12/39 1 July 2012 GIMPLE and RTL: Manipulating GIMPLE 12/39 Iterating Over GIMPLE Statements Iterating Over GIMPLE Statements • A basic block contains a doubly linked-list of GIMPLE statements • A basic block contains a doubly linked-list of GIMPLE statements • The statements are represented as GIMPLE tuples, and the operands are • The statements are represented as GIMPLE tuples, and the operands are represented by tree data structure represented by tree data structure • Processing of statements can be done through iterators • Processing of statements can be done through iterators basic_block bb; basic_block bb; gimple_stmt_iterator gsi; gimple_stmt_iterator gsi; FOR_EACH_BB (bb) FOR_EACH_BB (bb) { % { % for ( gsi=gsi_start_bb (bb); !gsi_end_p (gsi); % for ( gsi=gsi_start_bb (bb); !gsi_end_p (gsi); % gsi_next (&gsi)) gsi_next (&gsi)) find_pointer_assignmentsgsi_stmt (gsi)); find_pointer_assignmentsgsi_stmt (gsi)); } } Get the first statement of bb True if end reached Essential Abstractions in GCC GCC Resource Center, IIT Bombay Essential Abstractions in GCC GCC Resource Center, IIT Bombay