Code Size Reduction Using Global Code Motion

1
Code Size Reduction Using Global Code Motion
2
Overview

• Global Code Hoisting Sinking
• Global Code Merge
• Experimental Results
• Conclusions

3
Code Hoisting/Sinking - A Motivating Example
Function ulaw2alaw in benchmark
g711.c unsigned char ulaw2alaw(unsigned char
uval) uval 0xff return ((uval
0x80) ? (0xD5 (_u2a0xff
uval 1)) (0x55 (_u2a0x7f
uval 1)))
I3 and I12 can be hoisted to B1 I11 and I19 can
be sunk to B4
4
Problem Specification

• Given a single-entry, multiple-exit acyclic
  region R in control flow graph (CFG), global code
  motion tries to reduce the code size by
  eliminating identical instructions in multiple
  code paths from or to a single common point.
• Two Scenarios global code hoisting global code
  sinking
• Three constraints on legality of global code
  motion
• Local and global data dependences are preserved
• The scheduled basic block pairs are independent
• instruction integrity for each control flow path
  is preserved

5
Examples for code motion constraints

• It is illegal to hoist I1 from B2 and B3 to B1
  because the path B4?B3 will lose I1
• It is illegal to sink I2 from B1 and B4 to B3
  because the path B1?B2 will lose I2
• Instruction Integrity should be preserved, i.e.
  the number of executed instructions in any path
  is kept unchanged.

• It is illegal to hoist I1 from B2 and B3 to B1,
  because B2 and B3 are dependent.
• It is illegal to sink I2 from B1 and B2 to B3,
  because B1 and B2 are dependent
• The scheduled two BBs must be independent.

6
Algorithm Overview

• Every loop, ignoring the back edge, is a
  single-entry, single-exit region.
• The code motion algorithm is iterated on every
  nested loop body, from the innermost one to the
  outermost one. When processing the outer loop,
  the inner loop is reduced to a black box.

7
Several Key Problems

• How to identify the pairs of independent basic
  blocks which are bifurcated from a common
  immediate dominator, or joint to a common
  immediate post-dominator.
• Which instructions within a basic block are free
  to be hoisted and sunk without violating the
  existing data dependences.
• Within those movable instructions, how to
  determine whether they are identical.

8
Independence Vector
Definition 1 Two basic blocks in acyclic region
R are INDEPENDENT if and only if there is no
control path flowing through them.
Definition 2 An Dependent Vector IV(n) is a bit
vector for BB n in which the i th bit is set when
there exist a path flowing from BB i to BB n.
9
Identify Schedulable BB pairs
Problem How to traverse the CFG and maximize the
scheduling opportunity
Schedule (B5, B6), (B7, B8) before (B2, B3)
10
Identify Schedulable BB pairs (contd.)
Resulting list 5, 6, 9, 7, 10, 11, 8, 12, 13

• For code hoisting, evaluate the list from right
  to left
• For code sinking, evaluate the list from left to
  right

11
Code Motion Algorithm
12
Instruction Comparison
Code motion is constrained by data dependence

• Two instructions are identical when they have
  the identical operation and operands
• Register renaming may be used.

13
Global Code Merge
Problem Formulation Given two independent
instruction sequences and their DDGs, construct a
scheme to merge identical instructions from two
sequences under the data dependence constraints,
while obtaining the minimum number of
instructions.
14
A Solution Method
M1 Ø, F1 Ø,U1 all instructions in basic
block 1 M1 Ø, F1 Ø,U1 all instructions in
basic block 1 While U1 ? Ø U2 ? Ø do for
each instruction i in U1 if is
predecessors s in DDG satisfies s M1
U1 U1 i, F1 F1 i endfor for
each instruction j in U2 if j s
predecessor s in DDG satisfies s M2
U2 U2 j, F2 F2 j endfor for
each pair i F1 and j F2 if( ij)
keep one copy (i or j) in the merged
sequence F1 F1 i, F2 F2 j,
M1 M1 i, M2 M2 j for all
instructions left in F1 and F2 do if-conversion
M1 M1 F1, M2 M2 F2, F1 Ø, F2
Ø endfor endwhile
M the merged instruction set, including those
that hve been compared and merged F the free
intruction set, including those whose
predecessors in DDG belong to M U unmerged
instruction set
It hurts the performance. In order to balance
the code size and the performance, we let code
merge work only when the coalesced basic block
contains at more n C instructions. (Assume
there are n instructions and m instructions in
original two basic blocks, and n m ) C 5 is
an empirical value.
15
Experimental Results
16
Relative Performance of Code Hoisting/Sinking and
Code Merge
17
Conclusions

• Global code hoisting/sinking tries to reduce the
  code size by eliminating identical instructions
  in multiple code path.
• Global Code merge tries to merge two basic blocks
  in diamond-shape using if-conversion, keeping
  only one copy for those identical instructions.
• Testing in kylincc shows a code size reduction up
  to 15, with an average of approximately 5.