[PATCH] D30651: [InlineCost, -Oz] Don't take into account the penalty of a fast call of frequently used functions

[Evgeny Astigeevich via Phabricator via llvm-commits]

eastig created this revision.
Herald added subscribers: eraman, aemerson.

This patch introduces an inline heuristic to be used whilst the cost of inlining is calculated for calls at '-Oz' optimization level.

When the cost of inlining is calculated for a call which has the fast calling convention and it is a call of a frequently used function, the penalty of the call is not taken into account. By default a function is frequently used if a number of its users is greater than 3. This value can be change via the command-line option '-inline-freq-func-threshold'. The heuristic allows inlining of very small functions, like a few instructions.

**The rationale for the heuristic:**

1. Fast calls are usually optimized to minimize the call penalty, e.g. parameters are passed through registers.
2. Inlining of frequently used functions increases the code size.
3. It's based on real examples of code for microcontrollers (v6-M). We see an improvement of the code size of the examples ~6.5%.

**Benchmark results of the LNT testsuite:**

Code size of only three benchmarks were affected by the heuristic.

x86 (i7-4770):

|                                                  | Code size improvement | Perf regression |
| MultiSource/Applications/SPASS/SPASS             | 6.67%                 | ~0%             |
| MultiSource/Benchmarks/FreeBench/distray/distray | 2.46%                 | ~0%             |
| MultiSource/Applications/sqlite3/sqlite3         | 1.34%                 | ~0%             |

v7-A (Cortex-A53, Thumb)

|                                          | Code size improvement | Perf regression |
| MultiSource/Applications/SPASS/SPASS     | 4.93%                 | ~3%             |
| MultiSource/Applications/sqlite3/sqlite3 | 1.12%                 | ~1.5%           |

AArch64 (Cortex-A57)

|                                                  | Code size improvement | Perf regression |
| MultiSource/Applications/SPASS/SPASS             | 3.95%                 | ~3%             |
| MultiSource/Benchmarks/FreeBench/distray/distray | 1.17%                 | ~0%             |
| MultiSource/Applications/sqlite3/sqlite3         | 1.07%                 | ~0%             |


https://reviews.llvm.org/D30651

Files:
  include/llvm/Analysis/InlineCost.h
  lib/Analysis/InlineCost.cpp
  test/Transforms/Inline/minsize-fastcc-frq-func.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D30651.90686.patch
Type: text/x-patch
Size: 5850 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170306/46efd1bb/attachment.bin>