| - 2004-09-14
: No data for -Os between 2004-09-14 and 2004-10-16 because of a GCC bug (PR17560)
For more info see the following bug reports: PR17560, PR17607, PR17883, and PR17704.
- 2004-08-07
: Code size increase due to a patch for cross-jump handling.
A target independent code size increase of about 1% was observed when the patch that replaces a zero-nonzero check of matching instructions with a minimum threshold check in cross-jump handling was committed. For more info see [PATCH] Add minimum insn match to crossjump.
- 2004-05-13
: Tree SSA merge decreased code size and increased compilation time
Code size decrease of about 0.5% was observed when the tree-ssa-20020619-branch was merged into mainline. For more info see SSA for Trees.
- 2004-01-27
: Significant code size decrease on i386 (alignment of long string literals)
An i386 patch saves about 2.5% of code size (see the changes on i386). The patch decreases the alignment of long string literals from 32 bytes to sizeof (void *) when not optimizing for size and to 1 with -Os. For more info see the corresponding [PATCH] Decrease excessive string literal alignment on IA-32/AMD64 mail.
- 2004-01-14
: A minor code size increase on ARM and mips
This cgraphunit patch made a small code size increase on arm and mips with -Os (about 0.15%). The patch changed the ggc_collect condition level in cgraph_finalize_function. For more info see the corresponding [C++ patch] PR c++/12850 mail.
- 2004-01-05 - 2004-01-09
: Code size increase due to typo
The Updated version of -Winline warnings patch increased the code size for some targets due to typo (see, for example, the code size change for m68k and for mips). The problem was fixed.
- 2004-01-02
: Patch saves code size on ARM with -Os but introduces bootstrap failure
Some of the commited patches PR ice/13473, PR opt/12617, and/or PR debug/13367 introduced a bootstrap failure on ARM with -O2, and at the same time code size saves with -Os can be measured (see the changes on arm). For more info about the bootstrap failure see the New bootstrap failure on ARM systems mail.
- 2003-11-19
: Minor code size saves on ARM when optimizing for size
A small code size save has been observed for the ARM target, which was caused by a patch that improves the estimate of the code size for calls to libgcc's div & mod subroutines when optimizing for space. The PR middle-end/11821: Tweak arm_rtx_costs_1 mail contains the description and the patch. See the problem report for more info.
- 2003-10-21 - 2003-10-30
: Significant code size increase on ARM when optimizing for size due to Factorization of constants patch
A significant code size increase has been observed for the ARM target, which was caused by a patch that allows factorization of constants into addressing insns when optimizing for space. The ARM: allow factorization of constants into addressing insns when optimizing for space mail contains the description and the patch.
The patch has been reverted because of the Re: ARM: allow factorization of constants into addressing insns when optimizing for space mail.
- 2003-10-18 - 2003-10-28
: Significant code size increase on ARM due to Protect stack pointer patch
A significant code size increase has been observed for the ARM target, which was caused by a patch that protects the stack pointer from being corrupted on interrupt. The [ARM PATCH] Protect stack pointer mail contains the description of the problem.
The problem essentially disappeared after the optimize frame-pointer patch. For more information see the corresponding mail.
- 2003-10-12 - 2003-10-15
: Side effect of a patch
The [PATCH] Constant folding NEGATE_EXPR clean-ups patch increased the code size for all targets (see, for example, the code size increase for arm and for mips). The [PATCH] disable tree_swap_operands_p when optimizing for size in fold-const.c mail contains the description of the problem and a patch which disables some features when optimizing for size.
- 2003-10-11
: Code size increase on m68k due to interrupt_handler patch
The interrupt_handler - rev. 4 patch contained some improvements for the m68k target, but it also increased the code size by about 1%. The [m68k] Add __attribute__((interrupt_handler)) - rev. 4 mail contains the description of the problem.
- 2003-09-18 - 2003-09-25
: A minor code size increase for a short period
The [PATCH] Constant folding clean-ups patch contained some clean-ups in GCC's constant folding, but it also had size drawbacks (see, for example, the code size increase for arm and for mips). The [PATCH] PR bootstrap/12358: Operand reordering problems patch fixed the problem.
- 2003-09-04 - 2003-09-05
: Enabling unit-at-a-time in mainline resulted in major code size improvement for most targets
The unit-at-a-time flag has been made default in GCC mainline, which saved about 1% of code for arm and mips, and which produced a drawback of 0.4% for i386. The [RFC/RFA?] Enabling unit-at-a-time for -O2 and increasing inline limits mail contains the description of the patch.
- 2003-08-31 - 2003-09-01
: A patch for optimizing large jump tables benefited from CSiBE statistics
The condition for generating jump tables from switch statements has been modified to include the case when optimizing for size, which caused this arm benefit and this i386 benefit. The threshold value has been determined based on the CSiBE statistics. This Problem Report motivated the [PATCH] Optimizing large jump tables for switch statements patch.
Comment: a more sophisticated algorithm may be more effective that computes a cost/benefit value to decide when should a jump table be created.
- 2003-06-27 - 2003-06-29
: A problem with a patch caused code size changes for a short period
Whereas the [switch] Move more to opts.c from toplev.c patch reduced arm code size, it enlarged mips code. The [PATCH] Re-enable scheduler patch fixed the problem.
|
| Help |
Each observation lists the effective date, eventually the patch
that caused the change, diagram(s) showing the effects, and some detailed info.
|
|
|
