https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68198
Bug ID: 68198
Summary: Excessive code size, compile time and memory usage
bloat due to FSM threading in 453.povray
Product: gcc
Version: 6.0
Status: UNCONFIRMED
Keywords: compile-time-hog, memory-hog
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
CC: law at redhat dot com
Target Milestone: ---
Target: arm, aarch64
Created attachment 36642
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36642&action=edit
-fmem-report output
I see a large compile-time regression building 453.povray from SPEC2006 on arm
and aarch64.
The file in question is fnpovfpu.cpp.
Try compiling it with:
arm-none-linux-gnueabihf-g++ -DSPEC_CPU -DNDEBUG -Ofast -mcpu=cortex-a57
-mfpu=neon-vfpv4 -mfloat-abi=hard -mthumb -fno-strict-aliasing -w fnpovfpu.cpp
A gcc from around end of Sep 23rd takes about 1.4s to compile the file on my
machine.
Recent gcc at r229696 takes about 6 and a half minutes and consumes a peak of
9.1GB of memory on an x86_64 host.
I'm attaching the -fmem-report I get.
The offending function pov::POVFPU_RunDefault is a large switch statement
inside a while (true), so I'm suspecting the FSM threading code.
The resulting assembly code on arm is so large that it doesn't assemble with
many "branch offset out of range" errors.
I'm seeing a similar time and space explosion on aarch64.
Jeff, could you please take a look?
