SDK/Toolchains/ToolchainChange

Revision as of 12:07, 8 February 2011

Toolchain Change in MeeGo 1.2

ARM Toolchain will be changed in MeeGo release 1.2. This page describes the change and the effects of the change.

Background

This change effects to ARM cross compiler part of MeeGo toolchain.

The toolchain in current MeeGo 1.1 supports ARM architecture only partially what comes to Floating Point Unit (FPU) support.

gcc compiler provides three options to support floating points in option -mfloat-abi [1]:

• soft - Full software implementation of floating points. Fits the processors without the FPU. All floating point operations are emulated by the compiler
• softfp - FPU unit is in use but the code is compatible to full software code. Floating Point operations are passed to the FPU in integer registers
• hard - Floating Point arguments are passed directly to the FPU registers. No compatibility to soft and softfp.

Full utilization of the FPU gives performance improvement to applications using plenty of floating point operations - like games, location based services and graphics.

In order to take -mfloat-abi=hard option into use in MeeGo 1.2 the toolchain is needed to be updated to gcc-4.5.1 based.

The change proposal (SDK/Toolchains/ToolchainChangeProposal) was approved in MeeGo TSG on December 1, 2010.

Change Content

In MeeGo Release 1.2 and onwards the following is about to change:

• The toolchain is based on gcc-4.5.1 with binutils-2.20 (or newer) and glibc-2.10 (or newer)
• gcc option -mfloat-abi=hard is the default and mandatory
• There are two build configurations in the toolchain - Baseline and Optimized configurations. Baseline configuration maximize the compatibility and does not allow the use of NEON instructions nor Thumb2 instruction set. Optimized configuration allows the use of NEON instructions and Thumb2 instruction set.
• Third set is meant for N900 reference hardware. Since there is silicon bug in Cortex-A8 (see http://lists.meego.com/pipermail/meego-dev/2011-January/481145.html)

• Baseline configuration:

  * RPM architecture name: "armv7hl"
  * ARMv7 architecture (-march=armv7-a)
  * VFPv3-D16 is in use (-mfpu=vfpv3-d16)
  * Thumb2 instruction set not used (-mno-thumb)

• Optimized configuration (NEON and Thumb2 can be used)

  * RPM architecture name: "armv7tnhl"
  * ARMv7 architecture (-march=armv7-a)
  * NEON allowed (-mfpu=neon)
  * Thumb2 instruction set allowed (-mthumb)

• Configuration for N900 reference device:

  * RPM architecture name: "armv7nhl"
  * ARMv7 architecture (-march=armv7-a)
  * NEON allowed (-mfpu=neon)
  * Thumb2 instruction set not used (-mno-thumb)

• ARM RunFast is enabled in glibc by default

Impact of the Change

ARM Packages compiled in MeeGo Release 1.2 are not binary compatible with those packages that are compiled using gcc-option -mfloat-abi=softfp in MeeGo Release 1.0 or 1.1. So all the packages from Release 1.0 or 1.1 must be re-compiled using the toolchain of MeeGo 1.2.

Bugzilla

All hardfp related issues are tracked in the metabug #11429 (http://bugs.meego.com/showdependencytree.cgi?id=11429&hide_resolved=0)

References

[1] http://wiki.debian.org/ArmHardFloatPort/VfpComparison#DetailsonGCCfloating-pointoptions