BTW: I've forgot to mention: hard versus soft float abis are well
detected by configure itself. So this is not the problem in my patch.
The only issue is in testing/guessing for FPU/NEON. Good way may be to
grep /proc/cpuinfo but this is only applicable while doing native
compile on ARM-Linux platform. Is that what majority of GHC users will
do? I don't think so as I think the way for majority will be
cross-compilation but perhaps we can at least try it for those doing
native compiles? What do you think about it?
Thanks!
Karel
On 01/25/13 09:59 PM, Karel Gardas wrote:
On 01/25/13 09:38 PM, Simon Marlow wrote:
This would improve things for RPi, but we shouldn't be guessing here.
gcc has predefines for the appropriate features:
#define __ARM_PCS_VFP 1
indicates hard-float ABI, and
#define __VFP_FP__ 1
indicates VFP presences (2 or 3? I don't know)
The problem is that IIRC those defines are presented when you use
appropriate GCC's command-line option. It does not tell anything about
the target hardware itself. So that's why I'm just guessing in the
patch. But honestly speaking do you think there is any other GHC user
who would use GHC on ARMv6 and the hardware will be different than RPi?
I think a chance for this is very low since otherwise everybody is using
ARMv7 these days...
Anyway, if you don't like guessing, what about to add few configure
options which will drive the generation of ARM related bits into
settings file and keeping some good defaults like we use now? i.e.
ARMv7+VFPv3+NEON or ARMv6+VFPv2 or ARMv5 alone?
Thanks!
Karel
Cheers,
Simon
On 25/01/13 15:56, Karel Gardas wrote:
This patch adds support for Raspberry Pi board. The board is based
on Broadcom BCM2835, which includes ARM1176JZF-S CPU core. The core
is classical ARM11 so implementing ARMv6 ISA, but also provides VFPv2
FPU
unit. As this looks like fairly popular hardware, in this patch we
assume
detected ARMv6 hardware to support VFPv2. Also as Raspberry Pi is very
commonly
used with hard-float ABI Linux distribution, the patch relaxes a
condition
of "hard-float support for ARMv7 only" to any general ARM core.
---
aclocal.m4 | 2 +-
compiler/main/DriverPipeline.hs | 9 ++++++---
2 files changed, 7 insertions(+), 4 deletions(-)
diff --git a/aclocal.m4 b/aclocal.m4
index edbe389..b23568f 100644
--- a/aclocal.m4
+++ b/aclocal.m4
@@ -369,7 +369,7 @@ AC_DEFUN([GET_ARM_ISA],
[AC_DEFINE(arm_HOST_ARCH_PRE_ARMv7, 1, [ARM pre v7])
changequote(, )dnl
ARM_ISA=ARMv6
- ARM_ISA_EXT="[]"
+ ARM_ISA_EXT="[VFPv2]"
changequote([, ])dnl
],
[changequote(, )dnl
diff --git a/compiler/main/DriverPipeline.hs
b/compiler/main/DriverPipeline.hs
index c24bb51..81d0bc0 100644
--- a/compiler/main/DriverPipeline.hs
+++ b/compiler/main/DriverPipeline.hs
@@ -1454,14 +1454,17 @@ runPhase LlvmLlc input_fn dflags
else if (elem VFPv3D16 ext)
then ["-mattr=+v7,+vfp3,+d16"]
else []
+ ArchARM ARMv6 ext _ -> if (elem VFPv2 ext)
+ then ["-mattr=+v6,+vfp2"]
+ else ["-mattr=+v6"]
_ -> []
-- On Ubuntu/Debian with ARM hard float ABI, LLVM's llc still
-- compiles into soft-float ABI. We need to explicitly set abi
-- to hard
abiOpts = case platformArch (targetPlatform dflags) of
- ArchARM ARMv7 _ HARD -> ["-float-abi=hard"]
- ArchARM ARMv7 _ _ -> []
- _ -> []
+ ArchARM _ _ HARD -> ["-float-abi=hard"]
+ ArchARM _ _ _ -> []
+ _ -> []
sseOpts | isSse4_2Enabled dflags = ["-mattr=+sse42"]
| isSse2Enabled dflags = ["-mattr=+sse2"]
_______________________________________________
ghc-devs mailing list
ghc-devs@haskell.org
http://www.haskell.org/mailman/listinfo/ghc-devs
_______________________________________________
ghc-devs mailing list
ghc-devs@haskell.org
http://www.haskell.org/mailman/listinfo/ghc-devs
