[PATCH] add better support for Raspberry Pi board
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"] -- 1.7.3.2
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) 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"]
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"]
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
I've finished this now. I committed the second part of your patch (the DriverPipeline one), fixed up the "detection" of hard-float, which was based on the triple passed to --target but needed some tweaking, and I've made it detect VFPv2 from the C compiler. We should probably be doing the hard-float detection the same way. So now you don't need any special -optlc options for registerised compilation with LLVM on Raspberry Pi, it just works. Cheers, Simon On 25/01/13 21:23, Karel Gardas wrote:
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
Simon, thanks for this. I've done this patch originally based on the complains on haskell-user mailing list about compilation to RPi. Otherwise I'm not that found of this hardware class, but rather prefer working on ARMv7 or even ARMv8. Anyway, thanks! Karel On 01/29/13 11:27 AM, Simon Marlow wrote:
I've finished this now. I committed the second part of your patch (the DriverPipeline one), fixed up the "detection" of hard-float, which was based on the triple passed to --target but needed some tweaking, and I've made it detect VFPv2 from the C compiler. We should probably be doing the hard-float detection the same way.
So now you don't need any special -optlc options for registerised compilation with LLVM on Raspberry Pi, it just works.
Cheers, Simon
On 25/01/13 21:23, Karel Gardas wrote:
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
participants (2)
-
Karel Gardas -
Simon Marlow