Skip to contentSkip to conversation

ARM: Ensure that NEON code always compiles with Clang

PATCH Superseded Nathan Chancellor 6 participants 9 messages Dec 15, 2018 – Mar 11, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ARM: Ensure that NEON code always compiles with Clang

While building arm32 allyesconfig, I ran into the following errors:

arch/arm/lib/xor-neon.c:17:2: error: You should compile this file with '-mfloat-abi=softfp -mfpu=neon'

In file included from lib/raid6/neon1.c:27:
/home/nathan/cbl/prebuilt/lib/clang/8.0.0/include/arm_neon.h:28:2:
error: "NEON support not enabled"

Building V=1 showed NEON_FLAGS getting passed along to Clang but __ARM_NEON__ was not getting defined. Ultimately, it boils down to Clang only defining __ARM_NEON__ when targeting armv7, rather than armv6k, which is the '-march' value for allyesconfig.

From lib/Basic/Targets/ARM.cpp in the Clang source:

// This only gets set when Neon instructions are actually available, unlike // the VFP define, hence the soft float and arch check. This is subtly // different from gcc, we follow the intent which was that it should be set // when Neon instructions are actually available.

  if ((FPU & NeonFPU) && !SoftFloat && ArchVersion >= 7) {
    Builder.defineMacro("__ARM_NEON", "1");
    Builder.defineMacro("__ARM_NEON__");
    // current AArch32 NEON implementations do not support double-precision
    // floating-point even when it is present in VFP.
    Builder.defineMacro("__ARM_NEON_FP",
                        "0x" + Twine::utohexstr(HW_FP & ~HW_FP_DP));

}

Ard Biesheuvel recommended explicitly adding '-march=armv7-a' at the beginning of the NEON_FLAGS definitions so that __ARM_NEON__ always gets definined by Clang. This doesn't functionally change anything because that code will only run where NEON is supported, which is implicitly armv7.

---
 Documentation/arm/kernel_mode_neon.txt | 4 ++--
 arch/arm/lib/Makefile                  | 2 +-
 arch/arm/lib/xor-neon.c                | 2 +-
 lib/raid6/Makefile                     | 2 +-
 4 files changed, 5 insertions(+), 5 deletions(-)
Documentation/arm/kernel_mode_neon.txt +2 -2
index 525452726d31..b9e060c5b61e 100644
@@ -6,7 +6,7 @@ TL;DR summary
 * Use only NEON instructions, or VFP instructions that don't rely on support
   code
 * Isolate your NEON code in a separate compilation unit, and compile it with
-  '-mfpu=neon -mfloat-abi=softfp'
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp'
 * Put kernel_neon_begin() and kernel_neon_end() calls around the calls into your
   NEON code
 * Don't sleep in your NEON code, and be aware that it will be executed with
@@ -87,7 +87,7 @@ instructions appearing in unexpected places if no special care is taken.
 Therefore, the recommended and only supported way of using NEON/VFP in the
 kernel is by adhering to the following rules:
 * isolate the NEON code in a separate compilation unit and compile it with
-  '-mfpu=neon -mfloat-abi=softfp';
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp';
 * issue the calls to kernel_neon_begin(), kernel_neon_end() as well as the calls
   into the unit containing the NEON code from a compilation unit which is *not*
   built with the GCC flag '-mfpu=neon' set.
arch/arm/lib/Makefile +1 -1
index ad25fd1872c7..0bff0176db2c 100644
@@ -39,7 +39,7 @@ $(obj)/csumpartialcopy.o:	$(obj)/csumpartialcopygeneric.S
 $(obj)/csumpartialcopyuser.o:	$(obj)/csumpartialcopygeneric.S
 
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
-  NEON_FLAGS			:= -mfloat-abi=softfp -mfpu=neon
+  NEON_FLAGS			:= -march=armv7-a -mfloat-abi=softfp -mfpu=neon
   CFLAGS_xor-neon.o		+= $(NEON_FLAGS)
   obj-$(CONFIG_XOR_BLOCKS)	+= xor-neon.o
 endif
arch/arm/lib/xor-neon.c +1 -1
index a6741a895189..4600b62d845f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
 #endif
 
 /*
lib/raid6/Makefile +1 -1
index 2f8b61dfd9b0..bfec7c87c61e 100644
@@ -25,7 +25,7 @@ endif
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
 NEON_FLAGS := -ffreestanding
 ifeq ($(ARCH),arm)
-NEON_FLAGS += -mfloat-abi=softfp -mfpu=neon
+NEON_FLAGS += -march=armv7-a -mfloat-abi=softfp -mfpu=neon
 endif
 CFLAGS_recov_neon_inner.o += $(NEON_FLAGS)
 ifeq ($(ARCH),arm64)
Signature · 1 line
2.20.1
Acked-by on v1 1/1
#1 lore raw
Quoted 35 lines · Nathan Chancellor

While building arm32 allyesconfig, I ran into the following errors:

arch/arm/lib/xor-neon.c:17:2: error: You should compile this file with '-mfloat-abi=softfp -mfpu=neon'

In file included from lib/raid6/neon1.c:27:
/home/nathan/cbl/prebuilt/lib/clang/8.0.0/include/arm_neon.h:28:2:
error: "NEON support not enabled"

Building V=1 showed NEON_FLAGS getting passed along to Clang but __ARM_NEON__ was not getting defined. Ultimately, it boils down to Clang only defining __ARM_NEON__ when targeting armv7, rather than armv6k, which is the '-march' value for allyesconfig.

From lib/Basic/Targets/ARM.cpp in the Clang source:

// This only gets set when Neon instructions are actually available, unlike // the VFP define, hence the soft float and arch check. This is subtly // different from gcc, we follow the intent which was that it should be set // when Neon instructions are actually available.

  if ((FPU & NeonFPU) && !SoftFloat && ArchVersion >= 7) {
    Builder.defineMacro("__ARM_NEON", "1");
    Builder.defineMacro("__ARM_NEON__");
    // current AArch32 NEON implementations do not support double-precision
    // floating-point even when it is present in VFP.
    Builder.defineMacro("__ARM_NEON_FP",
                        "0x" + Twine::utohexstr(HW_FP & ~HW_FP_DP));

}

Ard Biesheuvel recommended explicitly adding '-march=armv7-a' at the beginning of the NEON_FLAGS definitions so that __ARM_NEON__ always gets definined by Clang. This doesn't functionally change anything because that code will only run where NEON is supported, which is implicitly armv7.

Quoted 3 lines

Did you test that this doesn't create issues with gcc e.g. complaints from the linker that objects have incompatible architecture specifications or similar annoyance? This already happened in the past but I forget the exact scenario. If you already did, or after you do validate with gcc as well, then you may add:

Quoted 72 lines
---
 Documentation/arm/kernel_mode_neon.txt | 4 ++--
 arch/arm/lib/Makefile                  | 2 +-
 arch/arm/lib/xor-neon.c                | 2 +-
 lib/raid6/Makefile                     | 2 +-
 4 files changed, 5 insertions(+), 5 deletions(-)
Documentation/arm/kernel_mode_neon.txt +2 -2
index 525452726d31..b9e060c5b61e 100644
@@ -6,7 +6,7 @@ TL;DR summary
 * Use only NEON instructions, or VFP instructions that don't rely on support
   code
 * Isolate your NEON code in a separate compilation unit, and compile it with
-  '-mfpu=neon -mfloat-abi=softfp'
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp'
 * Put kernel_neon_begin() and kernel_neon_end() calls around the calls into your
   NEON code
 * Don't sleep in your NEON code, and be aware that it will be executed with
@@ -87,7 +87,7 @@ instructions appearing in unexpected places if no special care is taken.
 Therefore, the recommended and only supported way of using NEON/VFP in the
 kernel is by adhering to the following rules:
 * isolate the NEON code in a separate compilation unit and compile it with
-  '-mfpu=neon -mfloat-abi=softfp';
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp';
 * issue the calls to kernel_neon_begin(), kernel_neon_end() as well as the calls
   into the unit containing the NEON code from a compilation unit which is *not*
   built with the GCC flag '-mfpu=neon' set.
arch/arm/lib/Makefile +1 -1
index ad25fd1872c7..0bff0176db2c 100644
@@ -39,7 +39,7 @@ $(obj)/csumpartialcopy.o:	$(obj)/csumpartialcopygeneric.S
 $(obj)/csumpartialcopyuser.o:	$(obj)/csumpartialcopygeneric.S
 
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
-  NEON_FLAGS			:= -mfloat-abi=softfp -mfpu=neon
+  NEON_FLAGS			:= -march=armv7-a -mfloat-abi=softfp -mfpu=neon
   CFLAGS_xor-neon.o		+= $(NEON_FLAGS)
   obj-$(CONFIG_XOR_BLOCKS)	+= xor-neon.o
 endif
arch/arm/lib/xor-neon.c +1 -1
index a6741a895189..4600b62d845f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
 #endif
 
 /*
lib/raid6/Makefile +1 -1
index 2f8b61dfd9b0..bfec7c87c61e 100644
@@ -25,7 +25,7 @@ endif
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
 NEON_FLAGS := -ffreestanding
 ifeq ($(ARCH),arm)
-NEON_FLAGS += -mfloat-abi=softfp -mfpu=neon
+NEON_FLAGS += -march=armv7-a -mfloat-abi=softfp -mfpu=neon
 endif
 CFLAGS_recov_neon_inner.o += $(NEON_FLAGS)
 ifeq ($(ARCH),arm64)
Signature · 1 line
2.20.1
#2 lore raw
Quoted 41 lines · Nicolas Pitre
Quoted 35 lines · Nathan Chancellor

While building arm32 allyesconfig, I ran into the following errors:

arch/arm/lib/xor-neon.c:17:2: error: You should compile this file with '-mfloat-abi=softfp -mfpu=neon'

In file included from lib/raid6/neon1.c:27:
/home/nathan/cbl/prebuilt/lib/clang/8.0.0/include/arm_neon.h:28:2:
error: "NEON support not enabled"

Building V=1 showed NEON_FLAGS getting passed along to Clang but __ARM_NEON__ was not getting defined. Ultimately, it boils down to Clang only defining __ARM_NEON__ when targeting armv7, rather than armv6k, which is the '-march' value for allyesconfig.

From lib/Basic/Targets/ARM.cpp in the Clang source:

// This only gets set when Neon instructions are actually available, unlike // the VFP define, hence the soft float and arch check. This is subtly // different from gcc, we follow the intent which was that it should be set // when Neon instructions are actually available.

  if ((FPU & NeonFPU) && !SoftFloat && ArchVersion >= 7) {
    Builder.defineMacro("__ARM_NEON", "1");
    Builder.defineMacro("__ARM_NEON__");
    // current AArch32 NEON implementations do not support double-precision
    // floating-point even when it is present in VFP.
    Builder.defineMacro("__ARM_NEON_FP",
                        "0x" + Twine::utohexstr(HW_FP & ~HW_FP_DP));

}

Ard Biesheuvel recommended explicitly adding '-march=armv7-a' at the beginning of the NEON_FLAGS definitions so that __ARM_NEON__ always gets definined by Clang. This doesn't functionally change anything because that code will only run where NEON is supported, which is implicitly armv7.

Quoted 3 lines
Quoted 9 lines

Did you test that this doesn't create issues with gcc e.g. complaints from the linker that objects have incompatible architecture specifications or similar annoyance? This already happened in the past but I forget the exact scenario. If you already did, or after you do validate with gcc as well, then you may add:

Hi Nicolas,

I was 99% sure that I checked GCC before sending this but I just did another run to confirm and everything links successfully. We still use binutils for assembling/linking the kernel so I assume that I would have seen a warning from ld.bfd even with Clang.

Thank you for the review!
Nathan

Quoted 73 lines
Quoted 73 lines
---
 Documentation/arm/kernel_mode_neon.txt | 4 ++--
 arch/arm/lib/Makefile                  | 2 +-
 arch/arm/lib/xor-neon.c                | 2 +-
 lib/raid6/Makefile                     | 2 +-
 4 files changed, 5 insertions(+), 5 deletions(-)
Documentation/arm/kernel_mode_neon.txt +2 -2
index 525452726d31..b9e060c5b61e 100644
@@ -6,7 +6,7 @@ TL;DR summary
 * Use only NEON instructions, or VFP instructions that don't rely on support
   code
 * Isolate your NEON code in a separate compilation unit, and compile it with
-  '-mfpu=neon -mfloat-abi=softfp'
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp'
 * Put kernel_neon_begin() and kernel_neon_end() calls around the calls into your
   NEON code
 * Don't sleep in your NEON code, and be aware that it will be executed with
@@ -87,7 +87,7 @@ instructions appearing in unexpected places if no special care is taken.
 Therefore, the recommended and only supported way of using NEON/VFP in the
 kernel is by adhering to the following rules:
 * isolate the NEON code in a separate compilation unit and compile it with
-  '-mfpu=neon -mfloat-abi=softfp';
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp';
 * issue the calls to kernel_neon_begin(), kernel_neon_end() as well as the calls
   into the unit containing the NEON code from a compilation unit which is *not*
   built with the GCC flag '-mfpu=neon' set.
arch/arm/lib/Makefile +1 -1
index ad25fd1872c7..0bff0176db2c 100644
@@ -39,7 +39,7 @@ $(obj)/csumpartialcopy.o:	$(obj)/csumpartialcopygeneric.S
 $(obj)/csumpartialcopyuser.o:	$(obj)/csumpartialcopygeneric.S
 
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
-  NEON_FLAGS			:= -mfloat-abi=softfp -mfpu=neon
+  NEON_FLAGS			:= -march=armv7-a -mfloat-abi=softfp -mfpu=neon
   CFLAGS_xor-neon.o		+= $(NEON_FLAGS)
   obj-$(CONFIG_XOR_BLOCKS)	+= xor-neon.o
 endif
arch/arm/lib/xor-neon.c +1 -1
index a6741a895189..4600b62d845f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
 #endif
 
 /*
lib/raid6/Makefile +1 -1
index 2f8b61dfd9b0..bfec7c87c61e 100644
@@ -25,7 +25,7 @@ endif
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
 NEON_FLAGS := -ffreestanding
 ifeq ($(ARCH),arm)
-NEON_FLAGS += -mfloat-abi=softfp -mfpu=neon
+NEON_FLAGS += -march=armv7-a -mfloat-abi=softfp -mfpu=neon
 endif
 CFLAGS_recov_neon_inner.o += $(NEON_FLAGS)
 ifeq ($(ARCH),arm64)
Signature · 1 line
2.20.1
Reviewed-by on v1 1/1
#3 lore raw
In reply to Nathan Chancellor
Quoted 36 lines · Nathan Chancellor

While building arm32 allyesconfig, I ran into the following errors:

arch/arm/lib/xor-neon.c:17:2: error: You should compile this file with '-mfloat-abi=softfp -mfpu=neon'

In file included from lib/raid6/neon1.c:27:
/home/nathan/cbl/prebuilt/lib/clang/8.0.0/include/arm_neon.h:28:2:
error: "NEON support not enabled"

Building V=1 showed NEON_FLAGS getting passed along to Clang but __ARM_NEON__ was not getting defined. Ultimately, it boils down to Clang only defining __ARM_NEON__ when targeting armv7, rather than armv6k, which is the '-march' value for allyesconfig.

From lib/Basic/Targets/ARM.cpp in the Clang source:

// This only gets set when Neon instructions are actually available, unlike // the VFP define, hence the soft float and arch check. This is subtly // different from gcc, we follow the intent which was that it should be set // when Neon instructions are actually available.

  if ((FPU & NeonFPU) && !SoftFloat && ArchVersion >= 7) {
    Builder.defineMacro("__ARM_NEON", "1");
    Builder.defineMacro("__ARM_NEON__");
    // current AArch32 NEON implementations do not support double-precision
    // floating-point even when it is present in VFP.
    Builder.defineMacro("__ARM_NEON_FP",
                        "0x" + Twine::utohexstr(HW_FP & ~HW_FP_DP));

}

Ard Biesheuvel recommended explicitly adding '-march=armv7-a' at the beginning of the NEON_FLAGS definitions so that __ARM_NEON__ always gets definined by Clang. This doesn't functionally change anything because that code will only run where NEON is supported, which is implicitly armv7.

Quoted 3 lines

Nathan,
Thanks for sending the patch, and thanks to Ard for the suggestion.

Quoted 71 lines
---
 Documentation/arm/kernel_mode_neon.txt | 4 ++--
 arch/arm/lib/Makefile                  | 2 +-
 arch/arm/lib/xor-neon.c                | 2 +-
 lib/raid6/Makefile                     | 2 +-
 4 files changed, 5 insertions(+), 5 deletions(-)
Documentation/arm/kernel_mode_neon.txt +2 -2
index 525452726d31..b9e060c5b61e 100644
@@ -6,7 +6,7 @@ TL;DR summary
 * Use only NEON instructions, or VFP instructions that don't rely on support
   code
 * Isolate your NEON code in a separate compilation unit, and compile it with
-  '-mfpu=neon -mfloat-abi=softfp'
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp'
 * Put kernel_neon_begin() and kernel_neon_end() calls around the calls into your
   NEON code
 * Don't sleep in your NEON code, and be aware that it will be executed with
@@ -87,7 +87,7 @@ instructions appearing in unexpected places if no special care is taken.
 Therefore, the recommended and only supported way of using NEON/VFP in the
 kernel is by adhering to the following rules:
 * isolate the NEON code in a separate compilation unit and compile it with
-  '-mfpu=neon -mfloat-abi=softfp';
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp';
 * issue the calls to kernel_neon_begin(), kernel_neon_end() as well as the calls
   into the unit containing the NEON code from a compilation unit which is *not*
   built with the GCC flag '-mfpu=neon' set.
arch/arm/lib/Makefile +1 -1
index ad25fd1872c7..0bff0176db2c 100644
@@ -39,7 +39,7 @@ $(obj)/csumpartialcopy.o:     $(obj)/csumpartialcopygeneric.S
 $(obj)/csumpartialcopyuser.o:  $(obj)/csumpartialcopygeneric.S
 
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
-  NEON_FLAGS                   := -mfloat-abi=softfp -mfpu=neon
+  NEON_FLAGS                   := -march=armv7-a -mfloat-abi=softfp -mfpu=neon
   CFLAGS_xor-neon.o            += $(NEON_FLAGS)
   obj-$(CONFIG_XOR_BLOCKS)     += xor-neon.o
 endif
arch/arm/lib/xor-neon.c +1 -1
index a6741a895189..4600b62d845f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
 #endif
 
 /*
lib/raid6/Makefile +1 -2
index 2f8b61dfd9b0..bfec7c87c61e 100644
@@ -25,7 +25,7 @@ endif
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
 NEON_FLAGS := -ffreestanding
 ifeq ($(ARCH),arm)
-NEON_FLAGS += -mfloat-abi=softfp -mfpu=neon
+NEON_FLAGS += -march=armv7-a -mfloat-abi=softfp -mfpu=neon
 endif
 CFLAGS_recov_neon_inner.o += $(NEON_FLAGS)
 ifeq ($(ARCH),arm64)
--

2.20.1

Signature · 2 lines
Thanks,
~Nick Desaulniers
#4 lore raw

PATCH RESEND 1/1 ARM: Ensure that NEON code always compiles with Clang

In reply to Nathan Chancellor

While building arm32 allyesconfig, I ran into the following errors:

arch/arm/lib/xor-neon.c:17:2: error: You should compile this file with '-mfloat-abi=softfp -mfpu=neon'

In file included from lib/raid6/neon1.c:27:
/home/nathan/cbl/prebuilt/lib/clang/8.0.0/include/arm_neon.h:28:2:
error: "NEON support not enabled"

Building V=1 showed NEON_FLAGS getting passed along to Clang but __ARM_NEON__ was not getting defined. Ultimately, it boils down to Clang only defining __ARM_NEON__ when targeting armv7, rather than armv6k, which is the '-march' value for allyesconfig.

Quoted 1 line

From lib/Basic/Targets/ARM.cpp in the Clang source:

// This only gets set when Neon instructions are actually available, unlike // the VFP define, hence the soft float and arch check. This is subtly // different from gcc, we follow the intent which was that it should be set // when Neon instructions are actually available.

  if ((FPU & NeonFPU) && !SoftFloat && ArchVersion >= 7) {
    Builder.defineMacro("__ARM_NEON", "1");
    Builder.defineMacro("__ARM_NEON__");
    // current AArch32 NEON implementations do not support double-precision
    // floating-point even when it is present in VFP.
    Builder.defineMacro("__ARM_NEON_FP",
                        "0x" + Twine::utohexstr(HW_FP & ~HW_FP_DP));

}

Ard Biesheuvel recommended explicitly adding '-march=armv7-a' at the beginning of the NEON_FLAGS definitions so that __ARM_NEON__ always gets definined by Clang. This doesn't functionally change anything because that code will only run where NEON is supported, which is implicitly armv7.

---

Resending with Nicolas's ack and Nick's review, to give others a chance to pitch in with review/testing before submitting it to the patch system (specifically Stefan, sorry I should have included you in the previous posting).

With this patch and:
* A tip of tree LLVM build (I used apt.llvm.org)
* Clang GCOV support: https://lore.kernel.org/lkml/[email protected]/
* Two minor hacks for unrelated issues that are still being worked through:
* https://raw.githubusercontent.com/nathanchance/patches/c313b2fa0efb/linux/build-hax/0003-DO-NOT-UPSTREAM-ARM-Don-t-select-HAVE_FUNCTION_TRACE.patch (see https://github.com/ClangBuiltLinux/linux/issues/35)
* https://gist.githubusercontent.com/nathanchance/b2f5a4015abade1a41e78d5fc3235c5b/raw/744321882ab05511331f26896bad7c9f0056a6a5/gistfile1.txt (see https://github.com/ClangBuiltLinux/linux/issues/325)

I can build and link a little endian allyesconfig ARM kernel. This patch alone works with GCC 8.2.0 and binutils 2.31.1.

 Documentation/arm/kernel_mode_neon.txt | 4 ++--
 arch/arm/lib/Makefile                  | 2 +-
 arch/arm/lib/xor-neon.c                | 2 +-
 lib/raid6/Makefile                     | 2 +-
 4 files changed, 5 insertions(+), 5 deletions(-)
Documentation/arm/kernel_mode_neon.txt +2 -2
index 525452726d31..b9e060c5b61e 100644
@@ -6,7 +6,7 @@ TL;DR summary
 * Use only NEON instructions, or VFP instructions that don't rely on support
   code
 * Isolate your NEON code in a separate compilation unit, and compile it with
-  '-mfpu=neon -mfloat-abi=softfp'
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp'
 * Put kernel_neon_begin() and kernel_neon_end() calls around the calls into your
   NEON code
 * Don't sleep in your NEON code, and be aware that it will be executed with
@@ -87,7 +87,7 @@ instructions appearing in unexpected places if no special care is taken.
 Therefore, the recommended and only supported way of using NEON/VFP in the
 kernel is by adhering to the following rules:
 * isolate the NEON code in a separate compilation unit and compile it with
-  '-mfpu=neon -mfloat-abi=softfp';
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp';
 * issue the calls to kernel_neon_begin(), kernel_neon_end() as well as the calls
   into the unit containing the NEON code from a compilation unit which is *not*
   built with the GCC flag '-mfpu=neon' set.
arch/arm/lib/Makefile +1 -1
index ad25fd1872c7..0bff0176db2c 100644
@@ -39,7 +39,7 @@ $(obj)/csumpartialcopy.o:	$(obj)/csumpartialcopygeneric.S
 $(obj)/csumpartialcopyuser.o:	$(obj)/csumpartialcopygeneric.S
 
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
-  NEON_FLAGS			:= -mfloat-abi=softfp -mfpu=neon
+  NEON_FLAGS			:= -march=armv7-a -mfloat-abi=softfp -mfpu=neon
   CFLAGS_xor-neon.o		+= $(NEON_FLAGS)
   obj-$(CONFIG_XOR_BLOCKS)	+= xor-neon.o
 endif
arch/arm/lib/xor-neon.c +1 -1
index 2c40aeab3eaa..c691b901092f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
 #endif
 
 /*
lib/raid6/Makefile +1 -1
index 4e90d443d1b0..e723eacf7868 100644
@@ -39,7 +39,7 @@ endif
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
 NEON_FLAGS := -ffreestanding
 ifeq ($(ARCH),arm)
-NEON_FLAGS += -mfloat-abi=softfp -mfpu=neon
+NEON_FLAGS += -march=armv7-a -mfloat-abi=softfp -mfpu=neon
 endif
 CFLAGS_recov_neon_inner.o += $(NEON_FLAGS)
 ifeq ($(ARCH),arm64)
Signature · 1 line
2.20.1
Reviewed-by on v1 1/1
#5 lore raw
Quoted 40 lines · Nathan Chancellor

While building arm32 allyesconfig, I ran into the following errors:

arch/arm/lib/xor-neon.c:17:2: error: You should compile this file with '-mfloat-abi=softfp -mfpu=neon'

In file included from lib/raid6/neon1.c:27:
/home/nathan/cbl/prebuilt/lib/clang/8.0.0/include/arm_neon.h:28:2:
error: "NEON support not enabled"

Building V=1 showed NEON_FLAGS getting passed along to Clang but __ARM_NEON__ was not getting defined. Ultimately, it boils down to Clang only defining __ARM_NEON__ when targeting armv7, rather than armv6k, which is the '-march' value for allyesconfig.

Quoted 1 line

From lib/Basic/Targets/ARM.cpp in the Clang source:

// This only gets set when Neon instructions are actually available, unlike // the VFP define, hence the soft float and arch check. This is subtly // different from gcc, we follow the intent which was that it should be set // when Neon instructions are actually available.

  if ((FPU & NeonFPU) && !SoftFloat && ArchVersion >= 7) {
    Builder.defineMacro("__ARM_NEON", "1");
    Builder.defineMacro("__ARM_NEON__");
    // current AArch32 NEON implementations do not support double-precision
    // floating-point even when it is present in VFP.
    Builder.defineMacro("__ARM_NEON_FP",
                        "0x" + Twine::utohexstr(HW_FP & ~HW_FP_DP));

}

Ard Biesheuvel recommended explicitly adding '-march=armv7-a' at the beginning of the NEON_FLAGS definitions so that __ARM_NEON__ always gets definined by Clang. This doesn't functionally change anything because that code will only run where NEON is supported, which is implicitly armv7.

Quoted 6 lines

---

Resending with Nicolas's ack and Nick's review, to give others a chance to pitch in with review/testing before submitting it to the patch system (specifically Stefan, sorry I should have included you in the previous posting).

No worries.

Looks good to me.

--
Stefan

Quoted 88 lines

With this patch and:
* A tip of tree LLVM build (I used apt.llvm.org)
* Clang GCOV support:
https://lore.kernel.org/lkml/[email protected]/
* Two minor hacks for unrelated issues that are still being worked through:
*
https://raw.githubusercontent.com/nathanchance/patches/c313b2fa0efb/linux/build-hax/0003-DO-NOT-UPSTREAM-ARM-Don-t-select-HAVE_FUNCTION_TRACE.patch
(see https://github.com/ClangBuiltLinux/linux/issues/35)
*
https://gist.githubusercontent.com/nathanchance/b2f5a4015abade1a41e78d5fc3235c5b/raw/744321882ab05511331f26896bad7c9f0056a6a5/gistfile1.txt
(see https://github.com/ClangBuiltLinux/linux/issues/325)

I can build and link a little endian allyesconfig ARM kernel. This patch alone works with GCC 8.2.0 and binutils 2.31.1.

 Documentation/arm/kernel_mode_neon.txt | 4 ++--
 arch/arm/lib/Makefile                  | 2 +-
 arch/arm/lib/xor-neon.c                | 2 +-
 lib/raid6/Makefile                     | 2 +-
 4 files changed, 5 insertions(+), 5 deletions(-)
Documentation/arm/kernel_mode_neon.txt +1 -1
index 525452726d31..b9e060c5b61e 100644
@@ -6,7 +6,7 @@ TL;DR summary
 * Use only NEON instructions, or VFP instructions that don't rely on support
   code
 * Isolate your NEON code in a separate compilation unit, and compile it with
-  '-mfpu=neon -mfloat-abi=softfp'
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp'
 * Put kernel_neon_begin() and kernel_neon_end() calls around the

calls into your
NEON code
* Don't sleep in your NEON code, and be aware that it will be executed with

Documentation/arm/kernel_mode_neon.txt +0 -0
@@ -87,7 +87,7 @@ instructions appearing in unexpected places if no

special care is taken.
Therefore, the recommended and only supported way of using NEON/VFP in the kernel is by adhering to the following rules:
* isolate the NEON code in a separate compilation unit and compile it with

Documentation/arm/kernel_mode_neon.txt +1 -1
-  '-mfpu=neon -mfloat-abi=softfp';
+  '-march=armv7-a -mfpu=neon -mfloat-abi=softfp';
 * issue the calls to kernel_neon_begin(), kernel_neon_end() as well

as the calls
into the unit containing the NEON code from a compilation unit which is *not* built with the GCC flag '-mfpu=neon' set.

arch/arm/lib/Makefile +1 -1
index ad25fd1872c7..0bff0176db2c 100644
@@ -39,7 +39,7 @@ $(obj)/csumpartialcopy.o:	$(obj)/csumpartialcopygeneric.S
 $(obj)/csumpartialcopyuser.o:	$(obj)/csumpartialcopygeneric.S
 
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
-  NEON_FLAGS			:= -mfloat-abi=softfp -mfpu=neon
+  NEON_FLAGS			:= -march=armv7-a -mfloat-abi=softfp -mfpu=neon
   CFLAGS_xor-neon.o		+= $(NEON_FLAGS)
   obj-$(CONFIG_XOR_BLOCKS)	+= xor-neon.o
 endif
arch/arm/lib/xor-neon.c +1 -2
index 2c40aeab3eaa..c691b901092f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a
-mfloat-abi=softfp -mfpu=neon'
 #endif
 
 /*
lib/raid6/Makefile +1 -1
index 4e90d443d1b0..e723eacf7868 100644
@@ -39,7 +39,7 @@ endif
 ifeq ($(CONFIG_KERNEL_MODE_NEON),y)
 NEON_FLAGS := -ffreestanding
 ifeq ($(ARCH),arm)
-NEON_FLAGS += -mfloat-abi=softfp -mfpu=neon
+NEON_FLAGS += -march=armv7-a -mfloat-abi=softfp -mfpu=neon
 endif
 CFLAGS_recov_neon_inner.o += $(NEON_FLAGS)
 ifeq ($(ARCH),arm64)
#6 lore raw
In reply to Nathan Chancellor

arch/arm/lib/xor-neon.c line 19

Quoted 13 lines · Nathan Chancellor

endif

arch/arm/lib/xor-neon.c +1 -1
index a6741a895189..4600b62d845f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
 #endif
 

I see this patch has made it in now, but I also see two other problems with the same file that prevent it from working right with clang:

- it triggers #warning This code requires at least version 4.6 of GCC
- As I reported in https://bugs.llvm.org/show_bug.cgi?id=40976, even when it builds cleanly, it does not get vectorized.

Has anyone actually managed to get this to do the right thing?

Arnd

#7 lore raw
Quoted 20 lines · Arnd Bergmann

arch/arm/lib/xor-neon.c line 19

Quoted 13 lines · Nathan Chancellor

endif

arch/arm/lib/xor-neon.c +1 -1
index a6741a895189..4600b62d845f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
 #endif
 

I see this patch has made it in now, but I also see two other problems with the same file that prevent it from working right with clang:

Quoted 1 line

- it triggers #warning This code requires at least version 4.6 of GCC

What is currently the oldest GCC we support for ARM?

Quoted 5 lines

- As I reported in https://bugs.llvm.org/show_bug.cgi?id=40976, even when it builds cleanly, it does not get vectorized.

Has anyone actually managed to get this to do the right thing?

On my Cortex-A57 under KVM, I get this at boot

[    0.002287] xor: measuring software checksum speed
[    0.100054]    arm4regs  :  5212.800 MB/sec
[    0.200131]    8regs     :  3472.800 MB/sec
[    0.300205]    32regs    :  3282.000 MB/sec
[    0.400281]    neon      :  7011.600 MB/sec

So that means that
a) the cost model is inaccurate, at least for some cores,
b) given that the code is only selected if it is faster, it would be nice if we could override the cost model based decisions made by the vectorizer.

#8 lore raw
Quoted 22 lines · Ard Biesheuvel
Quoted 20 lines · Arnd Bergmann

arch/arm/lib/xor-neon.c line 19

Quoted 13 lines · Nathan Chancellor

endif

arch/arm/lib/xor-neon.c +1 -1
index a6741a895189..4600b62d845f 100644
@@ -14,7 +14,7 @@
 MODULE_LICENSE("GPL");
 
 #ifndef __ARM_NEON__
-#error You should compile this file with '-mfloat-abi=softfp -mfpu=neon'
+#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
 #endif
 

I see this patch has made it in now, but I also see two other problems with the same file that prevent it from working right with clang:

Quoted 3 lines
Quoted 1 line

- it triggers #warning This code requires at least version 4.6 of GCC

What is currently the oldest GCC we support for ARM?

Linux overall requires gcc-4.6, so we could just as well drop this check, good point.

Arnd