Skip to contentSkip to conversation

arm64: Fix compiling with ancient compiler

PATCH Superseded Andrew Jones 2 participants 3 messages Feb 1, 2022 – Feb 2, 2022

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 arm64: Fix compiling with ancient compiler

When compiling with an ancient compiler (gcc-4.8.5-36.el7_6.2.aarch64) the build fails with

lib/libcflat.a(alloc.o): In function `mult_overflow':
/home/drjones/kvm-unit-tests/lib/alloc.c:19: undefined reference to `__multi3'

According to kernel commit fb8722735f50 ("arm64: support __int128 on gcc 5+") GCC5+ will not emit __multi3 for __int128 multiplication, so let's just fallback to the non-__int128 overflow check when we use gcc versions older than 5.

---
 lib/alloc.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
lib/alloc.c +2 -1
index f4266f5d064e..70228aa32c6c 100644
@@ -1,6 +1,7 @@
 #include "alloc.h"
 #include "asm/page.h"
 #include "bitops.h"
+#include <linux/compiler.h>
 
 void *malloc(size_t size)
 {
@@ -13,7 +14,7 @@ static bool mult_overflow(size_t a, size_t b)
 	/* 32 bit system, easy case: just use u64 */
 	return (u64)a * (u64)b >= (1ULL << 32);
 #else
-#ifdef __SIZEOF_INT128__
+#if defined(__SIZEOF_INT128__) && (!defined(__aarch64__) || GCC_VERSION >= 50000)
 	/* if __int128 is available use it (like the u64 case above) */
 	unsigned __int128 res = a;
 	res *= b;
Signature · 1 line
2.34.1
Acked-by on v1 1/1
#1 lore raw
Quoted 29 lines · Andrew Jones

When compiling with an ancient compiler (gcc-4.8.5-36.el7_6.2.aarch64) the build fails with

lib/libcflat.a(alloc.o): In function `mult_overflow':
/home/drjones/kvm-unit-tests/lib/alloc.c:19: undefined reference to `__multi3'

According to kernel commit fb8722735f50 ("arm64: support __int128 on gcc 5+") GCC5+ will not emit __multi3 for __int128 multiplication, so let's just fallback to the non-__int128 overflow check when we use gcc versions older than 5.

---
  lib/alloc.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
lib/alloc.c +1 -0
index f4266f5d064e..70228aa32c6c 100644
@@ -1,6 +1,7 @@
  #include "alloc.h"
  #include "asm/page.h"
  #include "bitops.h"
+#include <linux/compiler.h>
 
  void *malloc(size_t size)
  {

lib/alloc.c static bool mult_overflow(size_t a, size_t b) · line 20

Quoted 8 lines
lib/alloc.c +1 -1
@@ -13,7 +14,7 @@ static bool mult_overflow(size_t a, size_t b)
  	/* 32 bit system, easy case: just use u64 */
  	return (u64)a * (u64)b >= (1ULL << 32);
  #else
-#ifdef __SIZEOF_INT128__
+#if defined(__SIZEOF_INT128__) && (!defined(__aarch64__) || GCC_VERSION >= 50000)
  	/* if __int128 is available use it (like the u64 case above) */
  	unsigned __int128 res = a;
  	res *= b;

I'd also be OK if we'd simply declare GCC compiler versions < 5 as unsupported ... but since it's an easy fix:

#2 lore raw
Quoted 32 lines · Thomas Huth
Quoted 31 lines · Andrew Jones

When compiling with an ancient compiler (gcc-4.8.5-36.el7_6.2.aarch64) the build fails with

lib/libcflat.a(alloc.o): In function `mult_overflow':
/home/drjones/kvm-unit-tests/lib/alloc.c:19: undefined reference to `__multi3'

According to kernel commit fb8722735f50 ("arm64: support __int128 on gcc 5+") GCC5+ will not emit __multi3 for __int128 multiplication, so let's just fallback to the non-__int128 overflow check when we use gcc versions older than 5.

---
  lib/alloc.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
lib/alloc.c +1 -0
index f4266f5d064e..70228aa32c6c 100644
@@ -1,6 +1,7 @@
  #include "alloc.h"
  #include "asm/page.h"
  #include "bitops.h"
+#include <linux/compiler.h>
  void *malloc(size_t size)
  {
@@ -13,7 +14,7 @@ static bool mult_overflow(size_t a, size_t b)
  	/* 32 bit system, easy case: just use u64 */
  	return (u64)a * (u64)b >= (1ULL << 32);
  #else
Quoted 8 lines

lib/alloc.c static bool mult_overflow(size_t a, size_t b) · line 20

Quoted 5 lines
lib/alloc.c +1 -1
-#ifdef __SIZEOF_INT128__
+#if defined(__SIZEOF_INT128__) && (!defined(__aarch64__) || GCC_VERSION >= 50000)
  	/* if __int128 is available use it (like the u64 case above) */
  	unsigned __int128 res = a;
  	res *= b;

I'd also be OK if we'd simply declare GCC compiler versions < 5 as unsupported ... but since it's an easy fix:

I'd like to continue supporting this particular old compiler, because I only have one machine where I can run the unit tests with KVM AArch32, and on that machine I still have an old RHEL (RHEL7-Alt) installed. If it gets too difficult to support, then I'll try upgrading that machine's RHEL version or at least its compiler version.

Quoted 3 lines

Thanks, but I need a v2 since I just remembered clang. For v2, I see two options:

1) Just change the line as below and hope we don't need to worry about clang versions

#if defined(__SIZEOF_INT128__) && (!defined(__aarch64__) || defined(__clang__) || GCC_VERSION >= 50000)

2) Replace lib/alloc.c's mult_overflow() with lib/linux/compiler.h's
check_mul_overflow()

(2) seems like the better choice except for the fact that it lazily doesn't implement a fallback for compilers which don't support the builtins (and that's older than GCC7.1). Do we care? Of course I could combine the two implementations if we really want that fallback and the cleanup.

Thanks,
drew