Skip to contentSkip to conversation

Instruction set detection for clang.

PATCH In review Shu-Chun Weng 2 participants 8 messages Jul 23, 2020 – Sep 17, 2020

lore
Share & tools
lore
#0 lore raw

cover Instruction set detection for clang.

Currently when configuring QEMU with clang, AVX2, AVX512F, ATOMIC64, and ATOMIC128 are all disabled because the detection code is GCC-only. With these two patches, I am able to configure, build, and run tests with clang with all of the above enabled.

Shu-Chun Weng (2):
configure: avx2 and avx512f detection for clang
configure: atomic64/128 detection for clang

 configure           | 34 +++++++++++++++++++++++-----------
 util/bufferiszero.c | 33 +++++++++++++++++++++++----------
 2 files changed, 46 insertions(+), 21 deletions(-)
Signature · 1 line
2.28.0.rc0.105.gf9edc3c819-goog
#1 lore raw

PATCH 1/2 configure: avx2 and avx512f detection for clang

Since clang does not support "#pragma GCC", the instruction sets are always disabled. In this change, we

1. wrap "#pragma GCC" inside "#ifndef __clang__",
2. only retain them around "#include <{e,i,s}mmintrin.h>" to work around gcc bug,
3. and annotate each function with __attribute__((target(*))) which is recognized by both gcc and clang.

---
 configure           | 16 ++++++++++++++--
 util/bufferiszero.c | 33 +++++++++++++++++++++++----------
 2 files changed, 37 insertions(+), 12 deletions(-)
configure +14 -2
index 4bd80ed507..d9ce3aa5db 100755
@@ -5808,10 +5808,16 @@ fi
 
 if test "$cpuid_h" = "yes" && test "$avx2_opt" != "no"; then
   cat > $TMPC << EOF
+#include <cpuid.h>
+#ifndef __clang__
 #pragma GCC push_options
 #pragma GCC target("avx2")
-#include <cpuid.h>
+#endif
 #include <immintrin.h>
+#ifndef __clang__
+#pragma GCC pop_options
+#endif
+__attribute__((target("avx2")))
 static int bar(void *a) {
     __m256i x = *(__m256i *)a;
     return _mm256_testz_si256(x, x);
@@ -5835,10 +5841,16 @@ fi
 
 if test "$cpuid_h" = "yes" && test "$avx512f_opt" = "yes"; then
   cat > $TMPC << EOF
+#include <cpuid.h>
+#ifndef __clang__
 #pragma GCC push_options
 #pragma GCC target("avx512f")
-#include <cpuid.h>
+#endif
 #include <immintrin.h>
+#ifndef __clang__
+#pragma GCC pop_options
+#endif
+__attribute__((target("avx512f")))
 static int bar(void *a) {
     __m512i x = *(__m512i *)a;
     return _mm512_test_epi64_mask(x, x);
util/bufferiszero.c +23 -10
index 695bb4ce28..ca836b6e8c 100644
@@ -64,17 +64,18 @@ buffer_zero_int(const void *buf, size_t len)
 }
 
 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT) || defined(__SSE2__)
-/* Do not use push_options pragmas unnecessarily, because clang
- * does not support them.
- */
-#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
+#ifndef __clang__
 #pragma GCC push_options
 #pragma GCC target("sse2")
 #endif
 #include <emmintrin.h>
+#ifndef __clang__
+#pragma GCC pop_options
+#endif
 
 /* Note that each of these vectorized functions require len >= 64.  */
 
+__attribute__((target("sse2")))
 static bool
 buffer_zero_sse2(const void *buf, size_t len)
 {
@@ -104,19 +105,22 @@ buffer_zero_sse2(const void *buf, size_t len)
 
     return _mm_movemask_epi8(_mm_cmpeq_epi8(t, zero)) == 0xFFFF;
 }
-#if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
-#pragma GCC pop_options
-#endif
 
 #ifdef CONFIG_AVX2_OPT
 /* Note that due to restrictions/bugs wrt __builtin functions in gcc <= 4.8,
  * the includes have to be within the corresponding push_options region, and
  * therefore the regions themselves have to be ordered with increasing ISA.
  */
+#ifndef __clang__
 #pragma GCC push_options
 #pragma GCC target("sse4")
+#endif
 #include <smmintrin.h>
+#ifndef __clang__
+#pragma GCC pop_options
+#endif
 
+__attribute__((target("sse4")))
 static bool
 buffer_zero_sse4(const void *buf, size_t len)
 {
@@ -145,11 +149,16 @@ buffer_zero_sse4(const void *buf, size_t len)
     return _mm_testz_si128(t, t);
 }
 
-#pragma GCC pop_options
+#ifndef __clang__
 #pragma GCC push_options
 #pragma GCC target("avx2")
+#endif
 #include <immintrin.h>
+#ifndef __clang__
+#pragma GCC pop_options
+#endif
 
+__attribute__((target("avx2")))
 static bool
 buffer_zero_avx2(const void *buf, size_t len)
 {
@@ -176,14 +185,19 @@ buffer_zero_avx2(const void *buf, size_t len)
 
     return _mm256_testz_si256(t, t);
 }
-#pragma GCC pop_options
 #endif /* CONFIG_AVX2_OPT */
 
 #ifdef CONFIG_AVX512F_OPT
+#ifndef __clang__
 #pragma GCC push_options
 #pragma GCC target("avx512f")
+#endif
 #include <immintrin.h>
+#ifndef __clang__
+#pragma GCC pop_options
+#endif
 
+__attribute__((target("avx512f")))
 static bool
 buffer_zero_avx512(const void *buf, size_t len)
 {
@@ -210,7 +224,6 @@ buffer_zero_avx512(const void *buf, size_t len)
     return !_mm512_test_epi64_mask(t, t);
 
 }
-#pragma GCC pop_options
 #endif
 
 
Signature · 1 line
2.28.0.rc0.105.gf9edc3c819-goog
#2 lore raw

Re: [PATCH 1/2] configure: avx2 and avx512f detection for clang

Quoted 27 lines · Shu-Chun Weng

Since clang does not support "#pragma GCC", the instruction sets are always disabled. In this change, we

1. wrap "#pragma GCC" inside "#ifndef __clang__",
2. only retain them around "#include <{e,i,s}mmintrin.h>" to work around gcc bug,
3. and annotate each function with __attribute__((target(*))) which is recognized by both gcc and clang.

---
 configure           | 16 ++++++++++++++--
 util/bufferiszero.c | 33 +++++++++++++++++++++++----------
 2 files changed, 37 insertions(+), 12 deletions(-)
configure +2 -0
index 4bd80ed507..d9ce3aa5db 100755
@@ -5808,10 +5808,16 @@ fi
 
 if test "$cpuid_h" = "yes" && test "$avx2_opt" != "no"; then
   cat > $TMPC << EOF
+#include <cpuid.h>
+#ifndef __clang__
 #pragma GCC push_options
 #pragma GCC target("avx2")

configure fi · line 5823

Quoted 10 lines
configure +5 -1
-#include <cpuid.h>
+#endif
 #include <immintrin.h>
+#ifndef __clang__
+#pragma GCC pop_options
+#endif
+__attribute__((target("avx2")))
 static int bar(void *a) {
     __m256i x = *(__m256i *)a;
     return _mm256_testz_si256(x, x);

I wonder whether it would make more sense to pass "-mavx2" to the compile_object call afterwards and simply remove the #pragmas here? Did you try that already?

Thomas

#3 lore raw

Re: [PATCH 1/2] configure: avx2 and avx512f detection for clang

Do we have the flexibility to do that for util/bufferiszero.c as well? Otherwise, we are using different mechanisms to detect (compile test.c with -mavx2) and actually use (GCC pragma & __attribute__((target(*))))) the feature in production.

Shu-Chun

Quoted 46 lines · Thomas Huth
Quoted 27 lines · Shu-Chun Weng

Since clang does not support "#pragma GCC", the instruction sets are always disabled. In this change, we

1. wrap "#pragma GCC" inside "#ifndef __clang__",
2. only retain them around "#include <{e,i,s}mmintrin.h>" to work around gcc bug,
3. and annotate each function with __attribute__((target(*))) which is recognized by both gcc and clang.

---
 configure           | 16 ++++++++++++++--
 util/bufferiszero.c | 33 +++++++++++++++++++++++----------
 2 files changed, 37 insertions(+), 12 deletions(-)
configure +2 -0
index 4bd80ed507..d9ce3aa5db 100755
@@ -5808,10 +5808,16 @@ fi
 
 if test "$cpuid_h" = "yes" && test "$avx2_opt" != "no"; then
   cat > $TMPC << EOF
+#include <cpuid.h>
+#ifndef __clang__
 #pragma GCC push_options
 #pragma GCC target("avx2")

configure fi · line 5823

Quoted 10 lines
configure +5 -1
-#include <cpuid.h>
+#endif
 #include <immintrin.h>
+#ifndef __clang__
+#pragma GCC pop_options
+#endif
+__attribute__((target("avx2")))
 static int bar(void *a) {
     __m256i x = *(__m256i *)a;
     return _mm256_testz_si256(x, x);

I wonder whether it would make more sense to pass "-mavx2" to the compile_object call afterwards and simply remove the #pragmas here? Did you try that already?

Thomas

#4 lore raw

Re: [PATCH 1/2] configure: avx2 and avx512f detection for clang

Quoted 4 lines · Shu-Chun Weng

Do we have the flexibility to do that for util/bufferiszero.c as well? Otherwise, we are using different mechanisms to detect (compile test.c with -mavx2) and actually use (GCC pragma & __attribute__((target(*))))) the feature in production.

That's true ... so it's likely better to keep the pragmas in the configure script, indeed!

Thanks,
Thomas

Quoted 50 lines

Shu-Chun

On Wed, Jul 22, 2020 at 9:55 PM Thomas Huth <[email protected] <mailto:[email protected]>> wrote:

    On 23/07/2020 02.27, Shu-Chun Weng wrote:
    > Since clang does not support "#pragma GCC", the instruction sets are
    > always disabled. In this change, we
    >
    >  1. wrap "#pragma GCC" inside "#ifndef __clang__",
    >  2. only retain them around "#include <{e,i,s}mmintrin.h>" to work
    >     around gcc bug,
    >  3. and annotate each function with `__attribute__((target(*)))` which
    >     is recognized by both gcc and clang.
    >
    > Signed-off-by: Shu-Chun Weng <[email protected] <mailto:[email protected]>>
    > ---
    >  configure           | 16 ++++++++++++++--
    >  util/bufferiszero.c | 33 +++++++++++++++++++++++----------
    >  2 files changed, 37 insertions(+), 12 deletions(-)
    >
    > diff --git a/configure b/configure
    > index 4bd80ed507..d9ce3aa5db 100755
    > --- a/configure
    > +++ b/configure
    > @@ -5808,10 +5808,16 @@ fi
    > 
    >  if test "$cpuid_h" = "yes" && test "$avx2_opt" != "no"; then
    >    cat > $TMPC << EOF
    > +#include <cpuid.h>
    > +#ifndef __clang__
    >  #pragma GCC push_options
    >  #pragma GCC target("avx2")
    > -#include <cpuid.h>
    > +#endif
    >  #include <immintrin.h>
    > +#ifndef __clang__
    > +#pragma GCC pop_options
    > +#endif
    > +__attribute__((target("avx2")))
    >  static int bar(void *a) {
    >      __m256i x = *(__m256i *)a;
    >      return _mm256_testz_si256(x, x);

    I wonder whether it would make more sense to pass "-mavx2" to the
    compile_object call afterwards and simply remove the #pragmas here?
    Did you try that already?

     Thomas
#5 lore raw

PATCH 2/2 configure: atomic64/128 detection for clang

In reply to Shu-Chun Weng

The public interface for __atomic_* and __sync_* do not contain the explicit *_{number} versions:
https://gcc.gnu.org/onlinedocs/gcc/_005f_005fatomic-Builtins.html
https://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html

They appear to be GCC's internal symbols which happen to work. However, clang does not recognize them. Replace the existing usages with the _n versions (or no suffix) which are the documented API.

---
 configure | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)
configure +9 -9
index d9ce3aa5db..0613a049e9 100755
@@ -5894,9 +5894,9 @@ if test "$int128" = "yes"; then
 int main(void)
 {
   unsigned __int128 x = 0, y = 0;
-  y = __atomic_load_16(&x, 0);
-  __atomic_store_16(&x, y, 0);
-  __atomic_compare_exchange_16(&x, &y, x, 0, 0, 0);
+  y = __atomic_load_n(&x, 0);
+  __atomic_store_n(&x, y, 0);
+  __atomic_compare_exchange_n(&x, &y, x, 0, 0, 0);
   return 0;
 }
 EOF
@@ -5911,7 +5911,7 @@ if test "$int128" = yes && test "$atomic128" = no; then
 int main(void)
 {
   unsigned __int128 x = 0, y = 0;
-  __sync_val_compare_and_swap_16(&x, y, x);
+  __sync_val_compare_and_swap(&x, y, x);
   return 0;
 }
 EOF
@@ -5931,11 +5931,11 @@ int main(void)
 {
   uint64_t x = 0, y = 0;
 #ifdef __ATOMIC_RELAXED
-  y = __atomic_load_8(&x, 0);
-  __atomic_store_8(&x, y, 0);
-  __atomic_compare_exchange_8(&x, &y, x, 0, 0, 0);
-  __atomic_exchange_8(&x, y, 0);
-  __atomic_fetch_add_8(&x, y, 0);
+  y = __atomic_load_n(&x, 0);
+  __atomic_store_n(&x, y, 0);
+  __atomic_compare_exchange_n(&x, &y, x, 0, 0, 0);
+  __atomic_exchange_n(&x, y, 0);
+  __atomic_fetch_add(&x, y, 0);
 #else
   typedef char is_host64[sizeof(void *) >= sizeof(uint64_t) ? 1 : -1];
   __sync_lock_test_and_set(&x, y);
Signature · 1 line
2.28.0.rc0.105.gf9edc3c819-goog
#6 lore raw
In reply to Shu-Chun Weng

Ping: https://patchew.org/QEMU/[email protected]/

Quoted 19 lines · Shu-Chun Weng

Currently when configuring QEMU with clang, AVX2, AVX512F, ATOMIC64, and ATOMIC128 are all disabled because the detection code is GCC-only. With these
two patches, I am able to configure, build, and run tests with clang with all of
the above enabled.

Shu-Chun Weng (2):
configure: avx2 and avx512f detection for clang
configure: atomic64/128 detection for clang

 configure           | 34 +++++++++++++++++++++++-----------
 util/bufferiszero.c | 33 +++++++++++++++++++++++----------
 2 files changed, 46 insertions(+), 21 deletions(-)

--
2.28.0.rc0.105.gf9edc3c819-goog

#7 lore raw

Ping.

Quoted 23 lines · Shu-Chun Weng

Ping: https://patchew.org/QEMU/[email protected]/

Quoted 19 lines · Shu-Chun Weng

Currently when configuring QEMU with clang, AVX2, AVX512F, ATOMIC64, and ATOMIC128 are all disabled because the detection code is GCC-only. With these
two patches, I am able to configure, build, and run tests with clang with all of
the above enabled.

Shu-Chun Weng (2):
configure: avx2 and avx512f detection for clang
configure: atomic64/128 detection for clang

 configure           | 34 +++++++++++++++++++++++-----------
 util/bufferiszero.c | 33 +++++++++++++++++++++++----------
 2 files changed, 46 insertions(+), 21 deletions(-)

--
2.28.0.rc0.105.gf9edc3c819-goog