Hi Ruifeng,

>Building with gcc 10.2 with SVE extension enabled got error:
>
>{standard input}: Assembler messages:
>{standard input}:4002: Error: selected processor does not support `mov
>z3.b,#0'
>{standard input}:4003: Error: selected processor does not support
>`whilelo p1.b,xzr,x7'
>{standard input}:4005: Error: selected processor does not support `ld1b
>z0.b,p1/z,[x8]'
>{standard input}:4006: Error: selected processor does not support
>`whilelo p4.s,wzr,w7'
>
>This is because inline assembly code explicitly resets cpu model to
>not have SVE support. Thus SVE instructions generated by compiler
>auto vectorization got rejected by assembler.
>
>Fixed the issue by replacing inline assembly with equivalent atomic
>built-ins. Compiler will generate LSE instructions for cpu that has
>the extension.
>
>Fixes: 8a4f835971f5 ("common/octeontx2: add IO handling APIs")
>Cc: jer...@marvell.com
>Cc: sta...@dpdk.org
>
>Signed-off-by: Ruifeng Wang <ruifeng.w...@arm.com>
>---
> drivers/common/octeontx2/otx2_io_arm64.h | 37 +++--------------------
>-
> 1 file changed, 4 insertions(+), 33 deletions(-)
>
>diff --git a/drivers/common/octeontx2/otx2_io_arm64.h
>b/drivers/common/octeontx2/otx2_io_arm64.h
>index b5c85d9a6..8843a79b5 100644
>--- a/drivers/common/octeontx2/otx2_io_arm64.h
>+++ b/drivers/common/octeontx2/otx2_io_arm64.h
>@@ -24,55 +24,26 @@
> static __rte_always_inline uint64_t
> otx2_atomic64_add_nosync(int64_t incr, int64_t *ptr)
> {
>-      uint64_t result;
>-
>       /* Atomic add with no ordering */
>-      asm volatile (
>-              ".cpu  generic+lse\n"
>-              "ldadd %x[i], %x[r], [%[b]]"
>-              : [r] "=r" (result), "+m" (*ptr)
>-              : [i] "r" (incr), [b] "r" (ptr)
>-              : "memory");
>-      return result;
>+      return (uint64_t)__atomic_fetch_add(ptr, incr,
>__ATOMIC_RELAXED);
> }
>

Here LDADD acts as a way to interface to co-processors i.e. 
LDADD instruction opcode + specific io address are recognized by 
HW interceptor and dispatched to the specific coprocessor.

Leaving it to the compiler to use the correct instruction is a bad idea.
This breaks the arm64_armv8_linux_gcc build as it doesn't have the
+lse enabled.
__atomic_fetch_add will generate a different instruction with SVE 
enabled.

Instead can we add +sve to the first line to prevent outer loop from optimizing 
out 
the trap?

I tested with 10.2 and n2 config below change works fine.
-" .cpu          generic+lse\n"
+" .cpu         generic+lse+sve\n"

Regards,
Pavan.

> static __rte_always_inline uint64_t
> otx2_atomic64_add_sync(int64_t incr, int64_t *ptr)
> {
>-      uint64_t result;
>-
>-      /* Atomic add with ordering */
>-      asm volatile (
>-              ".cpu  generic+lse\n"
>-              "ldadda %x[i], %x[r], [%[b]]"
>-              : [r] "=r" (result), "+m" (*ptr)
>-              : [i] "r" (incr), [b] "r" (ptr)
>-              : "memory");
>-      return result;
>+      return (uint64_t)__atomic_fetch_add(ptr, incr,
>__ATOMIC_ACQUIRE);
> }
>
> static __rte_always_inline uint64_t
> otx2_lmt_submit(rte_iova_t io_address)
> {
>-      uint64_t result;
>-
>-      asm volatile (
>-              ".cpu  generic+lse\n"
>-              "ldeor xzr,%x[rf],[%[rs]]" :
>-               [rf] "=r"(result): [rs] "r"(io_address));
>-      return result;
>+      return __atomic_fetch_xor((uint64_t *)io_address, 0,
>__ATOMIC_RELAXED);
> }
>
> static __rte_always_inline uint64_t
> otx2_lmt_submit_release(rte_iova_t io_address)
> {
>-      uint64_t result;
>-
>-      asm volatile (
>-              ".cpu  generic+lse\n"
>-              "ldeorl xzr,%x[rf],[%[rs]]" :
>-               [rf] "=r"(result) : [rs] "r"(io_address));
>-      return result;
>+      return __atomic_fetch_xor((uint64_t *)io_address, 0,
>__ATOMIC_RELEASE);
> }
>
> static __rte_always_inline void
>--
>2.25.1

Reply via email to