libgcc: Mostly vectorize CIE encoding extraction for FDEs

Message ID 87mt9umxu4.fsf@oldenburg.str.redhat.com
State Not applicable
Headers
Series libgcc: Mostly vectorize CIE encoding extraction for FDEs |

Checks

Context Check Description
dj/TryBot-apply_patch fail Patch failed to apply to master at the time it was sent
dj/TryBot-32bit fail Patch series failed to apply

Commit Message

Florian Weimer Oct. 17, 2022, 12:36 p.m. UTC
  "zR" and "zPLR" are the most common augmentations.  Use a simple
SIMD-with-in-a-register technique to check for both augmentations,
and that the following variable-length integers have length 1, to
get more quickly at the encoding field.

libgcc/

	* unwind-dw2-fde.c (get_cie_encoding_slow): Rename from
	get_cie_encoding.  Mark as noinline.
	(get_cie_encoding): Add fast path for "zR" and "zPLR"
	augmentations.  Call get_cie_encoding_slow as a fall-back.

---
 libgcc/unwind-dw2-fde.c | 61 +++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 59 insertions(+), 2 deletions(-)


base-commit: de84a1e4b107b803ec3b064c3771a6ed8c0e201e
  

Comments

Florian Weimer Oct. 17, 2022, 1 p.m. UTC | #1
* Florian Weimer via Libc-alpha:

> "zR" and "zPLR" are the most common augmentations.  Use a simple
> SIMD-with-in-a-register technique to check for both augmentations,
> and that the following variable-length integers have length 1, to
> get more quickly at the encoding field.
>
> libgcc/
>
> 	* unwind-dw2-fde.c (get_cie_encoding_slow): Rename from
> 	get_cie_encoding.  Mark as noinline.
> 	(get_cie_encoding): Add fast path for "zR" and "zPLR"
> 	augmentations.  Call get_cie_encoding_slow as a fall-back.

Sorry, this should have to gcc-patches.  Will resend there.

Thanks,
Florian
  
Andreas Schwab Oct. 17, 2022, 1:03 p.m. UTC | #2
On Okt 17 2022, Florian Weimer via Libc-alpha wrote:

> +      unsigned long long value = *(const unsigned long long *) &cie->version;

Is that strict-aliasing safe?

I guess you wanted to send that to gcc-patches@.
  
Florian Weimer Oct. 17, 2022, 1:07 p.m. UTC | #3
* Andreas Schwab:

> On Okt 17 2022, Florian Weimer via Libc-alpha wrote:
>
>> +      unsigned long long value = *(const unsigned long long *) &cie->version;
>
> Is that strict-aliasing safe?

Should be, the data is never written to during the program's life-time.

libgcc doesn't use memcpy consistently for reading this constant data.
Instead it uses non-char types to read things.

> I guess you wanted to send that to gcc-patches@.

Right.

Thanks,
Florian
  

Patch

diff --git a/libgcc/unwind-dw2-fde.c b/libgcc/unwind-dw2-fde.c
index 3c0cc654ec0..4e3a54c5a1a 100644
--- a/libgcc/unwind-dw2-fde.c
+++ b/libgcc/unwind-dw2-fde.c
@@ -333,8 +333,10 @@  base_from_object (unsigned char encoding, const struct object *ob)
 /* Return the FDE pointer encoding from the CIE.  */
 /* ??? This is a subset of extract_cie_info from unwind-dw2.c.  */
 
-static int
-get_cie_encoding (const struct dwarf_cie *cie)
+/* Disable inlining because the function is only used as a slow path in
+   get_cie_encoding below.  */
+static int __attribute__ ((noinline))
+get_cie_encoding_slow (const struct dwarf_cie *cie)
 {
   const unsigned char *aug, *p;
   _Unwind_Ptr dummy;
@@ -389,6 +391,61 @@  get_cie_encoding (const struct dwarf_cie *cie)
     }
 }
 
+static inline int
+get_cie_encoding (const struct dwarf_cie *cie)
+{
+  /* Fast path for some augmentations and single-byte variable-length
+     integers.  Do this only for targets that align struct dwarf_cie to 8
+     bytes, which ensures that at least 8 bytes are available starting at
+     cie->version.  */
+#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ \
+  || __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  if (__alignof (*cie) == 8 && sizeof (unsigned long long) == 8)
+    {
+      unsigned long long value = *(const unsigned long long *) &cie->version;
+
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+#define C(x) __builtin_bswap64 (x)
+#else
+#define C(x) x
+#endif
+
+      /* Fast path for "zR".  Check for version 1, the "zR" string and that
+	 the sleb128/uleb128 values are single bytes.  In the comments
+	 below, '1', 'c', 'd', 'r', 'l' are version, code alignment, data
+	 alignment, return address column, augmentation length.  Note that
+	 with CIE version 1, the return address column is byte-encoded.  */
+      unsigned long long expected =
+	/*   1 z R 0 c d r l.  */
+	C (0x017a520000000000ULL);
+      unsigned long long mask =
+	/*   1 z R 0 c d r l.  */
+	C (0xffffffff80800080ULL);
+
+      if ((value & mask) == expected)
+	return cie->augmentation[7];
+
+      /* Fast path for "zPLR".  */
+      expected =
+	/*   1 z P L R 0 c d.  */
+	C (0x017a504c52000000ULL);
+      mask =
+	/*   1 z P L R 0 c d.  */
+	C (0xffffffffffff8080ULL);
+#undef C
+
+      /* Validate the augmentation length, and return the enconding after
+	 it.  No check for the return address column because it is
+	 byte-encoded with CIE version 1.  */
+      if (__builtin_expect ((value & mask) == expected
+			    && (cie->augmentation[8] & 0x80) == 0, 1))
+	  return cie->augmentation[9];
+    }
+#endif
+
+  return get_cie_encoding_slow (cie);
+}
+
 static inline int
 get_fde_encoding (const struct dwarf_fde *f)
 {