This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH][rs6000] avoid using unaligned vsx or lxvd2x/stxvd2x for memcpy/memmove inline expansion
- From: Aaron Sawdey <acsawdey at linux dot ibm dot com>
- To: GCC Patches <gcc-patches at gcc dot gnu dot org>, Segher Boessenkool <segher at kernel dot crashing dot org>, David Edelsohn <dje dot gcc at gmail dot com>, Bill Schmidt <wschmidt at linux dot ibm dot com>
- Date: Wed, 19 Dec 2018 13:53:05 -0600
- Subject: [PATCH][rs6000] avoid using unaligned vsx or lxvd2x/stxvd2x for memcpy/memmove inline expansion
Because of POWER9 dd2.1 issues with certain unaligned vsx instructions
to cache inhibited memory, here is a patch that keeps memmove (and memcpy)
inline expansion from doing unaligned vector or using vector load/store
other than lvx/stvx. More description of the issue is here:
https://patchwork.ozlabs.org/patch/814059/
OK for trunk if bootstrap/regtest ok?
Thanks!
Aaron
2018-12-19 Aaron Sawdey <acsawdey@linux.ibm.com>
* config/rs6000/rs6000-string.c (expand_block_move): Don't use
unaligned vsx and avoid lxvd2x/stxvd2x.
(gen_lvx_v4si_move): New function.
Index: gcc/config/rs6000/rs6000-string.c
===================================================================
--- gcc/config/rs6000/rs6000-string.c (revision 267055)
+++ gcc/config/rs6000/rs6000-string.c (working copy)
@@ -2669,6 +2669,35 @@
return true;
}
+/* Generate loads and stores for a move of v4si mode using lvx/stvx.
+ This uses altivec_{l,st}vx_<mode>_internal which use unspecs to
+ keep combine from changing what instruction gets used.
+
+ DEST is the destination for the data.
+ SRC is the source of the data for the move. */
+
+static rtx
+gen_lvx_v4si_move (rtx dest, rtx src)
+{
+ rtx rv = NULL;
+ if (MEM_P (dest))
+ {
+ gcc_assert (!MEM_P (src));
+ gcc_assert (GET_MODE (src) == V4SImode);
+ rv = gen_altivec_stvx_v4si_internal (dest, src);
+ }
+ else if (MEM_P (src))
+ {
+ gcc_assert (!MEM_P (dest));
+ gcc_assert (GET_MODE (dest) == V4SImode);
+ rv = gen_altivec_lvx_v4si_internal (dest, src);
+ }
+ else
+ gcc_unreachable ();
+
+ return rv;
+}
+
/* Expand a block move operation, and return 1 if successful. Return 0
if we should let the compiler generate normal code.
@@ -2721,11 +2750,11 @@
/* Altivec first, since it will be faster than a string move
when it applies, and usually not significantly larger. */
- if (TARGET_ALTIVEC && bytes >= 16 && (TARGET_EFFICIENT_UNALIGNED_VSX || align >= 128))
+ if (TARGET_ALTIVEC && bytes >= 16 && align >= 128)
{
move_bytes = 16;
mode = V4SImode;
- gen_func.mov = gen_movv4si;
+ gen_func.mov = gen_lvx_v4si_move;
}
else if (bytes >= 8 && TARGET_POWERPC64
&& (align >= 64 || !STRICT_ALIGNMENT))
--
Aaron Sawdey, Ph.D. acsawdey@linux.vnet.ibm.com
050-2/C113 (507) 253-7520 home: 507/263-0782
IBM Linux Technology Center - PPC Toolchain