* [COMMITTED] i386: Account for the memory read in V*QImode multiplication sequences
@ 2023-05-22 14:36 Uros Bizjak
0 siblings, 0 replies; only message in thread
From: Uros Bizjak @ 2023-05-22 14:36 UTC (permalink / raw)
To: gcc-patches
[-- Attachment #1: Type: text/plain, Size: 289 bytes --]
Add the cost of a memory read to the cost of V*QImode vector mult sequences.
gcc/ChangeLog:
* config/i386/i386.cc (ix86_multiplication_cost): Add
the cost of a memory read to the cost of V?QImode sequences.
Bootstrapped and regression tested on x86_64-linux-gnu {,-m32}.
Uros.
[-- Attachment #2: p.diff.txt --]
[-- Type: text/plain, Size: 2157 bytes --]
diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
index 6a4b3326219..a36e625342d 100644
--- a/gcc/config/i386/i386.cc
+++ b/gcc/config/i386/i386.cc
@@ -20463,27 +20463,42 @@ ix86_multiplication_cost (const struct processor_costs *cost,
{
case V4QImode:
case V8QImode:
- /* Partial V*QImode is emulated with 4-5 insns. */
- if ((TARGET_AVX512BW && TARGET_AVX512VL) || TARGET_XOP)
+ /* Partial V*QImode is emulated with 4-6 insns. */
+ if (TARGET_AVX512BW && TARGET_AVX512VL)
return ix86_vec_cost (mode, cost->mulss + cost->sse_op * 3);
+ else if (TARGET_AVX2)
+ return ix86_vec_cost (mode, cost->mulss + cost->sse_op * 5);
+ else if (TARGET_XOP)
+ return (ix86_vec_cost (mode, cost->mulss + cost->sse_op * 3)
+ + cost->sse_load[2]);
else
- return ix86_vec_cost (mode, cost->mulss + cost->sse_op * 4);
+ return (ix86_vec_cost (mode, cost->mulss + cost->sse_op * 4)
+ + cost->sse_load[2]);
case V16QImode:
/* V*QImode is emulated with 4-11 insns. */
if (TARGET_AVX512BW && TARGET_AVX512VL)
return ix86_vec_cost (mode, cost->mulss + cost->sse_op * 3);
+ else if (TARGET_AVX2)
+ return ix86_vec_cost (mode, cost->mulss * 2 + cost->sse_op * 8);
else if (TARGET_XOP)
- return ix86_vec_cost (mode, cost->mulss * 2 + cost->sse_op * 5);
- /* FALLTHRU */
+ return (ix86_vec_cost (mode, cost->mulss * 2 + cost->sse_op * 5)
+ + cost->sse_load[2]);
+ else
+ return (ix86_vec_cost (mode, cost->mulss * 2 + cost->sse_op * 7)
+ + cost->sse_load[2]);
+
case V32QImode:
- if (TARGET_AVX512BW && mode == V32QImode)
+ if (TARGET_AVX512BW)
return ix86_vec_cost (mode, cost->mulss + cost->sse_op * 3);
else
- return ix86_vec_cost (mode, cost->mulss * 2 + cost->sse_op * 7);
+ return (ix86_vec_cost (mode, cost->mulss * 2 + cost->sse_op * 7)
+ + cost->sse_load[3] * 2);
case V64QImode:
- return ix86_vec_cost (mode, cost->mulss * 2 + cost->sse_op * 9);
+ return (ix86_vec_cost (mode, cost->mulss * 2 + cost->sse_op * 9)
+ + cost->sse_load[3] * 2
+ + cost->sse_load[4] * 2);
case V4SImode:
/* pmulld is used in this case. No emulation is needed. */
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2023-05-22 14:36 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-05-22 14:36 [COMMITTED] i386: Account for the memory read in V*QImode multiplication sequences Uros Bizjak
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).