| Message ID | CAMZc-bwxBuVGPZ=UaeXqLmdBJnYZDzJSPcL4CRaC1fWMvsP6xQ@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | Increase rtx_cost of sse_to_integer in skylake_cost. | expand |
On 9/15/20 9:20 PM, Hongtao Liu via Gcc-patches wrote: > Hi: > Rtx cost of sse_to_integer would be used by pass_stv as a > measurement for the scalar-to-vector transformation. As > https://gcc.gnu.org/pipermail/gcc-patches/2019-August/528839.html > indicates, movement between sse regs and gprs should be much expensive > than movement inside gprs(which is 2 as default). This patch would > also fix "pr96861". > > Bootstrap is ok, regression test is ok for both "i386.exp=* > --target_board='unix{-m32,}'" and "i386.exp=* > --target_board='unix{-m32\ -march=cascadelake,-m64\ > -march=cascadelake}"". > No big impact on SPEC2017. > Ok for trunk? > > gcc/ChangeLog > > PR target/96861 > * config/i386/x86-tune-costs.h (skylake_cost): increase rtx > cost of sse_to_integer from 2 to 6. > > gcc/testsuite > > * gcc.target/i386/pr95021-3.c: Add -mtune=generic. I'll defer to HJ's judgement here. If he's OK with it, then it's fine by me. jeff
On Thu, Sep 17, 2020 at 12:10 PM Jeff Law <law@redhat.com> wrote: > > > On 9/15/20 9:20 PM, Hongtao Liu via Gcc-patches wrote: > > Hi: > > Rtx cost of sse_to_integer would be used by pass_stv as a > > measurement for the scalar-to-vector transformation. As > > https://gcc.gnu.org/pipermail/gcc-patches/2019-August/528839.html > > indicates, movement between sse regs and gprs should be much expensive > > than movement inside gprs(which is 2 as default). This patch would > > also fix "pr96861". > > > > Bootstrap is ok, regression test is ok for both "i386.exp=* > > --target_board='unix{-m32,}'" and "i386.exp=* > > --target_board='unix{-m32\ -march=cascadelake,-m64\ > > -march=cascadelake}"". > > No big impact on SPEC2017. > > Ok for trunk? > > > > gcc/ChangeLog > > > > PR target/96861 > > * config/i386/x86-tune-costs.h (skylake_cost): increase rtx > > cost of sse_to_integer from 2 to 6. > > > > gcc/testsuite > > > > * gcc.target/i386/pr95021-3.c: Add -mtune=generic. > > I'll defer to HJ's judgement here. If he's OK with it, then it's fine > by me. > I've talked to H.J and he agrees on this patch. So i'm going to check in this patch. > > jeff > >
On Thu, Sep 17, 2020 at 11:11 PM Hongtao Liu <crazylht@gmail.com> wrote: > > On Thu, Sep 17, 2020 at 12:10 PM Jeff Law <law@redhat.com> wrote: > > > > > > On 9/15/20 9:20 PM, Hongtao Liu via Gcc-patches wrote: > > > Hi: > > > Rtx cost of sse_to_integer would be used by pass_stv as a > > > measurement for the scalar-to-vector transformation. As > > > https://gcc.gnu.org/pipermail/gcc-patches/2019-August/528839.html > > > indicates, movement between sse regs and gprs should be much expensive > > > than movement inside gprs(which is 2 as default). This patch would > > > also fix "pr96861". > > > > > > Bootstrap is ok, regression test is ok for both "i386.exp=* > > > --target_board='unix{-m32,}'" and "i386.exp=* > > > --target_board='unix{-m32\ -march=cascadelake,-m64\ > > > -march=cascadelake}"". > > > No big impact on SPEC2017. > > > Ok for trunk? > > > > > > gcc/ChangeLog > > > > > > PR target/96861 > > > * config/i386/x86-tune-costs.h (skylake_cost): increase rtx > > > cost of sse_to_integer from 2 to 6. > > > > > > gcc/testsuite > > > > > > * gcc.target/i386/pr95021-3.c: Add -mtune=generic. > > > > I'll defer to HJ's judgement here. If he's OK with it, then it's fine > > by me. > > > > I've talked to H.J and he agrees on this patch. > So i'm going to check in this patch. I am checking this patch to add a testcase.
From 2a9a943ec56ca3ea1ba1a2447a32b103c2a1c790 Mon Sep 17 00:00:00 2001 From: liuhongt <hongtao.liu@intel.com> Date: Wed, 16 Sep 2020 10:53:52 +0800 Subject: [PATCH] Increase rtx cost of sse_to_integer in skylake_cost. As https://gcc.gnu.org/pipermail/gcc-patches/2019-August/528839.html indicates, movement between SSE and gpr should be much expensive than movement inside gpr(which is 2 as default). gcc/ChangeLog PR target/96861 * config/i386/x86-tune-costs.h (skylake_cost): increase rtx cost of sse_to_integer from 2 to 6. gcc/testsuite * gcc.target/i386/pr95021-3.c: Add -mtune=generic. --- gcc/config/i386/x86-tune-costs.h | 2 +- gcc/testsuite/gcc.target/i386/pr95021-3.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/gcc/config/i386/x86-tune-costs.h b/gcc/config/i386/x86-tune-costs.h index a782a9dd9e3..c289b6ba454 100644 --- a/gcc/config/i386/x86-tune-costs.h +++ b/gcc/config/i386/x86-tune-costs.h @@ -1769,7 +1769,7 @@ struct processor_costs skylake_cost = { {6, 6, 6, 10, 20}, /* cost of unaligned loads. */ {8, 8, 8, 8, 16}, /* cost of unaligned stores. */ 2, 2, 4, /* cost of moving XMM,YMM,ZMM register */ - 2, /* cost of moving SSE register to integer. */ + 6, /* cost of moving SSE register to integer. */ 20, 8, /* Gather load static, per_elt. */ 22, 10, /* Gather store static, per_elt. */ 64, /* size of l1 cache. */ diff --git a/gcc/testsuite/gcc.target/i386/pr95021-3.c b/gcc/testsuite/gcc.target/i386/pr95021-3.c index 1748161a77c..52f9e4569b3 100644 --- a/gcc/testsuite/gcc.target/i386/pr95021-3.c +++ b/gcc/testsuite/gcc.target/i386/pr95021-3.c @@ -1,5 +1,5 @@ /* { dg-do compile { target ia32 } } */ -/* { dg-options "-O2 -msse2 -mstv -mregparm=3 -W" } */ +/* { dg-options "-O2 -msse2 -mstv -mregparm=3 -W -mtune=generic" } */ /* { dg-final { scan-assembler "movq\[ \t\]+\[^\n\]*, %xmm" } } */ #include "pr95021-1.c" -- 2.18.1
Hi: Rtx cost of sse_to_integer would be used by pass_stv as a measurement for the scalar-to-vector transformation. As https://gcc.gnu.org/pipermail/gcc-patches/2019-August/528839.html indicates, movement between sse regs and gprs should be much expensive than movement inside gprs(which is 2 as default). This patch would also fix "pr96861". Bootstrap is ok, regression test is ok for both "i386.exp=* --target_board='unix{-m32,}'" and "i386.exp=* --target_board='unix{-m32\ -march=cascadelake,-m64\ -march=cascadelake}"". No big impact on SPEC2017. Ok for trunk? gcc/ChangeLog PR target/96861 * config/i386/x86-tune-costs.h (skylake_cost): increase rtx cost of sse_to_integer from 2 to 6. gcc/testsuite * gcc.target/i386/pr95021-3.c: Add -mtune=generic. -- BR, Hongtao