From 0d6a31f275dc73dabf9fd8742a4adaa9ecc13da6 Mon Sep 17 00:00:00 2001
From: Henrik Gramner <gramner@twoorioles.com>
Date: Thu, 12 Jan 2023 01:49:39 +0100
Subject: [PATCH] x86: Remove stack alignment compiler flags
The intent was good, but in practice it results in a significant
amount of problems due to various compiler bugs for negligible gains.
---
meson.build | 34 +++-------------------------------
src/meson.build | 27 +++++----------------------
src/x86/cdef_avx2.asm | 32 ++++++++------------------------
src/x86/ipred16_avx2.asm | 4 ----
src/x86/mc16_avx2.asm | 25 ++++---------------------
src/x86/mc_avx2.asm | 23 ++++-------------------
src/x86/msac.asm | 10 ----------
tests/libfuzzer/meson.build | 5 ++---
tests/meson.build | 3 +--
9 files changed, 27 insertions(+), 136 deletions(-)
diff --git a/meson.build b/meson.build
index cee8b791e..0402f73d3 100644
--- a/meson.build
+++ b/meson.build
@@ -319,45 +319,17 @@ if fuzzing_engine == 'libfuzzer'
add_project_arguments(cc.first_supported_argument(fuzzer_args), language : 'c')
endif
-# Stack alignments flags
-
-stackalign_flag = []
-stackrealign_flag = []
-
cdata.set10('ENDIANNESS_BIG', host_machine.endian() == 'big')
if host_machine.cpu_family().startswith('x86')
if get_option('stack_alignment') > 0
stack_alignment = get_option('stack_alignment')
- elif host_machine.cpu_family() == 'x86_64'
- if cc.has_argument('-mpreferred-stack-boundary=6')
- stackalign_flag = ['-mpreferred-stack-boundary=6']
- stackrealign_flag = ['-mincoming-stack-boundary=4']
- stack_alignment = 32
- elif cc.has_argument('-mstack-alignment=64')
- stackalign_flag = ['-mstack-alignment=64']
- stackrealign_flag = ['-mstackrealign']
- stack_alignment = 32
- else
- stack_alignment = 16
- endif
+ elif host_machine.cpu_family() == 'x86_64' or host_machine.system() in ['linux', 'darwin', 'ios', 'tvos']
+ stack_alignment = 16
else
- if host_machine.system() == 'linux' or host_machine.system() in ['darwin', 'ios', 'tvos']
- stack_alignment = 16
- elif cc.has_argument('-mpreferred-stack-boundary=4')
- stackalign_flag = ['-mpreferred-stack-boundary=4']
- stackrealign_flag = ['-mincoming-stack-boundary=2']
- stack_alignment = 16
- elif cc.has_argument('-mstack-alignment=16')
- stackalign_flag = ['-mstack-alignment=16']
- stackrealign_flag = ['-mstackrealign']
- stack_alignment = 16
- else
- stack_alignment = 4
- endif
+ stack_alignment = 4
endif
cdata_asm.set('STACK_ALIGNMENT', stack_alignment)
- cdata.set('STACK_ALIGNMENT', stack_alignment)
endif
cdata.set10('ARCH_AARCH64', host_machine.cpu_family() == 'aarch64' or host_machine.cpu() == 'arm64')
diff --git a/src/meson.build b/src/meson.build
index 719015496..2ace55e1a 100644
--- a/src/meson.build
+++ b/src/meson.build
@@ -37,6 +37,7 @@ libdav1d_sources = files(
'intra_edge.c',
'itx_1d.c',
'lf_mask.c',
+ 'lib.c',
'log.c',
'mem.c',
'msac.c',
@@ -47,6 +48,7 @@ libdav1d_sources = files(
'refmvs.c',
'scan.c',
'tables.c',
+ 'thread_task.c',
'warpmv.c',
'wedge.c',
)
@@ -74,14 +76,6 @@ libdav1d_arch_tmpl_sources = []
libdav1d_bitdepth_objs = []
-# libdav1d entrypoint source files
-# These source files contain library entry points and are
-# built with the stack-realign flag set, where necessary.
-libdav1d_entrypoints_sources = files(
- 'lib.c',
- 'thread_task.c'
-)
-
# ASM specific sources
libdav1d_asm_objs = []
# Arch-specific flags
@@ -245,7 +239,7 @@ endif
libdav1d_rc_obj = []
-libdav1d_flags = [stackalign_flag]
+libdav1d_flags = []
api_export_flags = []
#
@@ -280,18 +274,6 @@ endif
# Library definitions
#
-# Helper library for dav1d entrypoints
-libdav1d_entrypoints_objs = static_library('dav1d_entrypoint',
- libdav1d_entrypoints_sources,
- rev_target, config_h_target,
-
- include_directories : dav1d_inc_dirs,
- dependencies: [stdatomic_dependencies],
- c_args : [libdav1d_flags, stackrealign_flag, api_export_flags],
- install : false,
- build_by_default : false,
-).extract_all_objects(recursive: true)
-
# Helper library for each bitdepth
libdav1d_bitdepth_objs = []
foreach bitdepth : dav1d_bitdepths
@@ -330,10 +312,11 @@ libdav1d = library('dav1d',
libdav1d_sources,
libdav1d_asm_objs,
libdav1d_rc_obj,
+ rev_target,
+ config_h_target,
objects : [
libdav1d_bitdepth_objs,
- libdav1d_entrypoints_objs
],
include_directories : dav1d_inc_dirs,
diff --git a/src/x86/cdef_avx2.asm b/src/x86/cdef_avx2.asm
index c5c66c759..1f30f8a3b 100644
--- a/src/x86/cdef_avx2.asm
+++ b/src/x86/cdef_avx2.asm
@@ -396,21 +396,17 @@ SECTION .text
%macro CDEF_FILTER 2 ; w, h
INIT_YMM avx2
-cglobal cdef_filter_%1x%2_8bpc, 5, 10, 0, dst, stride, left, top, bot, \
+cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride, left, top, bot, \
pri, sec, dir, damping, edge
%assign stack_offset_entry stack_offset
mov edged, edgem
cmp edged, 0xf
jne .border_block
- PUSH r10
PUSH r11
+ PUSH r12
%if %2 == 4
- %assign regs_used 12
- %if STACK_ALIGNMENT < 32
- PUSH r%+regs_used
- %assign regs_used regs_used+1
- %endif
+%assign regs_used 13
ALLOC_STACK 0x60, 16
pmovzxbw xm0, [leftq+1]
vpermq m0, m0, q0110
@@ -420,23 +416,15 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 10, 0, dst, stride, left, top, bot, \
movu [rsp+0x28], m1
movu [rsp+0x40], m2
%elif %1 == 4
- PUSH r12
- %assign regs_used 13
- %if STACK_ALIGNMENT < 32
- PUSH r%+regs_used
- %assign regs_used regs_used+1
- %endif
+%assign regs_used 14
+ PUSH r13
ALLOC_STACK 8*2+%1*%2*1, 16
pmovzxwd m0, [leftq]
mova [rsp+0x10], m0
%else
- PUSH r12
+%assign regs_used 15
PUSH r13
- %assign regs_used 14
- %if STACK_ALIGNMENT < 32
- PUSH r%+regs_used
- %assign regs_used regs_used+1
- %endif
+ PUSH r14
ALLOC_STACK 8*4+%1*%2*2+32, 16
lea r11, [strideq*3]
movu xm4, [dstq+strideq*2]
@@ -1209,11 +1197,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 10, 0, dst, stride, left, top, bot, \
DEFINE_ARGS dst, stride, left, top, bot, pri, sec, stride3, dst4, edge
%define rstk rsp
%assign stack_offset stack_offset_entry
-%assign regs_used 10
-%if STACK_ALIGNMENT < 32
- PUSH r%+regs_used
- %assign regs_used regs_used+1
-%endif
+%assign regs_used 11
ALLOC_STACK 2*16+(%2+4)*32, 16
%define px rsp+2*16+2*32
diff --git a/src/x86/ipred16_avx2.asm b/src/x86/ipred16_avx2.asm
index 72300c2a4..7ddb18991 100644
--- a/src/x86/ipred16_avx2.asm
+++ b/src/x86/ipred16_avx2.asm
@@ -1936,11 +1936,7 @@ ALIGN function_align
.upsample_left: ; h4/h8
mova xm0, [tlq-16] ; 8 7 6 5 4 3 2 1
movu xm1, [tlq-14] ; 7 6 5 4 3 2 1 0
-%if STACK_ALIGNMENT < 32
vpbroadcastw xm4, r8m ; pixel_max
-%else
- vpbroadcastw xm4, r9m ; r8m -> r9m due to call
-%endif
cmp hd, 8
je .upsample_left_h8
pshufhw xm2, xm0, q2100 ; _ _ _ _ 4 4 3 2
diff --git a/src/x86/mc16_avx2.asm b/src/x86/mc16_avx2.asm
index 8b2ec4fa9..61eeaa100 100644
--- a/src/x86/mc16_avx2.asm
+++ b/src/x86/mc16_avx2.asm
@@ -2650,23 +2650,14 @@ cglobal prep_8tap_16bpc, 4, 8, 0, tmp, src, stride, w, h, mx, my
%ifidn %1, put
%assign isput 1
%assign isprep 0
- %if required_stack_alignment <= STACK_ALIGNMENT
-cglobal put_8tap_scaled_16bpc, 4, 15, 16, 0xe0, dst, ds, src, ss, w, h, mx, my, dx, dy, pxmax
- %else
cglobal put_8tap_scaled_16bpc, 4, 14, 16, 0xe0, dst, ds, src, ss, w, h, mx, my, dx, dy, pxmax
- %endif
%xdefine base_reg r12
mov r7d, pxmaxm
%else
%assign isput 0
%assign isprep 1
- %if required_stack_alignment <= STACK_ALIGNMENT
-cglobal prep_8tap_scaled_16bpc, 4, 15, 16, 0xe0, tmp, src, ss, w, h, mx, my, dx, dy, pxmax
- %xdefine tmp_stridem r14q
- %else
cglobal prep_8tap_scaled_16bpc, 4, 14, 16, 0xe0, tmp, src, ss, w, h, mx, my, dx, dy, pxmax
%define tmp_stridem qword [rsp+0xd0]
- %endif
%xdefine base_reg r11
%endif
lea base_reg, [%1_8tap_scaled_16bpc_avx2]
@@ -2698,15 +2689,9 @@ cglobal prep_8tap_scaled_16bpc, 4, 14, 16, 0xe0, tmp, src, ss, w, h, mx, my, dx,
DEFINE_ARGS dst, ds, src, ss, w, h, _, my, dx, dy, ss3
%define hm r6m
%endif
- %if required_stack_alignment > STACK_ALIGNMENT
- %define dsm [rsp+0x98]
- %define rX r1
- %define rXd r1d
- %else
- %define dsm dsq
- %define rX r14
- %define rXd r14d
- %endif
+ %define dsm [rsp+0x98]
+ %define rX r1
+ %define rXd r1d
%else ; prep
%if WIN64
mov r7d, hm
@@ -3580,9 +3565,7 @@ cglobal prep_8tap_scaled_16bpc, 4, 14, 16, 0xe0, tmp, src, ss, w, h, mx, my, dx,
; m1=mx, m7=pxmax, m10=h_rnd, m11=h_sh, m12=free
mov myd, mym
%if isput
- %if required_stack_alignment > STACK_ALIGNMENT
- %define dsm [rsp+0xb8]
- %endif
+ %define dsm [rsp+0xb8]
movifnidn dsm, dsq
mova [rsp+0xc0], xm7
%else
diff --git a/src/x86/mc_avx2.asm b/src/x86/mc_avx2.asm
index 2719ef361..3b208033b 100644
--- a/src/x86/mc_avx2.asm
+++ b/src/x86/mc_avx2.asm
@@ -2721,22 +2721,13 @@ cglobal prep_8tap_8bpc, 3, 8, 0, tmp, src, stride, w, h, mx, my, stride3
%macro MC_8TAP_SCALED 1
%ifidn %1, put
%assign isprep 0
- %if required_stack_alignment <= STACK_ALIGNMENT
-cglobal put_8tap_scaled_8bpc, 4, 15, 16, 112, dst, ds, src, ss, w, h, mx, my, dx, dy
- %else
cglobal put_8tap_scaled_8bpc, 4, 14, 16, 128, dst, ds, src, ss, w, h, mx, my, dx, dy
- %endif
%xdefine base_reg r12
%define rndshift 10
%else
%assign isprep 1
- %if required_stack_alignment <= STACK_ALIGNMENT
-cglobal prep_8tap_scaled_8bpc, 4, 15, 16, 128, tmp, src, ss, w, h, mx, my, dx, dy
- %xdefine tmp_stridem r14q
- %else
cglobal prep_8tap_scaled_8bpc, 4, 14, 16, 128, tmp, src, ss, w, h, mx, my, dx, dy
- %define tmp_stridem qword [rsp+120]
- %endif
+ %define tmp_stridem qword [rsp+120]
%xdefine base_reg r11
%define rndshift 6
%endif
@@ -2763,15 +2754,9 @@ cglobal prep_8tap_scaled_8bpc, 4, 14, 16, 128, tmp, src, ss, w, h, mx, my, dx, d
DEFINE_ARGS dst, ds, src, ss, w, h, _, my, dx, dy, ss3
%define hm r6m
%endif
- %if required_stack_alignment > STACK_ALIGNMENT
- %define dsm [rsp+112]
- %define rX r1
- %define rXd r1d
- %else
- %define dsm dsq
- %define rX r14
- %define rXd r14d
- %endif
+ %define dsm [rsp+112]
+ %define rX r1
+ %define rXd r1d
%else ; prep
%if WIN64
mov r7d, hm
diff --git a/src/x86/msac.asm b/src/x86/msac.asm
index 92a3a731d..9f05c921a 100644
--- a/src/x86/msac.asm
+++ b/src/x86/msac.asm
@@ -619,7 +619,6 @@ cglobal msac_decode_symbol_adapt16, 3, 6, 6
mov t3d, [t0+msac.update_cdf]
mov t4d, t2d
not t2
-%if STACK_ALIGNMENT < 32
mov r5, rsp
%if WIN64
and rsp, ~31
@@ -627,11 +626,6 @@ cglobal msac_decode_symbol_adapt16, 3, 6, 6
%else
and r5, ~31
%define buf r5-32
-%endif
-%elif WIN64
- sub rsp, 64
-%else
- %define buf rsp-56
%endif
psrlw m1, m0, 6
movd [buf-4], xm2
@@ -666,11 +660,7 @@ cglobal msac_decode_symbol_adapt16, 3, 6, 6
movzx t2d, word [buf+rax-2]
shr eax, 1
%if WIN64
-%if STACK_ALIGNMENT < 32
mov rsp, r5
-%else
- add rsp, 64
-%endif
%endif
vzeroupper
jmp m(msac_decode_symbol_adapt4, _sse2).renorm2
diff --git a/tests/libfuzzer/meson.build b/tests/libfuzzer/meson.build
index 35914033d..45d28562c 100644
--- a/tests/libfuzzer/meson.build
+++ b/tests/libfuzzer/meson.build
@@ -51,7 +51,6 @@ endif
dav1d_fuzzer = executable('dav1d_fuzzer',
dav1d_fuzzer_sources,
include_directories: dav1d_inc_dirs,
- c_args: [stackalign_flag, stackrealign_flag],
link_args: fuzzer_ldflags,
link_with : libdav1d,
build_by_default: true,
@@ -62,7 +61,7 @@ dav1d_fuzzer = executable('dav1d_fuzzer',
dav1d_fuzzer_mt = executable('dav1d_fuzzer_mt',
dav1d_fuzzer_sources,
include_directories: dav1d_inc_dirs,
- c_args: [stackalign_flag, stackrealign_flag, '-DDAV1D_MT_FUZZING'],
+ c_args: ['-DDAV1D_MT_FUZZING'],
link_args: fuzzer_ldflags,
link_with : libdav1d,
build_by_default: true,
@@ -92,7 +91,7 @@ if (objcopy.found() and
dav1d_fuzzer_mem = executable('dav1d_fuzzer_mem',
dav1d_fuzzer_sources + ['alloc_fail.c'],
include_directories: dav1d_inc_dirs,
- c_args: [stackalign_flag, stackrealign_flag, '-DDAV1D_ALLOC_FAIL'],
+ c_args: ['-DDAV1D_ALLOC_FAIL'],
link_args: fuzzer_ldflags + [join_paths(libdav1d_af.full_path())],
link_depends: libdav1d_af,
build_by_default: false,
diff --git a/tests/meson.build b/tests/meson.build
index 30358d237..f0f0eb713 100644
--- a/tests/meson.build
+++ b/tests/meson.build
@@ -54,7 +54,7 @@ if is_asm_enabled
'checkasm_bitdepth_@0@'.format(bitdepth),
checkasm_tmpl_sources,
include_directories: dav1d_inc_dirs,
- c_args: ['-DBITDEPTH=@0@'.format(bitdepth), stackalign_flag],
+ c_args: ['-DBITDEPTH=@0@'.format(bitdepth)],
install: false,
build_by_default: false,
)
@@ -87,7 +87,6 @@ if is_asm_enabled
],
include_directories: dav1d_inc_dirs,
- c_args: [stackalign_flag, stackrealign_flag],
build_by_default: false,
dependencies : [
thread_dependency,
--
GitLab