ref: c2292efcb35f45bced8ea13172ac8fbd6d612a38
parent: 7cb756ea5131285637259d264925c608a24d3443
author: Henrik Gramner <gramner@twoorioles.com>
date: Sun Dec 16 12:57:38 EST 2018
Implement support for PIC in x86-32 asm Convert all existing 32-bit SSSE3 asm to use PIC.
--- a/meson.build
+++ b/meson.build
@@ -256,13 +256,12 @@
cdata.set10('ARCH_X86_64', true) cdata_asm.set10('ARCH_X86_32', false) cdata.set10('ARCH_X86_32', false)-
- cdata_asm.set10('PIC', true)else
cdata_asm.set10('ARCH_X86_64', false) cdata.set10('ARCH_X86_64', false) cdata_asm.set10('ARCH_X86_32', true) cdata.set10('ARCH_X86_32', true)+ cdata_asm.set10('PIC', true)endif
else
cdata.set10('ARCH_X86', false)--- a/src/ext/x86/x86inc.asm
+++ b/src/ext/x86/x86inc.asm
@@ -89,16 +89,13 @@
%endif
%endmacro
-%if WIN64
- %define PIC
-%elif ARCH_X86_64 == 0
-; x86_32 doesn't require PIC.
-; Some distros prefer shared objects to be PIC, but nothing breaks if
-; the code contains a few textrels, so we'll skip that complexity.
- %undef PIC
-%endif
-%ifdef PIC
+%if ARCH_X86_64
+ %define PIC 1 ; always use PIC on x86-64
default rel
+%elifidn __OUTPUT_FORMAT__,win32
+ %define PIC 0 ; PIC isn't used on 32-bit Windows
+%elifndef PIC
+ %define PIC 0
%endif
%ifdef __NASM_VER__
@@ -219,6 +216,18 @@
%else
%define gprsize 4
%endif
+
+%macro LEA 2
+%if ARCH_X86_64
+ lea %1, [%2]
+%elif PIC
+ call $+5 ; special-cased to not affect the RSB on most CPU:s
+ pop %1
+ add %1, (%2)-$+1
+%else
+ mov %1, %2
+%endif
+%endmacro
%macro PUSH 1
push %1
--- a/src/x86/ipred_ssse3.asm
+++ b/src/x86/ipred_ssse3.asm
@@ -93,7 +93,7 @@
INIT_XMM ssse3
cglobal ipred_h, 3, 6, 2, dst, stride, tl, w, h, stride3
- lea r5, [ipred_h_ssse3_table]
+ LEA r5, ipred_h_ssse3_table
tzcnt wd, wm
movifnidn hd, hm
movsxd wq, [r5+wq*4]
--- a/src/x86/itx_ssse3.asm
+++ b/src/x86/itx_ssse3.asm
@@ -55,9 +55,15 @@
%define m(x) mangle(private_prefix %+ _ %+ x %+ SUFFIX)
+%if ARCH_X86_64
+%define o(x) x
+%else
+%define o(x) r5-$$+x ; PIC
+%endif
+
%macro ITX4_END 4-5 2048 ; row[1-4], rnd
%if %5
- mova m2, [pw_%5]
+ mova m2, [o(pw_%5)]
pmulhrsw m0, m2
pmulhrsw m1, m2
%endif
@@ -100,7 +106,6 @@
ret
%endmacro
-
; flags: 1 = swap, 2: coef_regs
%macro ITX_MUL2X_PACK 5-6 0 ; dst/src, tmp[1], rnd, coef[1-2], flags
%if %6 & 2
@@ -107,11 +112,11 @@
pmaddwd m%2, m%4, m%1
pmaddwd m%1, m%5
%elif %6 & 1
- pmaddwd m%2, m%1, [pw_%5_%4]
+ pmaddwd m%2, m%1, [o(pw_%5_%4)]
pmaddwd m%1, [pw_%4_m%5]
%else
- pmaddwd m%2, m%1, [pw_%4_m%5]
- pmaddwd m%1, [pw_%5_%4]
+ pmaddwd m%2, m%1, [o(pw_%4_m%5)]
+ pmaddwd m%1, [o(pw_%5_%4)]
%endif
paddd m%2, m%3
paddd m%1, m%3
@@ -126,13 +131,13 @@
paddw m0, m1
punpcklqdq m0, m3 ;high: in0-in2 ;low: in0+in2
- mova m3, [pd_2048]
+ mova m3, [o(pd_2048)]
ITX_MUL2X_PACK 2, 1, 3, 1567, 3784
%if %0 == 1
pmulhrsw m0, m%1
%else
- pmulhrsw m0, [pw_2896x8] ;high: t1 ;low: t0
+ pmulhrsw m0, [o(pw_2896x8)] ;high: t1 ;low: t0
%endif
psubsw m1, m0, m2 ;high: out2 ;low: out3
@@ -146,15 +151,14 @@
punpckhqdq m1, m1 ;
paddw m1, m0 ;low: in0 - in2 + in3
- pmaddwd m0, m2, [pw_1321_3803] ;1321 * in0 + 3803 * in2
- pmaddwd m2, [pw_2482_m1321] ;2482 * in0 - 1321 * in2
- pmaddwd m4, m3, [pw_3344_2482] ;3344 * in1 + 2482 * in3
- pmaddwd m5, m3, [pw_3344_m3803] ;3344 * in1 - 3803 * in3
+ pmaddwd m0, m2, [o(pw_1321_3803)] ;1321 * in0 + 3803 * in2
+ pmaddwd m2, [o(pw_2482_m1321)] ;2482 * in0 - 1321 * in2
+ pmaddwd m4, m3, [o(pw_3344_2482)] ;3344 * in1 + 2482 * in3
+ pmaddwd m5, m3, [o(pw_3344_m3803)];3344 * in1 - 3803 * in3
paddd m4, m0 ;t0 + t3
-
- pmaddwd m3, [pw_m6688_m3803] ;-2 * 3344 * in1 - 3803 * in3
- pmulhrsw m1, [pw_3344x8] ;low: out2
- mova m0, [pd_2048]
+ pmaddwd m3, [o(pw_m6688_m3803)] ;-2 * 3344 * in1 - 3803 * in3
+ pmulhrsw m1, [o(pw_3344x8)] ;low: out2
+ mova m0, [o(pd_2048)]
paddd m2, m0
paddd m0, m4 ;t0 + t3 + 2048
paddd m5, m2 ;t1 + t3 + 2048
@@ -169,9 +173,11 @@
%endmacro
%macro INV_TXFM_FN 4 ; type1, type2, fast_thresh, size
-cglobal inv_txfm_add_%1_%2_%4, 4, 5, 0, dst, stride, coeff, eob, tx2
+cglobal inv_txfm_add_%1_%2_%4, 4, 6, 0, dst, stride, coeff, eob, tx2
%undef cmp
- lea tx2q, [m(i%2_%4_internal).pass2]
+%if ARCH_X86_32
+ LEA r5, $$
+%endif
%if %3 > 0
cmp eobd, %3
jle %%end
@@ -179,7 +185,8 @@
test eobd, eobd
jz %%end
%endif
- call i%1_%4_internal
+ lea tx2q, [o(m(i%2_%4_internal).pass2)]
+ call m(i%1_%4_internal)
RET
ALIGN function_align
%%end:
@@ -188,10 +195,10 @@
%macro INV_TXFM_4X4_FN 2-3 -1 ; type1, type2, fast_thresh
INV_TXFM_FN %1, %2, %3, 4x4
%ifidn %1_%2, dct_identity
- mova m0, [pw_2896x8]
+ mova m0, [o(pw_2896x8)]
pmulhrsw m0, [coeffq]
paddw m0, m0
- pmulhrsw m0, [pw_5793x4]
+ pmulhrsw m0, [o(pw_5793x4)]
punpcklwd m0, m0
punpckhdq m1, m0, m0
punpckldq m0, m0
@@ -205,8 +212,8 @@
punpcklwd m0, m1
punpcklqdq m0, m0
paddw m0, m0
- pmulhrsw m0, [pw_5793x4]
- pmulhrsw m0, [pw_2896x8]
+ pmulhrsw m0, [o(pw_5793x4)]
+ pmulhrsw m0, [o(pw_2896x8)]
mova m1, m0
call m(iadst_4x4_internal).end
RET
@@ -214,17 +221,17 @@
pshuflw m0, [coeffq], q0000
punpcklqdq m0, m0
%ifidn %1, dct
- mova m1, [pw_2896x8]
+ mova m1, [o(pw_2896x8)]
pmulhrsw m0, m1
%elifidn %1, adst
- pmulhrsw m0, [iadst4_dconly1a]
+ pmulhrsw m0, [o(iadst4_dconly1a)]
%elifidn %1, flipadst
- pmulhrsw m0, [iadst4_dconly1b]
+ pmulhrsw m0, [o(iadst4_dconly1b)]
%endif
mov [coeffq], eobd ;0
%ifidn %2, dct
%ifnidn %1, dct
- pmulhrsw m0, [pw_2896x8]
+ pmulhrsw m0, [o(pw_2896x8)]
%else
pmulhrsw m0, m1
%endif
@@ -232,8 +239,8 @@
call m(iadst_4x4_internal).end2
RET
%else ; adst / flipadst
- pmulhrsw m1, m0, [iadst4_dconly2b]
- pmulhrsw m0, [iadst4_dconly2a]
+ pmulhrsw m1, m0, [o(iadst4_dconly2b)]
+ pmulhrsw m0, [o(iadst4_dconly2a)]
call m(i%2_4x4_internal).end2
RET
%endif
@@ -240,9 +247,13 @@
%endif
%endmacro
-
INIT_XMM ssse3
+INV_TXFM_4X4_FN dct, dct, 0
+INV_TXFM_4X4_FN dct, adst, 0
+INV_TXFM_4X4_FN dct, flipadst, 0
+INV_TXFM_4X4_FN dct, identity, 3
+
cglobal idct_4x4_internal, 0, 0, 4, dst, stride, coeff, eob, tx2
mova m0, [coeffq+16*0] ;high: in1 ;low: in0
mova m1, [coeffq+16*1] ;high: in3 ;low in2
@@ -249,7 +260,7 @@
IDCT4_1D_PACKED
- mova m2, [deint_shuf]
+ mova m2, [o(deint_shuf)]
shufps m3, m0, m1, q1331
shufps m0, m1, q0220
pshufb m0, m2 ;high: in1 ;low: in0
@@ -265,7 +276,10 @@
ITX4_END 0, 1, 3, 2
-INV_TXFM_4X4_FN dct, dct, 0
+INV_TXFM_4X4_FN adst, dct, 0
+INV_TXFM_4X4_FN adst, adst, 0
+INV_TXFM_4X4_FN adst, flipadst, 0
+INV_TXFM_4X4_FN adst, identity
cglobal iadst_4x4_internal, 0, 0, 6, dst, stride, coeff, eob, tx2
mova m0, [coeffq+16*0]
@@ -294,9 +308,10 @@
IADST4_1D_PACKED
ret
-INV_TXFM_4X4_FN adst, adst, 0
-INV_TXFM_4X4_FN dct, adst, 0
-INV_TXFM_4X4_FN adst, dct, 0
+INV_TXFM_4X4_FN flipadst, dct, 0
+INV_TXFM_4X4_FN flipadst, adst, 0
+INV_TXFM_4X4_FN flipadst, flipadst, 0
+INV_TXFM_4X4_FN flipadst, identity
cglobal iflipadst_4x4_internal, 0, 0, 6, dst, stride, coeff, eob, tx2
mova m0, [coeffq+16*0]
@@ -321,16 +336,15 @@
.end2:
ITX4_END 3, 2, 1, 0
-INV_TXFM_4X4_FN flipadst, flipadst, 0
-INV_TXFM_4X4_FN flipadst, dct, 0
-INV_TXFM_4X4_FN flipadst, adst, 0
-INV_TXFM_4X4_FN dct, flipadst, 0
-INV_TXFM_4X4_FN adst, flipadst, 0
+INV_TXFM_4X4_FN identity, dct, 3
+INV_TXFM_4X4_FN identity, adst
+INV_TXFM_4X4_FN identity, flipadst
+INV_TXFM_4X4_FN identity, identity
cglobal iidentity_4x4_internal, 0, 0, 6, dst, stride, coeff, eob, tx2
mova m0, [coeffq+16*0]
mova m1, [coeffq+16*1]
- mova m2, [pw_5793x4]
+ mova m2, [o(pw_5793x4)]
paddw m0, m0
paddw m1, m1
pmulhrsw m0, m2
@@ -343,20 +357,12 @@
jmp tx2q
.pass2:
- mova m2, [pw_5793x4]
+ mova m2, [o(pw_5793x4)]
paddw m0, m0
paddw m1, m1
pmulhrsw m0, m2
pmulhrsw m1, m2
jmp m(iadst_4x4_internal).end
-
-INV_TXFM_4X4_FN identity, identity
-INV_TXFM_4X4_FN identity, dct, 3
-INV_TXFM_4X4_FN identity, adst
-INV_TXFM_4X4_FN identity, flipadst
-INV_TXFM_4X4_FN dct, identity, 3
-INV_TXFM_4X4_FN adst, identity
-INV_TXFM_4X4_FN flipadst, identity
%macro IWHT4_1D_PACKED 0
punpckhqdq m3, m0, m1 ;low: in1 high: in3
--- a/src/x86/mc_ssse3.asm
+++ b/src/x86/mc_ssse3.asm
@@ -186,7 +186,7 @@
%endmacro
cglobal avg, 4, 7, 3, dst, stride, tmp1, tmp2, w, h, stride3
- lea r6, [avg_ssse3_table]
+ LEA r6, avg_ssse3_table
tzcnt wd, wm ; leading zeros
movifnidn hd, hm ; move h(stack) to h(register) if not already that register
movsxd wq, dword [r6+wq*4] ; push table entry matching the tile width (tzcnt) in widen reg
@@ -216,7 +216,7 @@
%define W_AVG_INC_PTR AVG_INC_PTR
cglobal w_avg, 4, 7, 6, dst, stride, tmp1, tmp2, w, h, stride3
- lea r6, [w_avg_ssse3_table]
+ LEA r6, w_avg_ssse3_table
tzcnt wd, wm
movifnidn hd, hm
movd m0, r6m
@@ -269,11 +269,12 @@
cglobal mask, 4, 7, 7, dst, stride, tmp1, tmp2, w, mask, stride3
%define hd dword r5m
%endif
- lea r6, [mask_ssse3_table]
+%define base r6-mask_ssse3_table
+ LEA r6, mask_ssse3_table
tzcnt wd, wm
movsxd wq, dword [r6+wq*4]
pxor m4, m4
- mova m5, [pw_2048+r6-mask_ssse3_table]
+ mova m5, [base+pw_2048]
add wq, r6
mov maskq, r6m
BIDIR_FN MASK
@@ -284,9 +285,9 @@
%define reg_pw_27 m9
%define reg_pw_2048 m10
%else
- %define reg_pw_8 [pw_8]
- %define reg_pw_27 [pw_26] ; 64 - 38
- %define reg_pw_2048 [pw_2048]
+ %define reg_pw_8 [base+pw_8]
+ %define reg_pw_27 [base+pw_26] ; 64 - 38
+ %define reg_pw_2048 [base+pw_2048]
%endif
%macro W_MASK_420_B 2 ; src_offset in bytes, mask_out
@@ -323,63 +324,60 @@
W_MASK_420_B (%1*16), %2
%endmacro
+%define base r6-w_mask_420_ssse3_table
%if ARCH_X86_64
; args: dst, stride, tmp1, tmp2, w, h, mask, sign
-cglobal w_mask_420, 4, 9, 11, dst, stride, tmp1, tmp2, w, h, mask, stride3
- lea r7, [w_mask_420_ssse3_table]
+cglobal w_mask_420, 4, 8, 11, dst, stride, tmp1, tmp2, w, h, mask
+ lea r6, [w_mask_420_ssse3_table]
mov wd, wm
- tzcnt r8d, wd
+ tzcnt r7d, wd
movifnidn hd, hm
- mov maskq, maskmp
movd m0, r7m
pshuflw m0, m0, q0000 ; sign
punpcklqdq m0, m0
- movsxd r8, dword [r7+r8*4]
- mova reg_pw_8, [pw_8]
- mova reg_pw_27, [pw_26] ; 64 - 38
- mova reg_pw_2048, [pw_2048]
- mova m6, [pw_258] ; 64 * 4 + 2
+ movsxd r7, [r6+r7*4]
+ mova reg_pw_8, [base+pw_8]
+ mova reg_pw_27, [base+pw_26] ; 64 - 38
+ mova reg_pw_2048, [base+pw_2048]
+ mova m6, [base+pw_258] ; 64 * 4 + 2
+ add r7, r6
+ mov maskq, maskmp
psubw m6, m0
- add r8, r7
W_MASK_420 0, 4
- lea stride3q, [strideq*3]
- jmp r8
- %define dst_bak r8
- %define loop_w r7
- %define orig_w wq
+ jmp r7
+ %define loop_w r7d
%else
-cglobal w_mask_420, 4, 7, 8, dst, stride, tmp1, tmp2, w, mask, stride3
- tzcnt r6d, r4m
- mov wd, w_mask_420_ssse3_table
- add wd, [wq+r6*4]
+cglobal w_mask_420, 4, 7, 8, dst, stride, tmp1, tmp2, w, mask
+ tzcnt wd, wm
+ LEA r6, w_mask_420_ssse3_table
+ mov wd, [r6+wq*4]
mov maskq, r6mp
movd m0, r7m
pshuflw m0, m0, q0000 ; sign
punpcklqdq m0, m0
- mova m6, [pw_258] ; 64 * 4 + 2
+ mova m6, [base+pw_258] ; 64 * 4 + 2
+ add wq, r6
psubw m6, m0
W_MASK_420 0, 4
- lea stride3q, [strideq*3]
jmp wd
- %define dst_bak r0m
- %define loop_w r6q
- %define orig_w r4m
- %define hd dword r5m
+ %define loop_w dword r0m
+ %define hd dword r5m
%endif
.w4_loop:
add tmp1q, 2*16
add tmp2q, 2*16
W_MASK_420 0, 4
- lea dstq, [dstq+strideq*4]
+ lea dstq, [dstq+strideq*2]
add maskq, 4
.w4:
movd [dstq ], m0 ; copy m0[0]
pshuflw m1, m0, q1032
movd [dstq+strideq*1], m1 ; copy m0[1]
+ lea dstq, [dstq+strideq*2]
punpckhqdq m0, m0
- movd [dstq+strideq*2], m0 ; copy m0[2]
+ movd [dstq+strideq*0], m0 ; copy m0[2]
psrlq m0, 32
- movd [dstq+stride3q ], m0 ; copy m0[3]
+ movd [dstq+strideq*1], m0 ; copy m0[3]
pshufd m5, m4, q3131; DBDB even lines repeated
pshufd m4, m4, q2020; CACA odd lines repeated
psubw m1, m6, m4 ; m9 == 64 * 4 + 2
@@ -409,20 +407,19 @@
jg .w8_loop
RET
.w16: ; w32/64/128
- mov dst_bak, dstq
- mov loop_w, orig_w ; use width as counter
%if ARCH_X86_32
- mov wq, orig_w ; because we altered it in 32bit setup
+ mov wd, wm ; because we altered it in 32bit setup
%endif
+ mov loop_w, wd ; use width as counter
jmp .w16ge_inner_loop_first
.w16ge_loop:
lea tmp1q, [tmp1q+wq*2] ; skip even line pixels
lea tmp2q, [tmp2q+wq*2] ; skip even line pixels
+ sub dstq, wq
+ mov loop_w, wd
lea dstq, [dstq+strideq*2]
- mov dst_bak, dstq
- mov loop_w, orig_w
.w16ge_inner_loop:
- W_MASK_420_B 0, 4
+ W_MASK_420_B 0, 4
.w16ge_inner_loop_first:
mova [dstq ], m0
W_MASK_420_B wq*2, 5 ; load matching even line (offset = widthpx * (16+16))
@@ -438,7 +435,6 @@
add dstq, 16
sub loop_w, 16
jg .w16ge_inner_loop
- mov dstq, dst_bak
sub hd, 2
jg .w16ge_loop
RET
@@ -470,7 +466,7 @@
cglobal blend, 3, 7, 7, dst, ds, tmp, w, h, mask
%define base r6-blend_ssse3_table
- lea r6, [blend_ssse3_table]
+ LEA r6, blend_ssse3_table
tzcnt wd, wm
movifnidn hd, hm
movifnidn maskq, maskmp
@@ -546,7 +542,7 @@
cglobal blend_v, 3, 6, 8, dst, ds, tmp, w, h, mask
%define base r5-blend_v_ssse3_table
- lea r5, [blend_v_ssse3_table]
+ LEA r5, blend_v_ssse3_table
tzcnt wd, wm
movifnidn hd, hm
movsxd wq, dword [r5+wq*4]
@@ -646,15 +642,21 @@
jg .w32_loop
RET
-cglobal blend_h, 4, 7, 6, dst, ds, tmp, w, h, mask
-%define base r5-blend_h_ssse3_table
- lea r5, [blend_h_ssse3_table]
+cglobal blend_h, 3, 7, 6, dst, ds, tmp, w, h, mask
+%define base t0-blend_h_ssse3_table
+%if ARCH_X86_32
+ ; We need to keep the PIC pointer for w4, reload wd from stack instead
+ DECLARE_REG_TMP 6
+%else
+ DECLARE_REG_TMP 5
mov r6d, wd
- tzcnt wd, wd
+%endif
+ LEA t0, blend_h_ssse3_table
+ tzcnt wd, wm
mov hd, hm
- movsxd wq, dword [r5+wq*4]
+ movsxd wq, dword [t0+wq*4]
mova m5, [base+pw_512]
- add wq, r5
+ add wq, t0
lea maskq, [base+obmc_masks+hq*4]
neg hq
jmp wq
@@ -678,7 +680,11 @@
jl .w2
RET
.w4:
+%if ARCH_X86_32
+ mova m3, [base+blend_shuf]
+%else
mova m3, [blend_shuf]
+%endif
.w4_loop:
movd m0, [dstq+dsq*0]
movd m2, [dstq+dsq*1]
@@ -716,6 +722,9 @@
RET
; w16/w32/w64/w128
.w16:
+%if ARCH_X86_32
+ mov r6d, wm
+%endif
sub dsq, r6
.w16_loop0:
movd m3, [maskq+hq*2]
--- a/tests/checkasm/x86/checkasm.asm
+++ b/tests/checkasm/x86/checkasm.asm
@@ -200,7 +200,7 @@
jz .ok
mov r3, eax
mov r4, edx
- lea r0, [error_message]
+ LEA r0, error_message
mov [esp], r0
call fail_func
mov edx, r4