From 9e7e5c1185537c774099392b3d9a2948ff6b17e9 Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Thu, 19 Sep 2019 22:18:43 +0200
Subject: [PATCH] Add gcc7-generated assembler version of caxpy for power8

to work around wrong code generated by gcc 8.3
---
 kernel/power/caxpy_power8.S | 460 ++++++++++++++++++++++++++++++++++++
 1 file changed, 460 insertions(+)
 create mode 100644 kernel/power/caxpy_power8.S

diff --git a/kernel/power/caxpy_power8.S b/kernel/power/caxpy_power8.S
new file mode 100644
index 000000000..02d694391
--- /dev/null
+++ b/kernel/power/caxpy_power8.S
@@ -0,0 +1,460 @@
+	.file	"caxpy.c"
+	.abiversion 2
+	.section	".text"
+	.align 2
+	.p2align 4,,15
+	.globl caxpy_k
+	.type	caxpy_k, @function
+caxpy_k:
+.LCF0:
+0:	addis 2,12,.TOC.-.LCF0@ha
+	addi 2,2,.TOC.-.LCF0@l
+	.localentry	caxpy_k,.-caxpy_k
+	mr. 7,3
+	ble 0,.L33
+	cmpdi 7,9,1
+	beq 7,.L41
+.L3:
+	mtctr 7
+	ld 7,96(1)
+	sldi 9,9,3
+	sldi 7,7,3
+	.p2align 4,,15
+.L14:
+	lfs 10,4(8)
+	lfs 11,0(8)
+	lfs 12,0(10)
+	lfs 0,4(10)
+	fmuls 10,2,10
+	fmsubs 11,11,1,10
+	fadds 12,12,11
+	stfs 12,0(10)
+	lfs 11,0(8)
+	lfs 12,4(8)
+	add 8,8,9
+	fmuls 11,2,11
+	fmadds 12,12,1,11
+	fadds 0,0,12
+	stfs 0,4(10)
+	add 10,10,7
+	bdnz .L14
+.L33:
+	li 3,0
+	blr
+	.p2align 4,,15
+.L41:
+	ld 6,96(1)
+	cmpdi 7,6,1
+	bne 7,.L3
+	rldicr. 4,7,0,59
+	std 31,-8(1)
+	li 11,0
+	bne 0,.L42
+.L4:
+	addi 6,11,8
+	subf 0,4,7
+	sldi 6,6,2
+	addi 9,6,-32
+	add 5,10,6
+	add 3,8,9
+	add 6,8,6
+	subfc 5,5,3
+	add 9,10,9
+	subfe 5,5,5
+	subfc 6,6,9
+	subfe 31,31,31
+	addi 6,5,1
+	addi 5,31,1
+	or 6,6,5
+	rlwinm 6,6,0,0xff
+	cmpwi 7,6,0
+	beq 7,.L7
+	sradi 6,4,63
+	srdi 5,7,63
+	subfc 31,7,4
+	adde 6,5,6
+	subfic 31,0,3
+	subfe 31,31,31
+	xori 6,6,0x1
+	neg 31,31
+	and 6,6,31
+	rlwinm 6,6,0,0xff
+	cmpwi 7,6,0
+	beq 7,.L7
+	cmpd 7,4,7
+	li 6,1
+	blt 7,.L43
+.L9:
+	addi 0,7,-1
+	subf 0,4,0
+	subfic 0,0,3
+	subfe 31,31,31
+	addi 0,31,1
+	rlwinm 0,0,0,0xff
+	cmpwi 7,0,0
+	bne 7,.L10
+	sradi 0,4,63
+	subfc 31,7,4
+	adde 5,5,0
+	rlwinm 5,5,0,0xff
+	cmpwi 7,5,0
+	bne 7,.L10
+	addi 0,6,-1
+	addis 31,2,.LC3@toc@ha
+	std 30,-16(1)
+	xscvdpspn 12,1
+	xscvdpspn 11,2
+	srdi. 30,0,2
+	addis 6,2,.LC2@toc@ha
+	addi 6,6,.LC2@toc@l
+	mtctr 30
+	addi 31,31,.LC3@toc@l
+	lxvd2x 42,0,6
+	li 5,16
+	li 6,0
+	lxvd2x 41,0,31
+	xxspltw 12,12,0
+	xxspltw 11,11,0
+	xxpermdi 42,42,42,2
+	xxpermdi 41,41,41,2
+	beq 0,.L44
+	.p2align 4,,15
+.L11:
+	lxvd2x 45,3,6
+	lxvd2x 33,3,5
+	lxvd2x 43,9,6
+	lxvd2x 0,9,5
+	xxpermdi 45,45,45,2
+	xxpermdi 33,33,33,2
+	xxpermdi 32,43,43,2
+	xxpermdi 43,0,0,2
+	vperm 12,1,13,10
+	vperm 1,1,13,9
+	vperm 13,11,0,10
+	vperm 11,11,0,9
+	xvmulsp 0,11,44
+	xvmulsp 32,11,33
+	xvmaddmsp 33,12,0
+	xvmsubasp 32,12,44
+	xvaddsp 45,32,45
+	xvaddsp 32,33,43
+	vmrglw 1,0,13
+	vmrghw 0,0,13
+	xxpermdi 0,33,33,2
+	xxpermdi 32,32,32,2
+	stxvd2x 0,9,6
+	addi 6,6,32
+	stxvd2x 32,9,5
+	addi 5,5,32
+	bdnz .L11
+	rldicr 0,0,0,61
+	ld 30,-16(1)
+	sldi 9,0,1
+	add 4,4,0
+	add 11,11,9
+.L10:
+	sldi 6,11,2
+	addi 9,4,1
+	addi 5,6,4
+	cmpd 7,7,9
+	lfsx 12,8,6
+	lfsx 0,10,6
+	addi 9,11,2
+	lfsx 11,8,5
+	fmuls 11,2,11
+	fmsubs 12,12,1,11
+	fadds 0,0,12
+	stfsx 0,10,6
+	lfsx 11,8,6
+	lfsx 12,8,5
+	lfsx 0,10,5
+	fmuls 11,2,11
+	fmadds 12,12,1,11
+	fadds 0,0,12
+	stfsx 0,10,5
+	ble 7,.L39
+	sldi 9,9,2
+	addi 6,4,2
+	addi 5,9,4
+	cmpd 7,7,6
+	lfsx 12,8,9
+	lfsx 0,10,9
+	addi 6,11,4
+	lfsx 11,8,5
+	fmuls 11,2,11
+	fmsubs 12,1,12,11
+	fadds 0,0,12
+	stfsx 0,10,9
+	lfsx 11,8,9
+	lfsx 12,8,5
+	lfsx 0,10,5
+	fmuls 11,2,11
+	fmadds 12,1,12,11
+	fadds 0,0,12
+	stfsx 0,10,5
+	ble 7,.L39
+	sldi 6,6,2
+	addi 4,4,3
+	addi 5,6,4
+	cmpd 7,7,4
+	lfsx 12,8,6
+	lfsx 0,10,6
+	addi 9,11,6
+	lfsx 11,8,5
+	fmuls 11,2,11
+	fmsubs 12,1,12,11
+	fadds 0,0,12
+	stfsx 0,10,6
+	lfsx 11,8,6
+	lfsx 12,8,5
+	lfsx 0,10,5
+	fmuls 11,2,11
+	fmadds 12,1,12,11
+	fadds 0,0,12
+	stfsx 0,10,5
+	ble 7,.L39
+	sldi 9,9,2
+	ld 31,-8(1)
+	addi 7,9,4
+	lfsx 12,8,9
+	lfsx 0,10,9
+	lfsx 11,8,7
+	fmuls 11,2,11
+	fmsubs 12,1,12,11
+	fadds 0,0,12
+	stfsx 0,10,9
+	lfsx 11,8,9
+	lfsx 12,8,7
+	lfsx 0,10,7
+	fmuls 2,2,11
+	fmadds 1,1,12,2
+	fadds 1,0,1
+	stfsx 1,10,7
+	b .L33
+.L43:
+	mr 6,0
+	b .L9
+.L7:
+	addi 10,4,1
+	cmpd 7,10,7
+	subf 10,4,7
+	mtctr 10
+	bgt 7,.L26
+	li 10,-1
+	rldicr 10,10,0,0
+	cmpd 7,7,10
+	beq 7,.L26
+	.p2align 4,,15
+.L13:
+	lfs 10,4(3)
+	lfs 11,0(3)
+	addi 9,9,8
+	addi 3,3,8
+	lfs 12,-8(9)
+	lfs 0,-4(9)
+	fmuls 10,2,10
+	fmsubs 11,1,11,10
+	fadds 12,12,11
+	stfs 12,-8(9)
+	lfs 11,-8(3)
+	lfs 12,-4(3)
+	fmuls 11,2,11
+	fmadds 12,1,12,11
+	fadds 0,0,12
+	stfs 0,-4(9)
+	bdnz .L13
+.L39:
+	ld 31,-8(1)
+	b .L33
+.L42:
+	fneg 0,2
+	xxpermdi 39,2,2,0
+	addis 9,2,.LANCHOR0@toc@ha
+	std 28,-32(1)
+	sradi. 28,4,1
+	addi 9,9,.LANCHOR0@toc@l
+	xscvdpspn 5,1
+	xvcvdpsp 39,39
+	lxvd2x 12,0,9
+	xxpermdi 32,0,0,0
+	xxspltw 5,5,0
+	xvcvdpsp 32,32
+	xxpermdi 12,12,12,2
+	vmrgew 7,7,0
+	beq 0,.L5
+	xxlnor 38,12,12
+	std 29,-24(1)
+	std 30,-16(1)
+	mr 6,8
+	mr 9,10
+	li 29,0
+	li 30,16
+	li 31,32
+	li 12,48
+	li 0,64
+	li 11,80
+	li 3,96
+	li 5,112
+	.p2align 4,,15
+.L6:
+	lxvd2x 6,0,9
+	lxvd2x 40,0,6
+	addi 29,29,8
+	lxvd2x 41,6,30
+	lxvd2x 42,6,31
+	cmpd 7,28,29
+	lxvd2x 43,6,12
+	lxvd2x 44,6,0
+	lxvd2x 45,6,11
+	lxvd2x 33,6,3
+	lxvd2x 32,6,5
+	lxvd2x 7,9,30
+	addi 6,6,128
+	lxvd2x 8,9,31
+	lxvd2x 9,9,12
+	xxpermdi 40,40,40,2
+	xxpermdi 6,6,6,2
+	lxvd2x 10,9,0
+	lxvd2x 11,9,11
+	xxpermdi 41,41,41,2
+	xxpermdi 42,42,42,2
+	lxvd2x 12,9,3
+	lxvd2x 0,9,5
+	xxpermdi 43,43,43,2
+	xxpermdi 44,44,44,2
+	xxpermdi 45,45,45,2
+	xxpermdi 33,33,33,2
+	xxpermdi 32,32,32,2
+	xxpermdi 7,7,7,2
+	xxpermdi 8,8,8,2
+	xxpermdi 9,9,9,2
+	xxpermdi 10,10,10,2
+	xxpermdi 11,11,11,2
+	xxpermdi 12,12,12,2
+	xxpermdi 0,0,0,2
+	xvmaddasp 6,5,40
+	xvmaddasp 7,5,41
+	xvmaddasp 8,5,42
+	xvmaddasp 9,5,43
+	xvmaddasp 10,5,44
+	xvmaddasp 11,5,45
+	xvmaddasp 12,5,33
+	xvmaddasp 0,5,32
+	vperm 8,8,8,6
+	vperm 9,9,9,6
+	vperm 10,10,10,6
+	vperm 11,11,11,6
+	vperm 12,12,12,6
+	vperm 13,13,13,6
+	vperm 1,1,1,6
+	vperm 0,0,0,6
+	xvmaddasp 6,39,40
+	xvmaddasp 7,39,41
+	xvmaddasp 8,39,42
+	xvmaddasp 9,39,43
+	xvmaddasp 10,39,44
+	xvmaddasp 11,39,45
+	xvmaddasp 12,39,33
+	xvmaddasp 0,39,32
+	xxpermdi 6,6,6,2
+	xxpermdi 7,7,7,2
+	xxpermdi 8,8,8,2
+	xxpermdi 9,9,9,2
+	stxvd2x 6,0,9
+	xxpermdi 10,10,10,2
+	stxvd2x 7,9,30
+	xxpermdi 11,11,11,2
+	stxvd2x 8,9,31
+	xxpermdi 12,12,12,2
+	stxvd2x 9,9,12
+	xxpermdi 0,0,0,2
+	stxvd2x 10,9,0
+	stxvd2x 11,9,11
+	stxvd2x 12,9,3
+	stxvd2x 0,9,5
+	addi 9,9,128
+	bgt 7,.L6
+	ld 29,-24(1)
+	ld 30,-16(1)
+.L5:
+	cmpd 7,7,4
+	ble 7,.L36
+	sldi 11,4,1
+	ld 28,-32(1)
+	b .L4
+.L36:
+	ld 28,-32(1)
+	ld 31,-8(1)
+	b .L33
+.L44:
+	li 31,1
+	mtctr 31
+	b .L11
+.L26:
+	li 10,1
+	mtctr 10
+	b .L13
+	.long 0
+	.byte 0,0,0,0,0,4,0,0
+	.size	caxpy_k,.-caxpy_k
+	.section	.rodata
+	.align 4
+	.set	.LANCHOR0,. + 0
+	.type	swap_mask_arr, @object
+	.size	swap_mask_arr, 16
+swap_mask_arr:
+	.byte	4
+	.byte	5
+	.byte	6
+	.byte	7
+	.byte	0
+	.byte	1
+	.byte	2
+	.byte	3
+	.byte	12
+	.byte	13
+	.byte	14
+	.byte	15
+	.byte	8
+	.byte	9
+	.byte	10
+	.byte	11
+	.section	.rodata.cst16,"aM",@progbits,16
+	.align 4
+.LC2:
+	.byte	31
+	.byte	30
+	.byte	29
+	.byte	28
+	.byte	23
+	.byte	22
+	.byte	21
+	.byte	20
+	.byte	15
+	.byte	14
+	.byte	13
+	.byte	12
+	.byte	7
+	.byte	6
+	.byte	5
+	.byte	4
+.LC3:
+	.byte	27
+	.byte	26
+	.byte	25
+	.byte	24
+	.byte	19
+	.byte	18
+	.byte	17
+	.byte	16
+	.byte	11
+	.byte	10
+	.byte	9
+	.byte	8
+	.byte	3
+	.byte	2
+	.byte	1
+	.byte	0
+	.ident	"GCC: (SUSE Linux) 7.3.1 20180323 [gcc-7-branch revision 258812]"
+	.gnu_attribute 4, 1
+	.section	.note.GNU-stack,"",@progbits