summaryrefslogtreecommitdiff
path: root/vere/ext/gmp/gen/aarch64-macos/mpn/rsh1add_n.s
diff options
context:
space:
mode:
authorpolwex <polwex@sortug.com>2025-10-05 21:56:51 +0700
committerpolwex <polwex@sortug.com>2025-10-05 21:56:51 +0700
commitfcedfddf00b3f994e4f4e40332ac7fc192c63244 (patch)
tree51d38e62c7bdfcc5f9a5e9435fe820c93cfc9a3d /vere/ext/gmp/gen/aarch64-macos/mpn/rsh1add_n.s
claude is gud
Diffstat (limited to 'vere/ext/gmp/gen/aarch64-macos/mpn/rsh1add_n.s')
-rw-r--r--vere/ext/gmp/gen/aarch64-macos/mpn/rsh1add_n.s175
1 files changed, 175 insertions, 0 deletions
diff --git a/vere/ext/gmp/gen/aarch64-macos/mpn/rsh1add_n.s b/vere/ext/gmp/gen/aarch64-macos/mpn/rsh1add_n.s
new file mode 100644
index 0000000..0e46013
--- /dev/null
+++ b/vere/ext/gmp/gen/aarch64-macos/mpn/rsh1add_n.s
@@ -0,0 +1,175 @@
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ .text
+ .align 3
+ .globl ___gmpn_rsh1add_n
+
+___gmpn_rsh1add_n:
+ lsr x6, x3, #2
+
+ tbz x3, #0, Lbx0
+
+Lbx1: ldr x5, [x1],#8
+ ldr x9, [x2],#8
+ tbnz x3, #1, Lb11
+
+Lb01: adds x13, x5, x9
+ and x10, x13, #1
+ cbz x6, L1
+ ldp x4, x5, [x1],#48
+ ldp x8, x9, [x2],#48
+ adcs x14, x4, x8
+ adcs x15, x5, x9
+ ldp x4, x5, [x1,#-32]
+ ldp x8, x9, [x2,#-32]
+ extr x17, x14, x13, #1
+ adcs x12, x4, x8
+ adcs x13, x5, x9
+ str x17, [x0], #24
+ sub x6, x6, #1
+ cbz x6, Lend
+ b Ltop
+
+L1: cset x14, cs
+ extr x17, x14, x13, #1
+ str x17, [x0]
+ mov x0, x10
+ ret
+
+Lb11: adds x15, x5, x9
+ and x10, x15, #1
+
+ ldp x4, x5, [x1],#32
+ ldp x8, x9, [x2],#32
+ adcs x12, x4, x8
+ adcs x13, x5, x9
+ cbz x6, L3
+ ldp x4, x5, [x1,#-16]
+ ldp x8, x9, [x2,#-16]
+ extr x17, x12, x15, #1
+ adcs x14, x4, x8
+ adcs x15, x5, x9
+ str x17, [x0], #8
+ b Lmid
+
+L3: extr x17, x12, x15, #1
+ str x17, [x0], #8
+ b L2
+
+Lbx0: tbz x3, #1, Lb00
+
+Lb10: ldp x4, x5, [x1],#32
+ ldp x8, x9, [x2],#32
+ adds x12, x4, x8
+ adcs x13, x5, x9
+ and x10, x12, #1
+ cbz x6, L2
+ ldp x4, x5, [x1,#-16]
+ ldp x8, x9, [x2,#-16]
+ adcs x14, x4, x8
+ adcs x15, x5, x9
+ b Lmid
+
+Lb00: ldp x4, x5, [x1],#48
+ ldp x8, x9, [x2],#48
+ adds x14, x4, x8
+ adcs x15, x5, x9
+ and x10, x14, #1
+ ldp x4, x5, [x1,#-32]
+ ldp x8, x9, [x2,#-32]
+ adcs x12, x4, x8
+ adcs x13, x5, x9
+ add x0, x0, #16
+ sub x6, x6, #1
+ cbz x6, Lend
+
+ .align 4
+Ltop: ldp x4, x5, [x1,#-16]
+ ldp x8, x9, [x2,#-16]
+ extr x16, x15, x14, #1
+ extr x17, x12, x15, #1
+ adcs x14, x4, x8
+ adcs x15, x5, x9
+ stp x16, x17, [x0,#-16]
+Lmid: ldp x4, x5, [x1],#32
+ ldp x8, x9, [x2],#32
+ extr x16, x13, x12, #1
+ extr x17, x14, x13, #1
+ adcs x12, x4, x8
+ adcs x13, x5, x9
+ stp x16, x17, [x0],#32
+ sub x6, x6, #1
+ cbnz x6, Ltop
+
+Lend: extr x16, x15, x14, #1
+ extr x17, x12, x15, #1
+ stp x16, x17, [x0,#-16]
+L2: cset x14, cs
+ extr x16, x13, x12, #1
+ extr x17, x14, x13, #1
+ stp x16, x17, [x0]
+
+Lret: mov x0, x10
+ ret
+