aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp1702
1 files changed, 850 insertions, 852 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp
index ddf776107a..452d647bb4 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp
@@ -77,7 +77,6 @@ void a64_hybrid_s8s32_dot_6x16 (
ka.N = N;
ka.B_ptr = B_ptr;
__asm__ __volatile__(
-
"1:" // Row loop
"cmp %x[M], #0x6\n"
"bge 171f\n"
@@ -165,11 +164,11 @@ void a64_hybrid_s8s32_dot_6x16 (
"14:" // Height 1: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 15f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
"cbnz x28, 16f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -186,37 +185,37 @@ void a64_hybrid_s8s32_dot_6x16 (
"blt 18f\n"
"17:" // Height 1: Multiply loop: Main loop head
".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q17, [x10, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ "ldr q16, [x10, #0x30]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ "ldr q17, [x10, #0x40]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ "ldr q16, [x10, #0x50]\n"
+ ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n"
+ "ldr q17, [x10, #0x60]\n"
+ ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n"
+ "ldr q16, [x10, #0x70]\n"
+ ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n"
+ "ldr q17, [x10, #0x80]\n"
+ ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n"
+ "ldr q16, [x10, #0x90]\n"
+ ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n"
+ "ldr q17, [x10, #0xa0]\n"
+ ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n"
+ "ldr q16, [x10, #0xb0]\n"
+ ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n"
+ "ldr q17, [x10, #0xc0]\n"
+ ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n"
+ "ldr q16, [x10, #0xd0]\n"
+ ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n"
+ "ldr q17, [x10, #0xe0]\n"
+ ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n"
+ "ldr q16, [x10, #0xf0]\n"
"sub x27, x27, #0x10\n"
"add x26, x26, #0x10\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
"cmp x27, #0x20\n"
"add x10, x10, #0x100\n"
@@ -226,37 +225,37 @@ void a64_hybrid_s8s32_dot_6x16 (
"bge 17b\n"
"18:" // Height 1: Multiply loop: Single iteration only
".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q17, [x10, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ "ldr q16, [x10, #0x30]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ "ldr q17, [x10, #0x40]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ "ldr q16, [x10, #0x50]\n"
+ ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n"
+ "ldr q17, [x10, #0x60]\n"
+ ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n"
+ "ldr q16, [x10, #0x70]\n"
+ ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n"
+ "ldr q17, [x10, #0x80]\n"
+ ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n"
+ "ldr q16, [x10, #0x90]\n"
+ ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n"
+ "ldr q17, [x10, #0xa0]\n"
+ ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n"
+ "ldr q16, [x10, #0xb0]\n"
+ ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n"
+ "ldr q17, [x10, #0xc0]\n"
+ ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n"
+ "ldr q16, [x10, #0xd0]\n"
+ ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n"
+ "ldr q17, [x10, #0xe0]\n"
+ ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n"
+ "ldr q16, [x10, #0xf0]\n"
"add x26, x26, #0x10\n"
"sub x27, x27, #0x10\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n"
"prfm pldl1keep, [x26, #0x80]\n"
"add x10, x10, #0x100\n"
"19:" // Height 1: Multiply loop: Main loop skip
@@ -264,17 +263,17 @@ void a64_hybrid_s8s32_dot_6x16 (
"cmp x27, #0x4\n"
"blt 21f\n"
"20:" // Height 1: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr q6, [x10, #0x0]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
+ "ldr s18, [x26], #0x4\n"
+ "ldr q16, [x10, #0x0]\n"
+ ".inst 0x4f92e208 // sdot v8.4s, v16.16b, v18.4b[0]\n"
"sub x27, x27, #0x4\n"
- "ldr q7, [x10, #0x10]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
+ "ldr q16, [x10, #0x10]\n"
+ "ldr q17, [x10, #0x20]\n"
+ ".inst 0x4f92e209 // sdot v9.4s, v16.16b, v18.4b[0]\n"
"cmp x27, #0x4\n"
- "ldr q7, [x10, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
+ "ldr q16, [x10, #0x30]\n"
+ ".inst 0x4f92e22a // sdot v10.4s, v17.16b, v18.4b[0]\n"
+ ".inst 0x4f92e20b // sdot v11.4s, v16.16b, v18.4b[0]\n"
"add x10, x10, #0x40\n"
"bge 20b\n"
"21:" // Height 1: Multiply loop: Skip odd blocks
@@ -287,14 +286,14 @@ void a64_hybrid_s8s32_dot_6x16 (
"22:" // Height 1: Multiply loop: Ragged operand read: partial_1_0
"ldr b0, [x26, #0x0]\n"
"23:" // Height 1: Multiply loop: Ragged operand read: Done
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- "ldr q7, [x10, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
+ "ldr q17, [x10, #0x0]\n"
+ "ldr q16, [x10, #0x10]\n"
+ ".inst 0x4f80e228 // sdot v8.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f80e209 // sdot v9.4s, v16.16b, v0.4b[0]\n"
+ "ldr q17, [x10, #0x20]\n"
+ "ldr q16, [x10, #0x30]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
"add x10, x10, #0x40\n"
"24:" // Height 1: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
@@ -465,12 +464,12 @@ void a64_hybrid_s8s32_dot_6x16 (
"48:" // Height 2: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 49f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
"cbnz x28, 50f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -478,7 +477,7 @@ void a64_hybrid_s8s32_dot_6x16 (
"b 50f\n"
"49:" // Height 2: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
+ "add x25, x26, x21\n"
"50:" // Height 2: input setup done
"cmp x27, #0x10\n"
"blt 53f\n"
@@ -491,137 +490,137 @@ void a64_hybrid_s8s32_dot_6x16 (
"51:" // Height 2: Multiply loop: Main loop head
".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q17, [x10, #0x20]\n"
"sub x27, x27, #0x10\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q16, [x10, #0x30]\n"
"add x26, x26, #0x10\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n"
+ "ldr q17, [x10, #0x40]\n"
"add x25, x25, #0x10\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n"
+ "ldr q16, [x10, #0x50]\n"
"cmp x27, #0x20\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
+ ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e22c // sdot v12.4s, v17.16b, v1.4b[1]\n"
+ "ldr q17, [x10, #0x60]\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
+ ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e20d // sdot v13.4s, v16.16b, v1.4b[1]\n"
+ "ldr q16, [x10, #0x70]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e22e // sdot v14.4s, v17.16b, v1.4b[1]\n"
+ "ldr q17, [x10, #0x80]\n"
+ ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e20f // sdot v15.4s, v16.16b, v1.4b[1]\n"
+ "ldr q16, [x10, #0x90]\n"
+ ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea2c // sdot v12.4s, v17.16b, v1.4b[2]\n"
+ "ldr q17, [x10, #0xa0]\n"
+ ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea0d // sdot v13.4s, v16.16b, v1.4b[2]\n"
+ "ldr q16, [x10, #0xb0]\n"
+ ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea2e // sdot v14.4s, v17.16b, v1.4b[2]\n"
+ "ldr q17, [x10, #0xc0]\n"
+ ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea0f // sdot v15.4s, v16.16b, v1.4b[2]\n"
+ "ldr q16, [x10, #0xd0]\n"
+ ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea2c // sdot v12.4s, v17.16b, v1.4b[3]\n"
+ "ldr q17, [x10, #0xe0]\n"
+ ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea0d // sdot v13.4s, v16.16b, v1.4b[3]\n"
+ "ldr q16, [x10, #0xf0]\n"
"add x10, x10, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
+ ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea2e // sdot v14.4s, v17.16b, v1.4b[3]\n"
"ldr q6, [x10, #0x0]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa1ea0f // sdot v15.4s, v16.16b, v1.4b[3]\n"
"ldr q1, [x25, #0x0]\n"
"ldr q7, [x10, #0x10]\n"
"bge 51b\n"
"52:" // Height 2: Multiply loop: Single iteration only
".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q17, [x10, #0x20]\n"
"add x26, x26, #0x10\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q16, [x10, #0x30]\n"
"add x25, x25, #0x10\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n"
+ "ldr q17, [x10, #0x40]\n"
"sub x27, x27, #0x10\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n"
+ "ldr q16, [x10, #0x50]\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
+ ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e22c // sdot v12.4s, v17.16b, v1.4b[1]\n"
+ "ldr q17, [x10, #0x60]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e20d // sdot v13.4s, v16.16b, v1.4b[1]\n"
+ "ldr q16, [x10, #0x70]\n"
+ ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e22e // sdot v14.4s, v17.16b, v1.4b[1]\n"
+ "ldr q17, [x10, #0x80]\n"
+ ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e20f // sdot v15.4s, v16.16b, v1.4b[1]\n"
+ "ldr q16, [x10, #0x90]\n"
+ ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea2c // sdot v12.4s, v17.16b, v1.4b[2]\n"
+ "ldr q17, [x10, #0xa0]\n"
+ ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea0d // sdot v13.4s, v16.16b, v1.4b[2]\n"
+ "ldr q16, [x10, #0xb0]\n"
+ ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea2e // sdot v14.4s, v17.16b, v1.4b[2]\n"
+ "ldr q17, [x10, #0xc0]\n"
+ ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea0f // sdot v15.4s, v16.16b, v1.4b[2]\n"
+ "ldr q16, [x10, #0xd0]\n"
+ ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea2c // sdot v12.4s, v17.16b, v1.4b[3]\n"
+ "ldr q17, [x10, #0xe0]\n"
+ ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea0d // sdot v13.4s, v16.16b, v1.4b[3]\n"
+ "ldr q16, [x10, #0xf0]\n"
"add x10, x10, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea2e // sdot v14.4s, v17.16b, v1.4b[3]\n"
+ ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea0f // sdot v15.4s, v16.16b, v1.4b[3]\n"
"53:" // Height 2: Multiply loop: Main loop skip
"cbz x27, 58f\n"
"cmp x27, #0x4\n"
"blt 55f\n"
"54:" // Height 2: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s19, [x26], #0x4\n"
+ "ldr s18, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ "ldr q17, [x10, #0x0]\n"
+ "ldr q16, [x10, #0x10]\n"
+ ".inst 0x4f93e228 // sdot v8.4s, v17.16b, v19.4b[0]\n"
+ ".inst 0x4f92e22c // sdot v12.4s, v17.16b, v18.4b[0]\n"
+ "ldr q17, [x10, #0x20]\n"
+ ".inst 0x4f93e209 // sdot v9.4s, v16.16b, v19.4b[0]\n"
+ ".inst 0x4f92e20d // sdot v13.4s, v16.16b, v18.4b[0]\n"
+ "ldr q16, [x10, #0x30]\n"
+ ".inst 0x4f93e22a // sdot v10.4s, v17.16b, v19.4b[0]\n"
+ ".inst 0x4f92e22e // sdot v14.4s, v17.16b, v18.4b[0]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f93e20b // sdot v11.4s, v16.16b, v19.4b[0]\n"
+ ".inst 0x4f92e20f // sdot v15.4s, v16.16b, v18.4b[0]\n"
"bge 54b\n"
"55:" // Height 2: Multiply loop: Skip odd blocks
"cbz x27, 58f\n"
@@ -636,19 +635,19 @@ void a64_hybrid_s8s32_dot_6x16 (
"ldr b0, [x26, #0x0]\n"
"ldr b1, [x25, #0x0]\n"
"57:" // Height 2: Multiply loop: Ragged operand read: Done
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ "ldr q17, [x10, #0x0]\n"
+ "ldr q16, [x10, #0x10]\n"
+ ".inst 0x4f80e228 // sdot v8.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f81e22c // sdot v12.4s, v17.16b, v1.4b[0]\n"
+ "ldr q17, [x10, #0x20]\n"
+ ".inst 0x4f80e209 // sdot v9.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e20d // sdot v13.4s, v16.16b, v1.4b[0]\n"
+ "ldr q16, [x10, #0x30]\n"
+ ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n"
+ ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n"
+ ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n"
"58:" // Height 2: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
@@ -866,13 +865,13 @@ void a64_hybrid_s8s32_dot_6x16 (
"82:" // Height 3: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 83f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
"cbnz x28, 84f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -881,8 +880,8 @@ void a64_hybrid_s8s32_dot_6x16 (
"b 84f\n"
"83:" // Height 3: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
- "add x24, x25, x20\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
"84:" // Height 3: input setup done
"cmp x27, #0x10\n"
"blt 87f\n"
@@ -899,75 +898,75 @@ void a64_hybrid_s8s32_dot_6x16 (
"sub x27, x27, #0x10\n"
"add x26, x26, #0x10\n"
".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q21, [x10, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
"add x25, x25, #0x10\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q20, [x10, #0x30]\n"
"add x24, x24, #0x10\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n"
+ ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n"
"cmp x27, #0x20\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
+ ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n"
+ "ldr q21, [x10, #0x40]\n"
+ ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
+ ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n"
+ "ldr q20, [x10, #0x50]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ ".inst 0x4fa0e2a8 // sdot v8.4s, v21.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e2ac // sdot v12.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e2b0 // sdot v16.4s, v21.16b, v2.4b[1]\n"
+ "ldr q21, [x10, #0x60]\n"
+ ".inst 0x4fa0e289 // sdot v9.4s, v20.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e28d // sdot v13.4s, v20.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e291 // sdot v17.4s, v20.16b, v2.4b[1]\n"
+ "ldr q20, [x10, #0x70]\n"
+ ".inst 0x4fa0e2aa // sdot v10.4s, v21.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e2ae // sdot v14.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e2b2 // sdot v18.4s, v21.16b, v2.4b[1]\n"
+ "ldr q21, [x10, #0x80]\n"
+ ".inst 0x4fa0e28b // sdot v11.4s, v20.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e28f // sdot v15.4s, v20.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e293 // sdot v19.4s, v20.16b, v2.4b[1]\n"
+ "ldr q20, [x10, #0x90]\n"
+ ".inst 0x4f80eaa8 // sdot v8.4s, v21.16b, v0.4b[2]\n"
+ ".inst 0x4f81eaac // sdot v12.4s, v21.16b, v1.4b[2]\n"
+ ".inst 0x4f82eab0 // sdot v16.4s, v21.16b, v2.4b[2]\n"
+ "ldr q21, [x10, #0xa0]\n"
+ ".inst 0x4f80ea89 // sdot v9.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea8d // sdot v13.4s, v20.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea91 // sdot v17.4s, v20.16b, v2.4b[2]\n"
+ "ldr q20, [x10, #0xb0]\n"
+ ".inst 0x4f80eaaa // sdot v10.4s, v21.16b, v0.4b[2]\n"
+ ".inst 0x4f81eaae // sdot v14.4s, v21.16b, v1.4b[2]\n"
+ ".inst 0x4f82eab2 // sdot v18.4s, v21.16b, v2.4b[2]\n"
+ "ldr q21, [x10, #0xc0]\n"
+ ".inst 0x4f80ea8b // sdot v11.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea8f // sdot v15.4s, v20.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea93 // sdot v19.4s, v20.16b, v2.4b[2]\n"
+ "ldr q20, [x10, #0xd0]\n"
+ ".inst 0x4fa0eaa8 // sdot v8.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eaac // sdot v12.4s, v21.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eab0 // sdot v16.4s, v21.16b, v2.4b[3]\n"
+ "ldr q21, [x10, #0xe0]\n"
+ ".inst 0x4fa0ea89 // sdot v9.4s, v20.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea8d // sdot v13.4s, v20.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea91 // sdot v17.4s, v20.16b, v2.4b[3]\n"
+ "ldr q20, [x10, #0xf0]\n"
"add x10, x10, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
+ ".inst 0x4fa0eaaa // sdot v10.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eaae // sdot v14.4s, v21.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eab2 // sdot v18.4s, v21.16b, v2.4b[3]\n"
"ldr q6, [x10, #0x0]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0ea8b // sdot v11.4s, v20.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa1ea8f // sdot v15.4s, v20.16b, v1.4b[3]\n"
"ldr q1, [x25, #0x0]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
+ ".inst 0x4fa2ea93 // sdot v19.4s, v20.16b, v2.4b[3]\n"
"ldr q2, [x24, #0x0]\n"
"ldr q7, [x10, #0x10]\n"
"bge 85b\n"
@@ -977,98 +976,98 @@ void a64_hybrid_s8s32_dot_6x16 (
"add x26, x26, #0x10\n"
"add x25, x25, #0x10\n"
".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q21, [x10, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
"add x24, x24, #0x10\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q20, [x10, #0x30]\n"
"sub x27, x27, #0x10\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n"
+ ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n"
"prfm pldl1keep, [x26, #0x80]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
+ ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n"
+ "ldr q21, [x10, #0x40]\n"
+ ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n"
+ "ldr q20, [x10, #0x50]\n"
+ ".inst 0x4fa0e2a8 // sdot v8.4s, v21.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e2ac // sdot v12.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e2b0 // sdot v16.4s, v21.16b, v2.4b[1]\n"
+ "ldr q21, [x10, #0x60]\n"
+ ".inst 0x4fa0e289 // sdot v9.4s, v20.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e28d // sdot v13.4s, v20.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e291 // sdot v17.4s, v20.16b, v2.4b[1]\n"
+ "ldr q20, [x10, #0x70]\n"
+ ".inst 0x4fa0e2aa // sdot v10.4s, v21.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e2ae // sdot v14.4s, v21.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e2b2 // sdot v18.4s, v21.16b, v2.4b[1]\n"
+ "ldr q21, [x10, #0x80]\n"
+ ".inst 0x4fa0e28b // sdot v11.4s, v20.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e28f // sdot v15.4s, v20.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e293 // sdot v19.4s, v20.16b, v2.4b[1]\n"
+ "ldr q20, [x10, #0x90]\n"
+ ".inst 0x4f80eaa8 // sdot v8.4s, v21.16b, v0.4b[2]\n"
+ ".inst 0x4f81eaac // sdot v12.4s, v21.16b, v1.4b[2]\n"
+ ".inst 0x4f82eab0 // sdot v16.4s, v21.16b, v2.4b[2]\n"
+ "ldr q21, [x10, #0xa0]\n"
+ ".inst 0x4f80ea89 // sdot v9.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea8d // sdot v13.4s, v20.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea91 // sdot v17.4s, v20.16b, v2.4b[2]\n"
+ "ldr q20, [x10, #0xb0]\n"
+ ".inst 0x4f80eaaa // sdot v10.4s, v21.16b, v0.4b[2]\n"
+ ".inst 0x4f81eaae // sdot v14.4s, v21.16b, v1.4b[2]\n"
+ ".inst 0x4f82eab2 // sdot v18.4s, v21.16b, v2.4b[2]\n"
+ "ldr q21, [x10, #0xc0]\n"
+ ".inst 0x4f80ea8b // sdot v11.4s, v20.16b, v0.4b[2]\n"
+ ".inst 0x4f81ea8f // sdot v15.4s, v20.16b, v1.4b[2]\n"
+ ".inst 0x4f82ea93 // sdot v19.4s, v20.16b, v2.4b[2]\n"
+ "ldr q20, [x10, #0xd0]\n"
+ ".inst 0x4fa0eaa8 // sdot v8.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eaac // sdot v12.4s, v21.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eab0 // sdot v16.4s, v21.16b, v2.4b[3]\n"
+ "ldr q21, [x10, #0xe0]\n"
+ ".inst 0x4fa0ea89 // sdot v9.4s, v20.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea8d // sdot v13.4s, v20.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea91 // sdot v17.4s, v20.16b, v2.4b[3]\n"
+ "ldr q20, [x10, #0xf0]\n"
"add x10, x10, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
+ ".inst 0x4fa0eaaa // sdot v10.4s, v21.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eaae // sdot v14.4s, v21.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eab2 // sdot v18.4s, v21.16b, v2.4b[3]\n"
+ ".inst 0x4fa0ea8b // sdot v11.4s, v20.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ea8f // sdot v15.4s, v20.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ea93 // sdot v19.4s, v20.16b, v2.4b[3]\n"
"87:" // Height 3: Multiply loop: Main loop skip
"cbz x27, 92f\n"
"cmp x27, #0x4\n"
"blt 89f\n"
"88:" // Height 3: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s24, [x26], #0x4\n"
+ "ldr s23, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr s2, [x24], #0x4\n"
- "ldr q6, [x10, #0x0]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr s22, [x24], #0x4\n"
+ "ldr q21, [x10, #0x0]\n"
+ ".inst 0x4f98e2a8 // sdot v8.4s, v21.16b, v24.4b[0]\n"
+ ".inst 0x4f97e2ac // sdot v12.4s, v21.16b, v23.4b[0]\n"
+ "ldr q20, [x10, #0x10]\n"
+ ".inst 0x4f96e2b0 // sdot v16.4s, v21.16b, v22.4b[0]\n"
+ "ldr q21, [x10, #0x20]\n"
+ ".inst 0x4f98e289 // sdot v9.4s, v20.16b, v24.4b[0]\n"
+ ".inst 0x4f97e28d // sdot v13.4s, v20.16b, v23.4b[0]\n"
+ ".inst 0x4f96e291 // sdot v17.4s, v20.16b, v22.4b[0]\n"
+ "ldr q20, [x10, #0x30]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
+ ".inst 0x4f98e2aa // sdot v10.4s, v21.16b, v24.4b[0]\n"
+ ".inst 0x4f97e2ae // sdot v14.4s, v21.16b, v23.4b[0]\n"
+ ".inst 0x4f96e2b2 // sdot v18.4s, v21.16b, v22.4b[0]\n"
+ ".inst 0x4f98e28b // sdot v11.4s, v20.16b, v24.4b[0]\n"
+ ".inst 0x4f97e28f // sdot v15.4s, v20.16b, v23.4b[0]\n"
+ ".inst 0x4f96e293 // sdot v19.4s, v20.16b, v22.4b[0]\n"
"bge 88b\n"
"89:" // Height 3: Multiply loop: Skip odd blocks
"cbz x27, 92f\n"
@@ -1086,23 +1085,23 @@ void a64_hybrid_s8s32_dot_6x16 (
"ldr b1, [x25, #0x0]\n"
"ldr b2, [x24, #0x0]\n"
"91:" // Height 3: Multiply loop: Ragged operand read: Done
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q21, [x10, #0x0]\n"
+ "ldr q20, [x10, #0x10]\n"
+ ".inst 0x4f80e2a8 // sdot v8.4s, v21.16b, v0.4b[0]\n"
+ ".inst 0x4f81e2ac // sdot v12.4s, v21.16b, v1.4b[0]\n"
+ ".inst 0x4f82e2b0 // sdot v16.4s, v21.16b, v2.4b[0]\n"
+ "ldr q21, [x10, #0x20]\n"
+ ".inst 0x4f80e289 // sdot v9.4s, v20.16b, v0.4b[0]\n"
+ ".inst 0x4f81e28d // sdot v13.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f82e291 // sdot v17.4s, v20.16b, v2.4b[0]\n"
+ "ldr q20, [x10, #0x30]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
+ ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n"
+ ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n"
+ ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n"
+ ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n"
+ ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n"
+ ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n"
"92:" // Height 3: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
@@ -1367,14 +1366,14 @@ void a64_hybrid_s8s32_dot_6x16 (
"116:" // Height 4: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 117f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
"cbnz x28, 118f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -1384,9 +1383,9 @@ void a64_hybrid_s8s32_dot_6x16 (
"b 118f\n"
"117:" // Height 4: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
- "add x24, x25, x20\n"
- "add x23, x24, x20\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
"118:" // Height 4: input setup done
"cmp x27, #0x10\n"
"blt 121f\n"
@@ -1405,7 +1404,7 @@ void a64_hybrid_s8s32_dot_6x16 (
"add x26, x26, #0x10\n"
".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q25, [x10, #0x20]\n"
"add x25, x25, #0x10\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -1413,85 +1412,85 @@ void a64_hybrid_s8s32_dot_6x16 (
"add x23, x23, #0x10\n"
".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q24, [x10, #0x30]\n"
"cmp x27, #0x20\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n"
+ ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n"
"prfm pldl1keep, [x26, #0x80]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
+ ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n"
+ ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n"
+ "ldr q25, [x10, #0x40]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n"
"prfm pldl1keep, [x23, #0x80]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n"
+ "ldr q24, [x10, #0x50]\n"
+ ".inst 0x4fa0e328 // sdot v8.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e32c // sdot v12.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e330 // sdot v16.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e334 // sdot v20.4s, v25.16b, v3.4b[1]\n"
+ "ldr q25, [x10, #0x60]\n"
+ ".inst 0x4fa0e309 // sdot v9.4s, v24.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e30d // sdot v13.4s, v24.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e311 // sdot v17.4s, v24.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e315 // sdot v21.4s, v24.16b, v3.4b[1]\n"
+ "ldr q24, [x10, #0x70]\n"
+ ".inst 0x4fa0e32a // sdot v10.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e32e // sdot v14.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e332 // sdot v18.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e336 // sdot v22.4s, v25.16b, v3.4b[1]\n"
+ "ldr q25, [x10, #0x80]\n"
+ ".inst 0x4fa0e30b // sdot v11.4s, v24.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e30f // sdot v15.4s, v24.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e313 // sdot v19.4s, v24.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e317 // sdot v23.4s, v24.16b, v3.4b[1]\n"
+ "ldr q24, [x10, #0x90]\n"
+ ".inst 0x4f80eb28 // sdot v8.4s, v25.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb2c // sdot v12.4s, v25.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb30 // sdot v16.4s, v25.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb34 // sdot v20.4s, v25.16b, v3.4b[2]\n"
+ "ldr q25, [x10, #0xa0]\n"
+ ".inst 0x4f80eb09 // sdot v9.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb0d // sdot v13.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb11 // sdot v17.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb15 // sdot v21.4s, v24.16b, v3.4b[2]\n"
+ "ldr q24, [x10, #0xb0]\n"
+ ".inst 0x4f80eb2a // sdot v10.4s, v25.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb2e // sdot v14.4s, v25.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb32 // sdot v18.4s, v25.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb36 // sdot v22.4s, v25.16b, v3.4b[2]\n"
+ "ldr q25, [x10, #0xc0]\n"
+ ".inst 0x4f80eb0b // sdot v11.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb0f // sdot v15.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb13 // sdot v19.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb17 // sdot v23.4s, v24.16b, v3.4b[2]\n"
+ "ldr q24, [x10, #0xd0]\n"
+ ".inst 0x4fa0eb28 // sdot v8.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb2c // sdot v12.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb30 // sdot v16.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb34 // sdot v20.4s, v25.16b, v3.4b[3]\n"
+ "ldr q25, [x10, #0xe0]\n"
+ ".inst 0x4fa0eb09 // sdot v9.4s, v24.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb0d // sdot v13.4s, v24.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb11 // sdot v17.4s, v24.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb15 // sdot v21.4s, v24.16b, v3.4b[3]\n"
+ "ldr q24, [x10, #0xf0]\n"
"add x10, x10, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n"
+ ".inst 0x4fa0eb2a // sdot v10.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb2e // sdot v14.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb32 // sdot v18.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb36 // sdot v22.4s, v25.16b, v3.4b[3]\n"
"ldr q6, [x10, #0x0]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0eb0b // sdot v11.4s, v24.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa1eb0f // sdot v15.4s, v24.16b, v1.4b[3]\n"
"ldr q1, [x25, #0x0]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
+ ".inst 0x4fa2eb13 // sdot v19.4s, v24.16b, v2.4b[3]\n"
"ldr q2, [x24, #0x0]\n"
- ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n"
+ ".inst 0x4fa3eb17 // sdot v23.4s, v24.16b, v3.4b[3]\n"
"ldr q3, [x23, #0x0]\n"
"ldr q7, [x10, #0x10]\n"
"bge 119b\n"
@@ -1502,7 +1501,7 @@ void a64_hybrid_s8s32_dot_6x16 (
"add x25, x25, #0x10\n"
".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q25, [x10, #0x20]\n"
"add x24, x24, #0x10\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -1510,112 +1509,112 @@ void a64_hybrid_s8s32_dot_6x16 (
"sub x27, x27, #0x10\n"
".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q24, [x10, #0x30]\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n"
+ ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n"
"prfm pldl1keep, [x25, #0x80]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
+ ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n"
+ ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n"
+ "ldr q25, [x10, #0x40]\n"
"prfm pldl1keep, [x23, #0x80]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n"
+ "ldr q24, [x10, #0x50]\n"
+ ".inst 0x4fa0e328 // sdot v8.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e32c // sdot v12.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e330 // sdot v16.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e334 // sdot v20.4s, v25.16b, v3.4b[1]\n"
+ "ldr q25, [x10, #0x60]\n"
+ ".inst 0x4fa0e309 // sdot v9.4s, v24.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e30d // sdot v13.4s, v24.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e311 // sdot v17.4s, v24.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e315 // sdot v21.4s, v24.16b, v3.4b[1]\n"
+ "ldr q24, [x10, #0x70]\n"
+ ".inst 0x4fa0e32a // sdot v10.4s, v25.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e32e // sdot v14.4s, v25.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e332 // sdot v18.4s, v25.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e336 // sdot v22.4s, v25.16b, v3.4b[1]\n"
+ "ldr q25, [x10, #0x80]\n"
+ ".inst 0x4fa0e30b // sdot v11.4s, v24.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e30f // sdot v15.4s, v24.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e313 // sdot v19.4s, v24.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e317 // sdot v23.4s, v24.16b, v3.4b[1]\n"
+ "ldr q24, [x10, #0x90]\n"
+ ".inst 0x4f80eb28 // sdot v8.4s, v25.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb2c // sdot v12.4s, v25.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb30 // sdot v16.4s, v25.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb34 // sdot v20.4s, v25.16b, v3.4b[2]\n"
+ "ldr q25, [x10, #0xa0]\n"
+ ".inst 0x4f80eb09 // sdot v9.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb0d // sdot v13.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb11 // sdot v17.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb15 // sdot v21.4s, v24.16b, v3.4b[2]\n"
+ "ldr q24, [x10, #0xb0]\n"
+ ".inst 0x4f80eb2a // sdot v10.4s, v25.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb2e // sdot v14.4s, v25.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb32 // sdot v18.4s, v25.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb36 // sdot v22.4s, v25.16b, v3.4b[2]\n"
+ "ldr q25, [x10, #0xc0]\n"
+ ".inst 0x4f80eb0b // sdot v11.4s, v24.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb0f // sdot v15.4s, v24.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb13 // sdot v19.4s, v24.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb17 // sdot v23.4s, v24.16b, v3.4b[2]\n"
+ "ldr q24, [x10, #0xd0]\n"
+ ".inst 0x4fa0eb28 // sdot v8.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb2c // sdot v12.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb30 // sdot v16.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb34 // sdot v20.4s, v25.16b, v3.4b[3]\n"
+ "ldr q25, [x10, #0xe0]\n"
+ ".inst 0x4fa0eb09 // sdot v9.4s, v24.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb0d // sdot v13.4s, v24.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb11 // sdot v17.4s, v24.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb15 // sdot v21.4s, v24.16b, v3.4b[3]\n"
+ "ldr q24, [x10, #0xf0]\n"
"add x10, x10, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n"
+ ".inst 0x4fa0eb2a // sdot v10.4s, v25.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb2e // sdot v14.4s, v25.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb32 // sdot v18.4s, v25.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb36 // sdot v22.4s, v25.16b, v3.4b[3]\n"
+ ".inst 0x4fa0eb0b // sdot v11.4s, v24.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb0f // sdot v15.4s, v24.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb13 // sdot v19.4s, v24.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb17 // sdot v23.4s, v24.16b, v3.4b[3]\n"
"121:" // Height 4: Multiply loop: Main loop skip
"cbz x27, 126f\n"
"cmp x27, #0x4\n"
"blt 123f\n"
"122:" // Height 4: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s29, [x26], #0x4\n"
+ "ldr s28, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr s2, [x24], #0x4\n"
- "ldr s3, [x23], #0x4\n"
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr s27, [x24], #0x4\n"
+ "ldr s26, [x23], #0x4\n"
+ "ldr q25, [x10, #0x0]\n"
+ "ldr q24, [x10, #0x10]\n"
+ ".inst 0x4f9de328 // sdot v8.4s, v25.16b, v29.4b[0]\n"
+ ".inst 0x4f9ce32c // sdot v12.4s, v25.16b, v28.4b[0]\n"
+ ".inst 0x4f9be330 // sdot v16.4s, v25.16b, v27.4b[0]\n"
+ ".inst 0x4f9ae334 // sdot v20.4s, v25.16b, v26.4b[0]\n"
+ "ldr q25, [x10, #0x20]\n"
+ ".inst 0x4f9de309 // sdot v9.4s, v24.16b, v29.4b[0]\n"
+ ".inst 0x4f9ce30d // sdot v13.4s, v24.16b, v28.4b[0]\n"
+ ".inst 0x4f9be311 // sdot v17.4s, v24.16b, v27.4b[0]\n"
+ ".inst 0x4f9ae315 // sdot v21.4s, v24.16b, v26.4b[0]\n"
+ "ldr q24, [x10, #0x30]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
+ ".inst 0x4f9de32a // sdot v10.4s, v25.16b, v29.4b[0]\n"
+ ".inst 0x4f9ce32e // sdot v14.4s, v25.16b, v28.4b[0]\n"
+ ".inst 0x4f9be332 // sdot v18.4s, v25.16b, v27.4b[0]\n"
+ ".inst 0x4f9ae336 // sdot v22.4s, v25.16b, v26.4b[0]\n"
+ ".inst 0x4f9de30b // sdot v11.4s, v24.16b, v29.4b[0]\n"
+ ".inst 0x4f9ce30f // sdot v15.4s, v24.16b, v28.4b[0]\n"
+ ".inst 0x4f9be313 // sdot v19.4s, v24.16b, v27.4b[0]\n"
+ ".inst 0x4f9ae317 // sdot v23.4s, v24.16b, v26.4b[0]\n"
"bge 122b\n"
"123:" // Height 4: Multiply loop: Skip odd blocks
"cbz x27, 126f\n"
@@ -1636,27 +1635,27 @@ void a64_hybrid_s8s32_dot_6x16 (
"ldr b2, [x24, #0x0]\n"
"ldr b3, [x23, #0x0]\n"
"125:" // Height 4: Multiply loop: Ragged operand read: Done
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q25, [x10, #0x0]\n"
+ "ldr q24, [x10, #0x10]\n"
+ ".inst 0x4f80e328 // sdot v8.4s, v25.16b, v0.4b[0]\n"
+ ".inst 0x4f81e32c // sdot v12.4s, v25.16b, v1.4b[0]\n"
+ ".inst 0x4f82e330 // sdot v16.4s, v25.16b, v2.4b[0]\n"
+ ".inst 0x4f83e334 // sdot v20.4s, v25.16b, v3.4b[0]\n"
+ "ldr q25, [x10, #0x20]\n"
+ ".inst 0x4f80e309 // sdot v9.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f81e30d // sdot v13.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f82e311 // sdot v17.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f83e315 // sdot v21.4s, v24.16b, v3.4b[0]\n"
+ "ldr q24, [x10, #0x30]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
+ ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n"
+ ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n"
+ ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n"
+ ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n"
+ ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n"
+ ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n"
+ ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n"
+ ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n"
"126:" // Height 4: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
@@ -1968,15 +1967,15 @@ void a64_hybrid_s8s32_dot_6x16 (
"150:" // Height 5: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 151f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
- "ldr x22, [x21, #0x20]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "ldr x22, [x20, #0x20]\n"
"cbnz x28, 152f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -1987,10 +1986,10 @@ void a64_hybrid_s8s32_dot_6x16 (
"b 152f\n"
"151:" // Height 5: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
- "add x24, x25, x20\n"
- "add x23, x24, x20\n"
- "add x22, x23, x20\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
+ "add x22, x23, x21\n"
"152:" // Height 5: input setup done
"cmp x27, #0x10\n"
"blt 155f\n"
@@ -2013,7 +2012,7 @@ void a64_hybrid_s8s32_dot_6x16 (
"add x25, x25, #0x10\n"
"add x24, x24, #0x10\n"
".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q29, [x10, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
"add x23, x23, #0x10\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2022,100 +2021,100 @@ void a64_hybrid_s8s32_dot_6x16 (
"cmp x27, #0x20\n"
".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q28, [x10, #0x30]\n"
"prfm pldl1keep, [x26, #0x80]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n"
"prfm pldl1keep, [x25, #0x80]\n"
"prfm pldl1keep, [x24, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
+ ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n"
"prfm pldl1keep, [x23, #0x80]\n"
"prfm pldl1keep, [x22, #0x80]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n"
- ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n"
- ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n"
- ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n"
- ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n"
+ "ldr q29, [x10, #0x40]\n"
+ ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n"
+ ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n"
+ "ldr q28, [x10, #0x50]\n"
+ ".inst 0x4fa0e3a8 // sdot v8.4s, v29.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e3ac // sdot v12.4s, v29.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e3b0 // sdot v16.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e3b4 // sdot v20.4s, v29.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e3b8 // sdot v24.4s, v29.16b, v4.4b[1]\n"
+ "ldr q29, [x10, #0x60]\n"
+ ".inst 0x4fa0e389 // sdot v9.4s, v28.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e38d // sdot v13.4s, v28.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e391 // sdot v17.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e395 // sdot v21.4s, v28.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e399 // sdot v25.4s, v28.16b, v4.4b[1]\n"
+ "ldr q28, [x10, #0x70]\n"
+ ".inst 0x4fa0e3aa // sdot v10.4s, v29.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e3ae // sdot v14.4s, v29.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e3b2 // sdot v18.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e3b6 // sdot v22.4s, v29.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e3ba // sdot v26.4s, v29.16b, v4.4b[1]\n"
+ "ldr q29, [x10, #0x80]\n"
+ ".inst 0x4fa0e38b // sdot v11.4s, v28.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e38f // sdot v15.4s, v28.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e393 // sdot v19.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e397 // sdot v23.4s, v28.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e39b // sdot v27.4s, v28.16b, v4.4b[1]\n"
+ "ldr q28, [x10, #0x90]\n"
+ ".inst 0x4f80eba8 // sdot v8.4s, v29.16b, v0.4b[2]\n"
+ ".inst 0x4f81ebac // sdot v12.4s, v29.16b, v1.4b[2]\n"
+ ".inst 0x4f82ebb0 // sdot v16.4s, v29.16b, v2.4b[2]\n"
+ ".inst 0x4f83ebb4 // sdot v20.4s, v29.16b, v3.4b[2]\n"
+ ".inst 0x4f84ebb8 // sdot v24.4s, v29.16b, v4.4b[2]\n"
+ "ldr q29, [x10, #0xa0]\n"
+ ".inst 0x4f80eb89 // sdot v9.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb8d // sdot v13.4s, v28.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb91 // sdot v17.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb95 // sdot v21.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84eb99 // sdot v25.4s, v28.16b, v4.4b[2]\n"
+ "ldr q28, [x10, #0xb0]\n"
+ ".inst 0x4f80ebaa // sdot v10.4s, v29.16b, v0.4b[2]\n"
+ ".inst 0x4f81ebae // sdot v14.4s, v29.16b, v1.4b[2]\n"
+ ".inst 0x4f82ebb2 // sdot v18.4s, v29.16b, v2.4b[2]\n"
+ ".inst 0x4f83ebb6 // sdot v22.4s, v29.16b, v3.4b[2]\n"
+ ".inst 0x4f84ebba // sdot v26.4s, v29.16b, v4.4b[2]\n"
+ "ldr q29, [x10, #0xc0]\n"
+ ".inst 0x4f80eb8b // sdot v11.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb8f // sdot v15.4s, v28.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb93 // sdot v19.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb97 // sdot v23.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84eb9b // sdot v27.4s, v28.16b, v4.4b[2]\n"
+ "ldr q28, [x10, #0xd0]\n"
+ ".inst 0x4fa0eba8 // sdot v8.4s, v29.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ebac // sdot v12.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ebb0 // sdot v16.4s, v29.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ebb4 // sdot v20.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ebb8 // sdot v24.4s, v29.16b, v4.4b[3]\n"
+ "ldr q29, [x10, #0xe0]\n"
+ ".inst 0x4fa0eb89 // sdot v9.4s, v28.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb8d // sdot v13.4s, v28.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb91 // sdot v17.4s, v28.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb95 // sdot v21.4s, v28.16b, v3.4b[3]\n"
+ ".inst 0x4fa4eb99 // sdot v25.4s, v28.16b, v4.4b[3]\n"
+ "ldr q28, [x10, #0xf0]\n"
"add x10, x10, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ebaa // sdot v10.4s, v29.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ebae // sdot v14.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ebb2 // sdot v18.4s, v29.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ebb6 // sdot v22.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ebba // sdot v26.4s, v29.16b, v4.4b[3]\n"
"ldr q6, [x10, #0x0]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
+ ".inst 0x4fa0eb8b // sdot v11.4s, v28.16b, v0.4b[3]\n"
"ldr q0, [x26, #0x0]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
+ ".inst 0x4fa1eb8f // sdot v15.4s, v28.16b, v1.4b[3]\n"
"ldr q1, [x25, #0x0]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
+ ".inst 0x4fa2eb93 // sdot v19.4s, v28.16b, v2.4b[3]\n"
"ldr q2, [x24, #0x0]\n"
- ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n"
+ ".inst 0x4fa3eb97 // sdot v23.4s, v28.16b, v3.4b[3]\n"
"ldr q3, [x23, #0x0]\n"
- ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n"
+ ".inst 0x4fa4eb9b // sdot v27.4s, v28.16b, v4.4b[3]\n"
"ldr q4, [x22, #0x0]\n"
"ldr q7, [x10, #0x10]\n"
"bge 153b\n"
@@ -2129,7 +2128,7 @@ void a64_hybrid_s8s32_dot_6x16 (
"add x24, x24, #0x10\n"
"add x23, x23, #0x10\n"
".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
+ "ldr q29, [x10, #0x20]\n"
".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
"add x22, x22, #0x10\n"
".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
@@ -2138,131 +2137,131 @@ void a64_hybrid_s8s32_dot_6x16 (
"prfm pldl1keep, [x26, #0x80]\n"
".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q28, [x10, #0x30]\n"
"prfm pldl1keep, [x25, #0x80]\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n"
"prfm pldl1keep, [x24, #0x80]\n"
"prfm pldl1keep, [x23, #0x80]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
+ ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n"
"prfm pldl1keep, [x22, #0x80]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x10, #0x40]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x10, #0x50]\n"
- ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n"
- "ldr q6, [x10, #0x60]\n"
- ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n"
- "ldr q7, [x10, #0x70]\n"
- ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n"
- "ldr q6, [x10, #0x80]\n"
- ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n"
- ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n"
- ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n"
- ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n"
- ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n"
- "ldr q7, [x10, #0x90]\n"
- ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n"
- ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n"
- "ldr q6, [x10, #0xa0]\n"
- ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n"
- ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n"
- "ldr q7, [x10, #0xb0]\n"
- ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n"
- ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n"
- ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n"
- ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n"
- "ldr q6, [x10, #0xc0]\n"
- ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n"
- ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n"
- ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n"
- ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n"
- ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n"
- "ldr q7, [x10, #0xd0]\n"
- ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n"
- "ldr q6, [x10, #0xe0]\n"
- ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n"
- "ldr q7, [x10, #0xf0]\n"
+ ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n"
+ "ldr q29, [x10, #0x40]\n"
+ ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n"
+ ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n"
+ "ldr q28, [x10, #0x50]\n"
+ ".inst 0x4fa0e3a8 // sdot v8.4s, v29.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e3ac // sdot v12.4s, v29.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e3b0 // sdot v16.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e3b4 // sdot v20.4s, v29.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e3b8 // sdot v24.4s, v29.16b, v4.4b[1]\n"
+ "ldr q29, [x10, #0x60]\n"
+ ".inst 0x4fa0e389 // sdot v9.4s, v28.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e38d // sdot v13.4s, v28.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e391 // sdot v17.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e395 // sdot v21.4s, v28.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e399 // sdot v25.4s, v28.16b, v4.4b[1]\n"
+ "ldr q28, [x10, #0x70]\n"
+ ".inst 0x4fa0e3aa // sdot v10.4s, v29.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e3ae // sdot v14.4s, v29.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e3b2 // sdot v18.4s, v29.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e3b6 // sdot v22.4s, v29.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e3ba // sdot v26.4s, v29.16b, v4.4b[1]\n"
+ "ldr q29, [x10, #0x80]\n"
+ ".inst 0x4fa0e38b // sdot v11.4s, v28.16b, v0.4b[1]\n"
+ ".inst 0x4fa1e38f // sdot v15.4s, v28.16b, v1.4b[1]\n"
+ ".inst 0x4fa2e393 // sdot v19.4s, v28.16b, v2.4b[1]\n"
+ ".inst 0x4fa3e397 // sdot v23.4s, v28.16b, v3.4b[1]\n"
+ ".inst 0x4fa4e39b // sdot v27.4s, v28.16b, v4.4b[1]\n"
+ "ldr q28, [x10, #0x90]\n"
+ ".inst 0x4f80eba8 // sdot v8.4s, v29.16b, v0.4b[2]\n"
+ ".inst 0x4f81ebac // sdot v12.4s, v29.16b, v1.4b[2]\n"
+ ".inst 0x4f82ebb0 // sdot v16.4s, v29.16b, v2.4b[2]\n"
+ ".inst 0x4f83ebb4 // sdot v20.4s, v29.16b, v3.4b[2]\n"
+ ".inst 0x4f84ebb8 // sdot v24.4s, v29.16b, v4.4b[2]\n"
+ "ldr q29, [x10, #0xa0]\n"
+ ".inst 0x4f80eb89 // sdot v9.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb8d // sdot v13.4s, v28.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb91 // sdot v17.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb95 // sdot v21.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84eb99 // sdot v25.4s, v28.16b, v4.4b[2]\n"
+ "ldr q28, [x10, #0xb0]\n"
+ ".inst 0x4f80ebaa // sdot v10.4s, v29.16b, v0.4b[2]\n"
+ ".inst 0x4f81ebae // sdot v14.4s, v29.16b, v1.4b[2]\n"
+ ".inst 0x4f82ebb2 // sdot v18.4s, v29.16b, v2.4b[2]\n"
+ ".inst 0x4f83ebb6 // sdot v22.4s, v29.16b, v3.4b[2]\n"
+ ".inst 0x4f84ebba // sdot v26.4s, v29.16b, v4.4b[2]\n"
+ "ldr q29, [x10, #0xc0]\n"
+ ".inst 0x4f80eb8b // sdot v11.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4f81eb8f // sdot v15.4s, v28.16b, v1.4b[2]\n"
+ ".inst 0x4f82eb93 // sdot v19.4s, v28.16b, v2.4b[2]\n"
+ ".inst 0x4f83eb97 // sdot v23.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84eb9b // sdot v27.4s, v28.16b, v4.4b[2]\n"
+ "ldr q28, [x10, #0xd0]\n"
+ ".inst 0x4fa0eba8 // sdot v8.4s, v29.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ebac // sdot v12.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ebb0 // sdot v16.4s, v29.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ebb4 // sdot v20.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ebb8 // sdot v24.4s, v29.16b, v4.4b[3]\n"
+ "ldr q29, [x10, #0xe0]\n"
+ ".inst 0x4fa0eb89 // sdot v9.4s, v28.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb8d // sdot v13.4s, v28.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb91 // sdot v17.4s, v28.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb95 // sdot v21.4s, v28.16b, v3.4b[3]\n"
+ ".inst 0x4fa4eb99 // sdot v25.4s, v28.16b, v4.4b[3]\n"
+ "ldr q28, [x10, #0xf0]\n"
"add x10, x10, #0x100\n"
- ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n"
- ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n"
- ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n"
- ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n"
- ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n"
- ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n"
+ ".inst 0x4fa0ebaa // sdot v10.4s, v29.16b, v0.4b[3]\n"
+ ".inst 0x4fa1ebae // sdot v14.4s, v29.16b, v1.4b[3]\n"
+ ".inst 0x4fa2ebb2 // sdot v18.4s, v29.16b, v2.4b[3]\n"
+ ".inst 0x4fa3ebb6 // sdot v22.4s, v29.16b, v3.4b[3]\n"
+ ".inst 0x4fa4ebba // sdot v26.4s, v29.16b, v4.4b[3]\n"
+ ".inst 0x4fa0eb8b // sdot v11.4s, v28.16b, v0.4b[3]\n"
+ ".inst 0x4fa1eb8f // sdot v15.4s, v28.16b, v1.4b[3]\n"
+ ".inst 0x4fa2eb93 // sdot v19.4s, v28.16b, v2.4b[3]\n"
+ ".inst 0x4fa3eb97 // sdot v23.4s, v28.16b, v3.4b[3]\n"
+ ".inst 0x4fa4eb9b // sdot v27.4s, v28.16b, v4.4b[3]\n"
"155:" // Height 5: Multiply loop: Main loop skip
"cbz x27, 160f\n"
"cmp x27, #0x4\n"
"blt 157f\n"
"156:" // Height 5: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
+ "ldr s2, [x26], #0x4\n"
"ldr s1, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr s2, [x24], #0x4\n"
- "ldr s3, [x23], #0x4\n"
- "ldr s4, [x22], #0x4\n"
- "ldr q6, [x10, #0x0]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr s0, [x24], #0x4\n"
+ "ldr s31, [x23], #0x4\n"
+ "ldr s30, [x22], #0x4\n"
+ "ldr q29, [x10, #0x0]\n"
+ ".inst 0x4f82e3a8 // sdot v8.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f81e3ac // sdot v12.4s, v29.16b, v1.4b[0]\n"
+ "ldr q28, [x10, #0x10]\n"
+ ".inst 0x4f80e3b0 // sdot v16.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f9fe3b4 // sdot v20.4s, v29.16b, v31.4b[0]\n"
+ ".inst 0x4f9ee3b8 // sdot v24.4s, v29.16b, v30.4b[0]\n"
+ "ldr q29, [x10, #0x20]\n"
+ ".inst 0x4f82e389 // sdot v9.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f81e38d // sdot v13.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f80e391 // sdot v17.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f9fe395 // sdot v21.4s, v28.16b, v31.4b[0]\n"
+ ".inst 0x4f9ee399 // sdot v25.4s, v28.16b, v30.4b[0]\n"
+ "ldr q28, [x10, #0x30]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
+ ".inst 0x4f82e3aa // sdot v10.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n"
+ ".inst 0x4f80e3b2 // sdot v18.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f9fe3b6 // sdot v22.4s, v29.16b, v31.4b[0]\n"
+ ".inst 0x4f9ee3ba // sdot v26.4s, v29.16b, v30.4b[0]\n"
+ ".inst 0x4f82e38b // sdot v11.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f80e393 // sdot v19.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f9fe397 // sdot v23.4s, v28.16b, v31.4b[0]\n"
+ ".inst 0x4f9ee39b // sdot v27.4s, v28.16b, v30.4b[0]\n"
"bge 156b\n"
"157:" // Height 5: Multiply loop: Skip odd blocks
"cbz x27, 160f\n"
@@ -2286,31 +2285,31 @@ void a64_hybrid_s8s32_dot_6x16 (
"ldr b3, [x23, #0x0]\n"
"ldr b4, [x22, #0x0]\n"
"159:" // Height 5: Multiply loop: Ragged operand read: Done
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q29, [x10, #0x0]\n"
+ "ldr q28, [x10, #0x10]\n"
+ ".inst 0x4f80e3a8 // sdot v8.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f81e3ac // sdot v12.4s, v29.16b, v1.4b[0]\n"
+ ".inst 0x4f82e3b0 // sdot v16.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f83e3b4 // sdot v20.4s, v29.16b, v3.4b[0]\n"
+ ".inst 0x4f84e3b8 // sdot v24.4s, v29.16b, v4.4b[0]\n"
+ "ldr q29, [x10, #0x20]\n"
+ ".inst 0x4f80e389 // sdot v9.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f81e38d // sdot v13.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f82e391 // sdot v17.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f83e395 // sdot v21.4s, v28.16b, v3.4b[0]\n"
+ ".inst 0x4f84e399 // sdot v25.4s, v28.16b, v4.4b[0]\n"
+ "ldr q28, [x10, #0x30]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
+ ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n"
+ ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n"
+ ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n"
+ ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n"
+ ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n"
+ ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n"
+ ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n"
+ ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n"
+ ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n"
"160:" // Height 5: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
@@ -2672,16 +2671,16 @@ void a64_hybrid_s8s32_dot_6x16 (
"184:" // Height 6: String loop
"ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
"ldr w27, [x20, x28, LSL #0x2]\n"
- "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
"tbz %x[flags], #3, 185f\n"
- "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n"
- "add x21, x21, x20, LSL #3\n"
- "ldr x26, [x21, #0x0]\n"
- "ldr x25, [x21, #0x8]\n"
- "ldr x24, [x21, #0x10]\n"
- "ldr x23, [x21, #0x18]\n"
- "ldr x22, [x21, #0x20]\n"
- "ldr x21, [x21, #0x28]\n"
+ "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x26, [x20, #0x0]\n"
+ "ldr x25, [x20, #0x8]\n"
+ "ldr x24, [x20, #0x10]\n"
+ "ldr x23, [x20, #0x18]\n"
+ "ldr x22, [x20, #0x20]\n"
+ "ldr x21, [x20, #0x28]\n"
"cbnz x28, 186f\n"
"ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
"add x26, x26, x20\n"
@@ -2693,11 +2692,11 @@ void a64_hybrid_s8s32_dot_6x16 (
"b 186f\n"
"185:" // Height 6: setup direct input
"mov x26, %x[input_ptr]\n"
- "add x25, x26, x20\n"
- "add x24, x25, x20\n"
- "add x23, x24, x20\n"
- "add x22, x23, x20\n"
- "add x21, x22, x20\n"
+ "add x25, x26, x21\n"
+ "add x24, x25, x21\n"
+ "add x23, x24, x21\n"
+ "add x22, x23, x21\n"
+ "add x21, x22, x21\n"
"186:" // Height 6: input setup done
"cmp x27, #0x10\n"
"blt 189f\n"
@@ -2976,43 +2975,43 @@ void a64_hybrid_s8s32_dot_6x16 (
"cmp x27, #0x4\n"
"blt 191f\n"
"190:" // Height 6: Multiply loop: Odd block loop
- "ldr s0, [x26], #0x4\n"
- "ldr s1, [x25], #0x4\n"
+ "ldr s7, [x26], #0x4\n"
+ "ldr s6, [x25], #0x4\n"
"sub x27, x27, #0x4\n"
"cmp x27, #0x4\n"
- "ldr s2, [x24], #0x4\n"
- "ldr s3, [x23], #0x4\n"
- "ldr s4, [x22], #0x4\n"
- "ldr s5, [x21], #0x4\n"
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr s5, [x24], #0x4\n"
+ "ldr s4, [x23], #0x4\n"
+ "ldr s3, [x22], #0x4\n"
+ "ldr s2, [x21], #0x4\n"
+ "ldr q1, [x10, #0x0]\n"
+ "ldr q0, [x10, #0x10]\n"
+ ".inst 0x4f87e028 // sdot v8.4s, v1.16b, v7.4b[0]\n"
+ ".inst 0x4f86e02c // sdot v12.4s, v1.16b, v6.4b[0]\n"
+ ".inst 0x4f85e030 // sdot v16.4s, v1.16b, v5.4b[0]\n"
+ ".inst 0x4f84e034 // sdot v20.4s, v1.16b, v4.4b[0]\n"
+ ".inst 0x4f83e038 // sdot v24.4s, v1.16b, v3.4b[0]\n"
+ ".inst 0x4f82e03c // sdot v28.4s, v1.16b, v2.4b[0]\n"
+ "ldr q1, [x10, #0x20]\n"
+ ".inst 0x4f87e009 // sdot v9.4s, v0.16b, v7.4b[0]\n"
+ ".inst 0x4f86e00d // sdot v13.4s, v0.16b, v6.4b[0]\n"
+ ".inst 0x4f85e011 // sdot v17.4s, v0.16b, v5.4b[0]\n"
+ ".inst 0x4f84e015 // sdot v21.4s, v0.16b, v4.4b[0]\n"
+ ".inst 0x4f83e019 // sdot v25.4s, v0.16b, v3.4b[0]\n"
+ ".inst 0x4f82e01d // sdot v29.4s, v0.16b, v2.4b[0]\n"
+ "ldr q0, [x10, #0x30]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
- ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n"
+ ".inst 0x4f87e02a // sdot v10.4s, v1.16b, v7.4b[0]\n"
+ ".inst 0x4f86e02e // sdot v14.4s, v1.16b, v6.4b[0]\n"
+ ".inst 0x4f85e032 // sdot v18.4s, v1.16b, v5.4b[0]\n"
+ ".inst 0x4f84e036 // sdot v22.4s, v1.16b, v4.4b[0]\n"
+ ".inst 0x4f83e03a // sdot v26.4s, v1.16b, v3.4b[0]\n"
+ ".inst 0x4f82e03e // sdot v30.4s, v1.16b, v2.4b[0]\n"
+ ".inst 0x4f87e00b // sdot v11.4s, v0.16b, v7.4b[0]\n"
+ ".inst 0x4f86e00f // sdot v15.4s, v0.16b, v6.4b[0]\n"
+ ".inst 0x4f85e013 // sdot v19.4s, v0.16b, v5.4b[0]\n"
+ ".inst 0x4f84e017 // sdot v23.4s, v0.16b, v4.4b[0]\n"
+ ".inst 0x4f83e01b // sdot v27.4s, v0.16b, v3.4b[0]\n"
+ ".inst 0x4f82e01f // sdot v31.4s, v0.16b, v2.4b[0]\n"
"bge 190b\n"
"191:" // Height 6: Multiply loop: Skip odd blocks
"cbz x27, 194f\n"
@@ -3039,35 +3038,35 @@ void a64_hybrid_s8s32_dot_6x16 (
"ldr b4, [x22, #0x0]\n"
"ldr b5, [x21, #0x0]\n"
"193:" // Height 6: Multiply loop: Ragged operand read: Done
- "ldr q6, [x10, #0x0]\n"
- "ldr q7, [x10, #0x10]\n"
- ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n"
- "ldr q6, [x10, #0x20]\n"
- ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n"
- ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n"
- "ldr q7, [x10, #0x30]\n"
+ "ldr q7, [x10, #0x0]\n"
+ "ldr q6, [x10, #0x10]\n"
+ ".inst 0x4f80e0e8 // sdot v8.4s, v7.16b, v0.4b[0]\n"
+ ".inst 0x4f81e0ec // sdot v12.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f82e0f0 // sdot v16.4s, v7.16b, v2.4b[0]\n"
+ ".inst 0x4f83e0f4 // sdot v20.4s, v7.16b, v3.4b[0]\n"
+ ".inst 0x4f84e0f8 // sdot v24.4s, v7.16b, v4.4b[0]\n"
+ ".inst 0x4f85e0fc // sdot v28.4s, v7.16b, v5.4b[0]\n"
+ "ldr q7, [x10, #0x20]\n"
+ ".inst 0x4f80e0c9 // sdot v9.4s, v6.16b, v0.4b[0]\n"
+ ".inst 0x4f81e0cd // sdot v13.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f82e0d1 // sdot v17.4s, v6.16b, v2.4b[0]\n"
+ ".inst 0x4f83e0d5 // sdot v21.4s, v6.16b, v3.4b[0]\n"
+ ".inst 0x4f84e0d9 // sdot v25.4s, v6.16b, v4.4b[0]\n"
+ ".inst 0x4f85e0dd // sdot v29.4s, v6.16b, v5.4b[0]\n"
+ "ldr q6, [x10, #0x30]\n"
"add x10, x10, #0x40\n"
- ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n"
- ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n"
- ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n"
- ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n"
- ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n"
- ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n"
- ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n"
- ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n"
- ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n"
- ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n"
- ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n"
+ ".inst 0x4f80e0ea // sdot v10.4s, v7.16b, v0.4b[0]\n"
+ ".inst 0x4f81e0ee // sdot v14.4s, v7.16b, v1.4b[0]\n"
+ ".inst 0x4f82e0f2 // sdot v18.4s, v7.16b, v2.4b[0]\n"
+ ".inst 0x4f83e0f6 // sdot v22.4s, v7.16b, v3.4b[0]\n"
+ ".inst 0x4f84e0fa // sdot v26.4s, v7.16b, v4.4b[0]\n"
+ ".inst 0x4f85e0fe // sdot v30.4s, v7.16b, v5.4b[0]\n"
+ ".inst 0x4f80e0cb // sdot v11.4s, v6.16b, v0.4b[0]\n"
+ ".inst 0x4f81e0cf // sdot v15.4s, v6.16b, v1.4b[0]\n"
+ ".inst 0x4f82e0d3 // sdot v19.4s, v6.16b, v2.4b[0]\n"
+ ".inst 0x4f83e0d7 // sdot v23.4s, v6.16b, v3.4b[0]\n"
+ ".inst 0x4f84e0db // sdot v27.4s, v6.16b, v4.4b[0]\n"
+ ".inst 0x4f85e0df // sdot v31.4s, v6.16b, v5.4b[0]\n"
"194:" // Height 6: Multiply loop: No odd multiplies
"ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
"add x28, x28, #0x1\n"
@@ -3254,7 +3253,6 @@ void a64_hybrid_s8s32_dot_6x16 (
"madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
"b 1b\n"
"206:" // Exit
-
: [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
: [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"