diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp | 268 |
1 files changed, 134 insertions, 134 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp index e2ff9a214e..063084eb3c 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -87,228 +87,228 @@ void sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl( activation_min, activation_max); __asm__ __volatile__( - "ldr x20, [%x[params_struct], %[offsetof_args_outptrs]]\n" + "ldr x19, [%x[params_struct], %[offsetof_args_outptrs]]\n" ".inst 0xd503477f // SMSTART ZA\n" - "add x16, %x[params_struct], %[offsetof_Args_inptrs]\n" + "add x15, %x[params_struct], %[offsetof_Args_inptrs]\n" "ptrue p3.b\n" - "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n" + "ldr x14, [%x[params_struct], %[offsetof_args_params]]\n" ".inst 0x25207810 // ptrue pn8.b\n" - "ld1w { z19.s }, p3/Z, [x15]\n" - "addvl x15, x15, #1\n" - "ldp x14, x13, [x20, #0x0]\n" - "cntw x12\n" - ".inst 0xa040c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15]\n" - "addvl x15, x15, #4\n" - "ldp x11, x10, [x20, #0x10]\n" - "mov x9, #0x0\n" + "ld1w { z19.s }, p3/Z, [x14]\n" + "addvl x14, x14, #1\n" + "ldp x13, x12, [x19, #0x0]\n" + "cntw x11\n" + ".inst 0xa040c1c0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x14]\n" + "addvl x14, x14, #4\n" + "ldp x10, x9, [x19, #0x10]\n" + "mov x28, #0x0\n" "whilelt p2.s, XZR, %x[n_channels]\n" - ".inst 0xa040c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n" - "ldp x28, x27, [x16, #0x0]\n" - "addvl x15, x15, #4\n" - "cmp x12, %x[n_channels]\n" + ".inst 0xa040c1c4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x14]\n" + "ldp x27, x26, [x15, #0x0]\n" + "addvl x14, x14, #4\n" + "cmp x11, %x[n_channels]\n" "ld1rw { z18.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" - "ldp x26, x25, [x16, #0x10]\n" + "ldp x25, x24, [x15, #0x10]\n" "ld1rw { z17.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" - "sub x24, XZR, x12\n" - "ldp x23, x22, [x16, #0x20]\n" - "ld1w { z8.s }, p3/Z, [x15]\n" - "addvl x15, x15, #1\n" - "ldp x21, x20, [x16, #0x30]\n" - "ld1w { z9.s }, p2/Z, [x28, x9, LSL #2]\n" - "ld1w { z10.s }, p2/Z, [x27, x9, LSL #2]\n" - "ld1w { z11.s }, p2/Z, [x26, x9, LSL #2]\n" - "ld1w { z12.s }, p2/Z, [x25, x9, LSL #2]\n" - "ld1w { z13.s }, p2/Z, [x23, x9, LSL #2]\n" - "ld1w { z14.s }, p2/Z, [x22, x9, LSL #2]\n" - "ld1w { z15.s }, p2/Z, [x21, x9, LSL #2]\n" - "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n" + "sub x23, XZR, x11\n" + "ldp x22, x21, [x15, #0x20]\n" + "ld1w { z8.s }, p3/Z, [x14]\n" + "addvl x14, x14, #1\n" + "ldp x20, x19, [x15, #0x30]\n" + "ld1w { z9.s }, p2/Z, [x27, x28, LSL #2]\n" + "ld1w { z10.s }, p2/Z, [x26, x28, LSL #2]\n" + "ld1w { z11.s }, p2/Z, [x25, x28, LSL #2]\n" + "ld1w { z12.s }, p2/Z, [x24, x28, LSL #2]\n" + "ld1w { z13.s }, p2/Z, [x22, x28, LSL #2]\n" + "ld1w { z14.s }, p2/Z, [x21, x28, LSL #2]\n" + "ld1w { z15.s }, p2/Z, [x20, x28, LSL #2]\n" + "ld1w { z16.s }, p2/Z, [x19, x28, LSL #2]\n" "bge 2f\n" "1:" // Channel loop "movprfx z28, z19\n fmla z28.s, p3/M, z8.s, z9.s\n" "movprfx z29, z19\n fmla z29.s, p3/M, z6.s, z9.s\n" - "ldr x28, [x16, #0x40]\n" - "whilelt p1.s, x12, %x[n_channels]\n" + "ldr x27, [x15, #0x40]\n" + "whilelt p1.s, x11, %x[n_channels]\n" "fmla z28.s, p3/M, z0.s, z10.s\n" "fmla z29.s, p3/M, z1.s, z12.s\n" - "ldr x27, [x16, #0x48]\n" - "ld1w { z12.s }, p2/Z, [x27, x9, LSL #2]\n" + "ldr x26, [x15, #0x48]\n" + "ld1w { z12.s }, p2/Z, [x26, x28, LSL #2]\n" "fmla z28.s, p3/M, z1.s, z11.s\n" "fmla z29.s, p3/M, z2.s, z13.s\n" - "ld1w { z11.s }, p2/Z, [x28, x9, LSL #2]\n" - "ldr x26, [x16, #0x50]\n" + "ld1w { z11.s }, p2/Z, [x27, x28, LSL #2]\n" + "ldr x25, [x15, #0x50]\n" "fmla z28.s, p3/M, z3.s, z14.s\n" "fmla z29.s, p3/M, z0.s, z16.s\n" - "ld1w { z13.s }, p2/Z, [x26, x9, LSL #2]\n" - "ldr x25, [x16, #0x58]\n" + "ld1w { z13.s }, p2/Z, [x25, x28, LSL #2]\n" + "ldr x24, [x15, #0x58]\n" "fmla z28.s, p3/M, z4.s, z15.s\n" "fmla z29.s, p3/M, z4.s, z11.s\n" - "ldr x20, [x16, #0x78]\n" - "ld1w { z14.s }, p2/Z, [x25, x9, LSL #2]\n" + "ldr x19, [x15, #0x78]\n" + "ld1w { z14.s }, p2/Z, [x24, x28, LSL #2]\n" "fmla z28.s, p3/M, z2.s, z16.s\n" "fmla z29.s, p3/M, z5.s, z12.s\n" - "ldr x23, [x16, #0x60]\n" - "ld1w { z15.s }, p2/Z, [x23, x9, LSL #2]\n" + "ldr x22, [x15, #0x60]\n" + "ld1w { z15.s }, p2/Z, [x22, x28, LSL #2]\n" "movprfx z30, z19\n fmla z30.s, p3/M, z2.s, z9.s\n" "movprfx z31, z19\n fmla z31.s, p3/M, z0.s, z9.s\n" - "ldr x28, [x16, #0x80]\n" - "ld1w { z12.s }, p2/Z, [x28, x9, LSL #2]\n" + "ldr x27, [x15, #0x80]\n" + "ld1w { z12.s }, p2/Z, [x27, x28, LSL #2]\n" "fmla z28.s, p3/M, z5.s, z13.s\n" "fmla z29.s, p3/M, z3.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x20, x9, LSL #2]\n" - "ldr x22, [x16, #0x68]\n" + "ld1w { z13.s }, p2/Z, [x19, x28, LSL #2]\n" + "ldr x21, [x15, #0x68]\n" "fmla z30.s, p3/M, z3.s, z14.s\n" "fmla z31.s, p3/M, z4.s, z13.s\n" - "ldr x27, [x16, #0x88]\n" - "ld1w { z11.s }, p2/Z, [x22, x9, LSL #2]\n" + "ldr x26, [x15, #0x88]\n" + "ld1w { z11.s }, p2/Z, [x21, x28, LSL #2]\n" "fmla z30.s, p3/M, z0.s, z15.s\n" "fmla z31.s, p3/M, z1.s, z12.s\n" - "ld1w { z14.s }, p2/Z, [x27, x9, LSL #2]\n" - "ldr x21, [x16, #0x70]\n" - "ldr x25, [x16, #0x98]\n" + "ld1w { z14.s }, p2/Z, [x26, x28, LSL #2]\n" + "ldr x20, [x15, #0x70]\n" + "ldr x24, [x15, #0x98]\n" "fmla z30.s, p3/M, z4.s, z11.s\n" "fmla z31.s, p3/M, z5.s, z14.s\n" - "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n" + "ld1w { z16.s }, p2/Z, [x20, x28, LSL #2]\n" "fmla z28.s, p3/M, z6.s, z15.s\n" - "ld1w { z11.s }, p2/Z, [x25, x9, LSL #2]\n" - "ldr x26, [x16, #0x90]\n" + "ld1w { z11.s }, p2/Z, [x24, x28, LSL #2]\n" + "ldr x25, [x15, #0x90]\n" "fmla z30.s, p3/M, z1.s, z16.s\n" - "ldr x22, [x16, #0xa8]\n" + "ldr x21, [x15, #0xa8]\n" "fmla z31.s, p3/M, z2.s, z11.s\n" "fmla z28.s, p3/M, z7.s, z16.s\n" - "ld1w { z15.s }, p2/Z, [x26, x9, LSL #2]\n" - "ld1w { z16.s }, p2/Z, [x22, x9, LSL #2]\n" - "ldr x23, [x16, #0xa0]\n" + "ld1w { z15.s }, p2/Z, [x25, x28, LSL #2]\n" + "ld1w { z16.s }, p2/Z, [x21, x28, LSL #2]\n" + "ldr x22, [x15, #0xa0]\n" "fmla z30.s, p3/M, z6.s, z15.s\n" "fmla z31.s, p3/M, z3.s, z16.s\n" - "ldr x21, [x16, #0xb0]\n" - "ld1w { z13.s }, p2/Z, [x23, x9, LSL #2]\n" + "ldr x20, [x15, #0xb0]\n" + "ld1w { z13.s }, p2/Z, [x22, x28, LSL #2]\n" "fmla z30.s, p3/M, z7.s, z13.s\n" "fmla z29.s, p3/M, z7.s, z12.s\n" - "ld1w { z14.s }, p2/Z, [x21, x9, LSL #2]\n" - "ldr x20, [x16, #0xb8]\n" + "ld1w { z14.s }, p2/Z, [x20, x28, LSL #2]\n" + "ldr x19, [x15, #0xb8]\n" "fmla z31.s, p3/M, z7.s, z14.s\n" "fmla z30.s, p3/M, z5.s, z16.s\n" - "ld1w { z15.s }, p2/Z, [x20, x9, LSL #2]\n" - "ldr x28, [x16, #0xc0]\n" + "ld1w { z15.s }, p2/Z, [x19, x28, LSL #2]\n" + "ldr x27, [x15, #0xc0]\n" "fmla z31.s, p3/M, z6.s, z15.s\n" "fmla z29.s, p3/M, z8.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x28, x9, LSL #2]\n" - "ldp x28, x27, [x16, #0x0]\n" + "ld1w { z11.s }, p2/Z, [x27, x28, LSL #2]\n" + "ldp x27, x26, [x15, #0x0]\n" "fmla z30.s, p3/M, z8.s, z15.s\n" "fmla z31.s, p3/M, z8.s, z11.s\n" - "ldp x26, x25, [x16, #0x10]\n" - "ld1w { z19.s }, p3/Z, [x15]\n" - "addvl x15, x15, #1\n" - "incw x9\n" - "ldp x23, x22, [x16, #0x20]\n" - "ld1w { z9.s }, p1/Z, [x28, x12, LSL #2]\n" - "incw x24\n" + "ldp x25, x24, [x15, #0x10]\n" + "ld1w { z19.s }, p3/Z, [x14]\n" + "addvl x14, x14, #1\n" + "incw x28\n" + "ldp x22, x21, [x15, #0x20]\n" + "ld1w { z9.s }, p1/Z, [x27, x11, LSL #2]\n" + "incw x23\n" "mov p0.b, p2.b\n" - "ldp x21, x20, [x16, #0x30]\n" - "ld1w { z10.s }, p1/Z, [x27, x12, LSL #2]\n" - "whilelt p2.s, x9, %x[n_channels]\n" + "ldp x20, x19, [x15, #0x30]\n" + "ld1w { z10.s }, p1/Z, [x26, x11, LSL #2]\n" + "whilelt p2.s, x28, %x[n_channels]\n" ".inst 0xc1b1ca5c // fclamp { z28.s-z31.s }, z18.s, z17.s\n" - "ld1w { z11.s }, p1/Z, [x26, x12, LSL #2]\n" - "st1w { z28.s }, p0, [x14, x24, LSL #2]\n" - "ld1w { z12.s }, p1/Z, [x25, x12, LSL #2]\n" - "st1w { z29.s }, p0, [x13, x24, LSL #2]\n" - "ld1w { z13.s }, p1/Z, [x23, x12, LSL #2]\n" - "st1w { z30.s }, p0, [x11, x24, LSL #2]\n" - "ld1w { z14.s }, p1/Z, [x22, x12, LSL #2]\n" - "st1w { z31.s }, p0, [x10, x24, LSL #2]\n" - "ld1w { z15.s }, p1/Z, [x21, x12, LSL #2]\n" - "ld1w { z16.s }, p1/Z, [x20, x12, LSL #2]\n" - "incw x12\n" - "cmp x12, %x[n_channels]\n" - ".inst 0xa040c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15]\n" - "addvl x15, x15, #4\n" - ".inst 0xa040c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n" - "addvl x15, x15, #4\n" - "ld1w { z8.s }, p3/Z, [x15]\n" - "addvl x15, x15, #1\n" + "ld1w { z11.s }, p1/Z, [x25, x11, LSL #2]\n" + "st1w { z28.s }, p0, [x13, x23, LSL #2]\n" + "ld1w { z12.s }, p1/Z, [x24, x11, LSL #2]\n" + "st1w { z29.s }, p0, [x12, x23, LSL #2]\n" + "ld1w { z13.s }, p1/Z, [x22, x11, LSL #2]\n" + "st1w { z30.s }, p0, [x10, x23, LSL #2]\n" + "ld1w { z14.s }, p1/Z, [x21, x11, LSL #2]\n" + "st1w { z31.s }, p0, [x9, x23, LSL #2]\n" + "ld1w { z15.s }, p1/Z, [x20, x11, LSL #2]\n" + "ld1w { z16.s }, p1/Z, [x19, x11, LSL #2]\n" + "incw x11\n" + "cmp x11, %x[n_channels]\n" + ".inst 0xa040c1c0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x14]\n" + "addvl x14, x14, #4\n" + ".inst 0xa040c1c4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x14]\n" + "addvl x14, x14, #4\n" + "ld1w { z8.s }, p3/Z, [x14]\n" + "addvl x14, x14, #1\n" "blt 1b\n" "2:" // Channel tail "movprfx z28, z19\n fmla z28.s, p3/M, z8.s, z9.s\n" "movprfx z29, z19\n fmla z29.s, p3/M, z6.s, z9.s\n" - "ldr x28, [x16, #0x40]\n" - "incw x24\n" + "ldr x27, [x15, #0x40]\n" + "incw x23\n" "fmla z28.s, p3/M, z0.s, z10.s\n" "fmla z29.s, p3/M, z1.s, z12.s\n" - "ldr x27, [x16, #0x48]\n" - "ld1w { z12.s }, p2/Z, [x27, x9, LSL #2]\n" + "ldr x26, [x15, #0x48]\n" + "ld1w { z12.s }, p2/Z, [x26, x28, LSL #2]\n" "fmla z28.s, p3/M, z1.s, z11.s\n" "fmla z29.s, p3/M, z2.s, z13.s\n" - "ld1w { z11.s }, p2/Z, [x28, x9, LSL #2]\n" - "ldr x26, [x16, #0x50]\n" + "ld1w { z11.s }, p2/Z, [x27, x28, LSL #2]\n" + "ldr x25, [x15, #0x50]\n" "fmla z28.s, p3/M, z3.s, z14.s\n" "fmla z29.s, p3/M, z0.s, z16.s\n" - "ld1w { z13.s }, p2/Z, [x26, x9, LSL #2]\n" - "ldr x25, [x16, #0x58]\n" + "ld1w { z13.s }, p2/Z, [x25, x28, LSL #2]\n" + "ldr x24, [x15, #0x58]\n" "fmla z28.s, p3/M, z4.s, z15.s\n" "fmla z29.s, p3/M, z4.s, z11.s\n" - "ldr x20, [x16, #0x78]\n" - "ld1w { z14.s }, p2/Z, [x25, x9, LSL #2]\n" + "ldr x19, [x15, #0x78]\n" + "ld1w { z14.s }, p2/Z, [x24, x28, LSL #2]\n" "fmla z28.s, p3/M, z2.s, z16.s\n" "fmla z29.s, p3/M, z5.s, z12.s\n" - "ldr x23, [x16, #0x60]\n" - "ld1w { z15.s }, p2/Z, [x23, x9, LSL #2]\n" + "ldr x22, [x15, #0x60]\n" + "ld1w { z15.s }, p2/Z, [x22, x28, LSL #2]\n" "movprfx z30, z19\n fmla z30.s, p3/M, z2.s, z9.s\n" "movprfx z31, z19\n fmla z31.s, p3/M, z0.s, z9.s\n" - "ldr x28, [x16, #0x80]\n" - "ld1w { z12.s }, p2/Z, [x28, x9, LSL #2]\n" + "ldr x27, [x15, #0x80]\n" + "ld1w { z12.s }, p2/Z, [x27, x28, LSL #2]\n" "fmla z28.s, p3/M, z5.s, z13.s\n" "fmla z29.s, p3/M, z3.s, z13.s\n" - "ld1w { z13.s }, p2/Z, [x20, x9, LSL #2]\n" - "ldr x22, [x16, #0x68]\n" + "ld1w { z13.s }, p2/Z, [x19, x28, LSL #2]\n" + "ldr x21, [x15, #0x68]\n" "fmla z30.s, p3/M, z3.s, z14.s\n" "fmla z31.s, p3/M, z4.s, z13.s\n" - "ldr x27, [x16, #0x88]\n" - "ld1w { z11.s }, p2/Z, [x22, x9, LSL #2]\n" + "ldr x26, [x15, #0x88]\n" + "ld1w { z11.s }, p2/Z, [x21, x28, LSL #2]\n" "fmla z30.s, p3/M, z0.s, z15.s\n" "fmla z31.s, p3/M, z1.s, z12.s\n" - "ld1w { z14.s }, p2/Z, [x27, x9, LSL #2]\n" - "ldr x21, [x16, #0x70]\n" - "ldr x25, [x16, #0x98]\n" + "ld1w { z14.s }, p2/Z, [x26, x28, LSL #2]\n" + "ldr x20, [x15, #0x70]\n" + "ldr x24, [x15, #0x98]\n" "fmla z30.s, p3/M, z4.s, z11.s\n" "fmla z31.s, p3/M, z5.s, z14.s\n" - "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n" + "ld1w { z16.s }, p2/Z, [x20, x28, LSL #2]\n" "fmla z28.s, p3/M, z6.s, z15.s\n" - "ld1w { z11.s }, p2/Z, [x25, x9, LSL #2]\n" - "ldr x26, [x16, #0x90]\n" + "ld1w { z11.s }, p2/Z, [x24, x28, LSL #2]\n" + "ldr x25, [x15, #0x90]\n" "fmla z30.s, p3/M, z1.s, z16.s\n" - "ldr x22, [x16, #0xa8]\n" + "ldr x21, [x15, #0xa8]\n" "fmla z31.s, p3/M, z2.s, z11.s\n" "fmla z28.s, p3/M, z7.s, z16.s\n" - "ld1w { z15.s }, p2/Z, [x26, x9, LSL #2]\n" - "ld1w { z16.s }, p2/Z, [x22, x9, LSL #2]\n" - "ldr x23, [x16, #0xa0]\n" + "ld1w { z15.s }, p2/Z, [x25, x28, LSL #2]\n" + "ld1w { z16.s }, p2/Z, [x21, x28, LSL #2]\n" + "ldr x22, [x15, #0xa0]\n" "fmla z30.s, p3/M, z6.s, z15.s\n" "fmla z31.s, p3/M, z3.s, z16.s\n" - "ldr x21, [x16, #0xb0]\n" - "ld1w { z13.s }, p2/Z, [x23, x9, LSL #2]\n" + "ldr x20, [x15, #0xb0]\n" + "ld1w { z13.s }, p2/Z, [x22, x28, LSL #2]\n" "fmla z30.s, p3/M, z7.s, z13.s\n" "fmla z29.s, p3/M, z7.s, z12.s\n" - "ld1w { z14.s }, p2/Z, [x21, x9, LSL #2]\n" - "ldr x20, [x16, #0xb8]\n" + "ld1w { z14.s }, p2/Z, [x20, x28, LSL #2]\n" + "ldr x19, [x15, #0xb8]\n" "fmla z31.s, p3/M, z7.s, z14.s\n" "fmla z30.s, p3/M, z5.s, z16.s\n" - "ld1w { z15.s }, p2/Z, [x20, x9, LSL #2]\n" - "ldr x28, [x16, #0xc0]\n" + "ld1w { z15.s }, p2/Z, [x19, x28, LSL #2]\n" + "ldr x27, [x15, #0xc0]\n" "fmla z31.s, p3/M, z6.s, z15.s\n" "fmla z29.s, p3/M, z8.s, z11.s\n" - "ld1w { z11.s }, p2/Z, [x28, x9, LSL #2]\n" + "ld1w { z11.s }, p2/Z, [x27, x28, LSL #2]\n" "fmla z30.s, p3/M, z8.s, z15.s\n" "fmla z31.s, p3/M, z8.s, z11.s\n" "mov p0.b, p2.b\n" ".inst 0xc1b1ca5c // fclamp { z28.s-z31.s }, z18.s, z17.s\n" - "st1w { z28.s }, p0, [x14, x24, LSL #2]\n" - "st1w { z29.s }, p0, [x13, x24, LSL #2]\n" - "st1w { z30.s }, p0, [x11, x24, LSL #2]\n" - "st1w { z31.s }, p0, [x10, x24, LSL #2]\n" + "st1w { z28.s }, p0, [x13, x23, LSL #2]\n" + "st1w { z29.s }, p0, [x12, x23, LSL #2]\n" + "st1w { z30.s }, p0, [x10, x23, LSL #2]\n" + "st1w { z31.s }, p0, [x9, x23, LSL #2]\n" ".inst 0xd503467f // SMSTOP\n" : : [n_channels] "r" ((unsigned long) n_channels), [offsetof_Args_inptrs] "I" (offsetof(Args, inptrs)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_args_params] "I" (offsetof(Args, params)), [params_struct] "r" (¶ms_struct) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } |