diff options
author | Maxim Yurchuk <maxim-yurchuk@ydb.tech> | 2024-10-18 20:31:38 +0300 |
---|---|---|
committer | GitHub <noreply@github.com> | 2024-10-18 20:31:38 +0300 |
commit | 2a74bac2d2d3bccb4e10120f1ead805640ec9dd0 (patch) | |
tree | 047e4818ced5aaf73f58517629e5260b5291f9f0 /contrib/libs/cxxsupp/builtins/hexagon/common_entry_exit_legacy.S | |
parent | 2d9656823e9521d8c29ea4c9a1d0eab78391abfc (diff) | |
parent | 3d834a1923bbf9403cd4a448e7f32b670aa4124f (diff) | |
download | ydb-2a74bac2d2d3bccb4e10120f1ead805640ec9dd0.tar.gz |
Merge pull request #10502 from ydb-platform/mergelibs-241016-1210
Library import 241016-1210
Diffstat (limited to 'contrib/libs/cxxsupp/builtins/hexagon/common_entry_exit_legacy.S')
-rw-r--r-- | contrib/libs/cxxsupp/builtins/hexagon/common_entry_exit_legacy.S | 156 |
1 files changed, 156 insertions, 0 deletions
diff --git a/contrib/libs/cxxsupp/builtins/hexagon/common_entry_exit_legacy.S b/contrib/libs/cxxsupp/builtins/hexagon/common_entry_exit_legacy.S new file mode 100644 index 0000000000..8a60445732 --- /dev/null +++ b/contrib/libs/cxxsupp/builtins/hexagon/common_entry_exit_legacy.S @@ -0,0 +1,156 @@ +//===----------------------Hexagon builtin routine ------------------------===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===----------------------------------------------------------------------===// + + +// Functions that implement common sequences in function prologues and epilogues +// used to save code size + + .macro FUNCTION_BEGIN name + .text + .globl \name + .type \name, @function + .falign +\name: + .endm + + .macro FUNCTION_END name + .size \name, . - \name + .endm + + .macro FALLTHROUGH_TAIL_CALL name0 name1 + .size \name0, . - \name0 + .globl \name1 + .type \name1, @function + .falign +\name1: + .endm + + + + +// Save r27:26 at fp+#-8, r25:24 at fp+#-16, r23:22 at fp+#-24, r21:20 at +// fp+#-32, r19:18 at fp+#-40, and r17:16 at fp+#-48. + + + + +// The compiler knows that the __save_* functions clobber LR. No other +// registers should be used without informing the compiler. + +// Since we can only issue one store per packet, we don't hurt performance by +// simply jumping to the right point in this sequence of stores. + +FUNCTION_BEGIN __save_r27_through_r16 + memd(fp+#-48) = r17:16 +FALLTHROUGH_TAIL_CALL __save_r27_through_r16 __save_r27_through_r18 + memd(fp+#-40) = r19:18 +FALLTHROUGH_TAIL_CALL __save_r27_through_r18 __save_r27_through_r20 + memd(fp+#-32) = r21:20 +FALLTHROUGH_TAIL_CALL __save_r27_through_r20 __save_r27_through_r22 + memd(fp+#-24) = r23:22 +FALLTHROUGH_TAIL_CALL __save_r27_through_r22 __save_r27_through_r24 + memd(fp+#-16) = r25:24 + { + memd(fp+#-8) = r27:26 + jumpr lr + } +FUNCTION_END __save_r27_through_r24 + + + + +// For each of the *_before_sibcall functions, jumpr lr is executed in parallel +// with deallocframe. That way, the return gets the old value of lr, which is +// where these functions need to return, and at the same time, lr gets the value +// it needs going into the sibcall. + +FUNCTION_BEGIN __restore_r27_through_r20_and_deallocframe_before_sibcall + { + r21:20 = memd(fp+#-32) + r23:22 = memd(fp+#-24) + } +FALLTHROUGH_TAIL_CALL __restore_r27_through_r20_and_deallocframe_before_sibcall __restore_r27_through_r24_and_deallocframe_before_sibcall + { + r25:24 = memd(fp+#-16) + jump __restore_r27_through_r26_and_deallocframe_before_sibcall + } +FUNCTION_END __restore_r27_through_r24_and_deallocframe_before_sibcall + + + + +FUNCTION_BEGIN __restore_r27_through_r16_and_deallocframe_before_sibcall + r17:16 = memd(fp+#-48) +FALLTHROUGH_TAIL_CALL __restore_r27_through_r16_and_deallocframe_before_sibcall __restore_r27_through_r18_and_deallocframe_before_sibcall + { + r19:18 = memd(fp+#-40) + r21:20 = memd(fp+#-32) + } +FALLTHROUGH_TAIL_CALL __restore_r27_through_r18_and_deallocframe_before_sibcall __restore_r27_through_r22_and_deallocframe_before_sibcall + { + r23:22 = memd(fp+#-24) + r25:24 = memd(fp+#-16) + } +FALLTHROUGH_TAIL_CALL __restore_r27_through_r22_and_deallocframe_before_sibcall __restore_r27_through_r26_and_deallocframe_before_sibcall + { + r27:26 = memd(fp+#-8) + deallocframe + jumpr lr + } +FUNCTION_END __restore_r27_through_r26_and_deallocframe_before_sibcall + + + + +// Here we use the extra load bandwidth to restore LR early, allowing the return +// to occur in parallel with the deallocframe. + +FUNCTION_BEGIN __restore_r27_through_r16_and_deallocframe + { + r17:16 = memd(fp+#-48) + r19:18 = memd(fp+#-40) + } +FALLTHROUGH_TAIL_CALL __restore_r27_through_r16_and_deallocframe __restore_r27_through_r20_and_deallocframe + { + r21:20 = memd(fp+#-32) + r23:22 = memd(fp+#-24) + } +FALLTHROUGH_TAIL_CALL __restore_r27_through_r20_and_deallocframe __restore_r27_through_r24_and_deallocframe + { + lr = memw(fp+#4) + r25:24 = memd(fp+#-16) + } + { + r27:26 = memd(fp+#-8) + deallocframe + jumpr lr + } +FUNCTION_END __restore_r27_through_r24_and_deallocframe + + + + +// Here the load bandwidth is maximized for all three functions. + +FUNCTION_BEGIN __restore_r27_through_r18_and_deallocframe + { + r19:18 = memd(fp+#-40) + r21:20 = memd(fp+#-32) + } +FALLTHROUGH_TAIL_CALL __restore_r27_through_r18_and_deallocframe __restore_r27_through_r22_and_deallocframe + { + r23:22 = memd(fp+#-24) + r25:24 = memd(fp+#-16) + } +FALLTHROUGH_TAIL_CALL __restore_r27_through_r22_and_deallocframe __restore_r27_through_r26_and_deallocframe + { + r27:26 = memd(fp+#-8) + deallocframe + } + jumpr lr +FUNCTION_END __restore_r27_through_r26_and_deallocframe |