diff options
Diffstat (limited to 'third_party/aom/build/make/thumb.pm')
-rw-r--r-- | third_party/aom/build/make/thumb.pm | 71 |
1 files changed, 71 insertions, 0 deletions
diff --git a/third_party/aom/build/make/thumb.pm b/third_party/aom/build/make/thumb.pm new file mode 100644 index 000000000..8248694e9 --- /dev/null +++ b/third_party/aom/build/make/thumb.pm @@ -0,0 +1,71 @@ +#!/usr/bin/env perl +## +## Copyright (c) 2016, Alliance for Open Media. All rights reserved +## +## This source code is subject to the terms of the BSD 2 Clause License and +## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +## was not distributed with this source code in the LICENSE file, you can +## obtain it at www.aomedia.org/license/software. If the Alliance for Open +## Media Patent License 1.0 was not distributed with this source code in the +## PATENTS file, you can obtain it at www.aomedia.org/license/patent. +## + +package thumb; + +sub FixThumbInstructions($$) +{ + my $short_branches = $_[1]; + my $branch_shift_offset = $short_branches ? 1 : 0; + + # Write additions with shifts, such as "add r10, r11, lsl #8", + # in three operand form, "add r10, r10, r11, lsl #8". + s/(add\s+)(r\d+),\s*(r\d+),\s*(lsl #\d+)/$1$2, $2, $3, $4/g; + + # Convert additions with a non-constant shift into a sequence + # with left shift, addition and a right shift (to restore the + # register to the original value). Currently the right shift + # isn't necessary in the code base since the values in these + # registers aren't used, but doing the shift for consistency. + # This converts instructions such as "add r12, r12, r5, lsl r4" + # into the sequence "lsl r5, r4", "add r12, r12, r5", "lsr r5, r4". + s/^(\s*)(add)(\s+)(r\d+),\s*(r\d+),\s*(r\d+),\s*lsl (r\d+)/$1lsl$3$6, $7\n$1$2$3$4, $5, $6\n$1lsr$3$6, $7/g; + + # Convert loads with right shifts in the indexing into a + # sequence of an add, load and sub. This converts + # "ldrb r4, [r9, lr, asr #1]" into "add r9, r9, lr, asr #1", + # "ldrb r9, [r9]", "sub r9, r9, lr, asr #1". + s/^(\s*)(ldrb)(\s+)(r\d+),\s*\[(\w+),\s*(\w+),\s*(asr #\d+)\]/$1add $3$5, $5, $6, $7\n$1$2$3$4, [$5]\n$1sub $3$5, $5, $6, $7/g; + + # Convert register indexing with writeback into a separate add + # instruction. This converts "ldrb r12, [r1, r2]!" into + # "ldrb r12, [r1, r2]", "add r1, r1, r2". + s/^(\s*)(ldrb)(\s+)(r\d+),\s*\[(\w+),\s*(\w+)\]!/$1$2$3$4, [$5, $6]\n$1add $3$5, $6/g; + + # Convert negative register indexing into separate sub/add instructions. + # This converts "ldrne r4, [src, -pstep, lsl #1]" into + # "subne src, src, pstep, lsl #1", "ldrne r4, [src]", + # "addne src, src, pstep, lsl #1". In a couple of cases where + # this is used, it's used for two subsequent load instructions, + # where a hand-written version of it could merge two subsequent + # add and sub instructions. + s/^(\s*)((ldr|str|pld)(ne)?)(\s+)(r\d+,\s*)?\[(\w+), -([^\]]+)\]/$1sub$4$5$7, $7, $8\n$1$2$5$6\[$7\]\n$1add$4$5$7, $7, $8/g; + + # Convert register post indexing to a separate add instruction. + # This converts "ldrneb r9, [r0], r2" into "ldrneb r9, [r0]", + # "addne r0, r0, r2". + s/^(\s*)((ldr|str)(ne)?[bhd]?)(\s+)(\w+),(\s*\w+,)?\s*\[(\w+)\],\s*(\w+)/$1$2$5$6,$7 [$8]\n$1add$4$5$8, $8, $9/g; + + # Convert a conditional addition to the pc register into a series of + # instructions. This converts "addlt pc, pc, r3, lsl #2" into + # "itttt lt", "movlt.n r12, pc", "addlt.w r12, #12", + # "addlt.w r12, r12, r3, lsl #2", "movlt.n pc, r12". + # This assumes that r12 is free at this point. + s/^(\s*)addlt(\s+)pc,\s*pc,\s*(\w+),\s*lsl\s*#(\d+)/$1itttt$2lt\n$1movlt.n$2r12, pc\n$1addlt.w$2r12, #12\n$1addlt.w$2r12, r12, $3, lsl #($4-$branch_shift_offset)\n$1movlt.n$2pc, r12/g; + + # Convert "mov pc, lr" into "bx lr", since the former only works + # for switching from arm to thumb (and only in armv7), but not + # from thumb to arm. + s/mov(\s*)pc\s*,\s*lr/bx$1lr/g; +} + +1; |