michael@0: #!/usr/bin/perl
michael@0: ##
michael@0: ##  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
michael@0: ##
michael@0: ##  Use of this source code is governed by a BSD-style license
michael@0: ##  that can be found in the LICENSE file in the root of the source
michael@0: ##  tree. An additional intellectual property rights grant can be found
michael@0: ##  in the file PATENTS.  All contributing project authors may
michael@0: ##  be found in the AUTHORS file in the root of the source tree.
michael@0: ##
michael@0: 
michael@0: package thumb;
michael@0: 
michael@0: sub FixThumbInstructions($$)
michael@0: {
michael@0:     my $short_branches = $_[1];
michael@0:     my $branch_shift_offset = $short_branches ? 1 : 0;
michael@0: 
michael@0:     # Write additions with shifts, such as "add r10, r11, lsl #8",
michael@0:     # in three operand form, "add r10, r10, r11, lsl #8".
michael@0:     s/(add\s+)(r\d+),\s*(r\d+),\s*(lsl #\d+)/$1$2, $2, $3, $4/g;
michael@0: 
michael@0:     # Convert additions with a non-constant shift into a sequence
michael@0:     # with left shift, addition and a right shift (to restore the
michael@0:     # register to the original value). Currently the right shift
michael@0:     # isn't necessary in the code base since the values in these
michael@0:     # registers aren't used, but doing the shift for consitency.
michael@0:     # This converts instructions such as "add r12, r12, r5, lsl r4"
michael@0:     # into the sequence "lsl r5, r4", "add r12, r12, r5", "lsr r5, r4".
michael@0:     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;
michael@0: 
michael@0:     # Convert loads with right shifts in the indexing into a
michael@0:     # sequence of an add, load and sub. This converts
michael@0:     # "ldrb r4, [r9, lr, asr #1]" into "add r9, r9, lr, asr #1",
michael@0:     # "ldrb r9, [r9]", "sub r9, r9, lr, asr #1".
michael@0:     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;
michael@0: 
michael@0:     # Convert register indexing with writeback into a separate add
michael@0:     # instruction. This converts "ldrb r12, [r1, r2]!" into
michael@0:     # "ldrb r12, [r1, r2]", "add r1, r1, r2".
michael@0:     s/^(\s*)(ldrb)(\s+)(r\d+),\s*\[(\w+),\s*(\w+)\]!/$1$2$3$4, [$5, $6]\n$1add $3$5, $6/g;
michael@0: 
michael@0:     # Convert negative register indexing into separate sub/add instructions.
michael@0:     # This converts "ldrne r4, [src, -pstep, lsl #1]" into
michael@0:     # "subne src, src, pstep, lsl #1", "ldrne r4, [src]",
michael@0:     # "addne src, src, pstep, lsl #1". In a couple of cases where
michael@0:     # this is used, it's used for two subsequent load instructions,
michael@0:     # where a hand-written version of it could merge two subsequent
michael@0:     # add and sub instructions.
michael@0:     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;
michael@0: 
michael@0:     # Convert register post indexing to a separate add instruction.
michael@0:     # This converts "ldrneb r9, [r0], r2" into "ldrneb r9, [r0]",
michael@0:     # "add r0, r2".
michael@0:     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;
michael@0: 
michael@0:     # Convert a conditional addition to the pc register into a series of
michael@0:     # instructions. This converts "addlt pc, pc, r3, lsl #2" into
michael@0:     # "itttt lt", "movlt.n r12, pc", "addlt.w r12, #12",
michael@0:     # "addlt.w r12, r12, r3, lsl #2", "movlt.n pc, r12".
michael@0:     # This assumes that r12 is free at this point.
michael@0:     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;
michael@0: 
michael@0:     # Convert "mov pc, lr" into "bx lr", since the former only works
michael@0:     # for switching from arm to thumb (and only in armv7), but not
michael@0:     # from thumb to arm.
michael@0:     s/mov(\s*)pc\s*,\s*lr/bx$1lr/g;
michael@0: }
michael@0: 
michael@0: 1;