tree 5ef49d918bab3726375d49ff2353177075fde425
parent 76eaa97157f6204e04fa1d79529420760f7a408a
author Taylor Simpson <ltaylorsimpson@gmail.com> 1706783619 -0700
committer Brian Cain <bcain@quicinc.com> 1714951327 -0700

Hexagon (target/hexagon) Enable more short-circuit packets (scalar core)

Look for read-after-write instead of overlap of reads and writes

Here is an example with overalp but no read-after-write:
0x000200fc:  0x38103876	{	R0 = add(R0,R1); R6 = add(R6,R7) }

BEFORE:
 ---- 00000000000200fc
 mov_i32 loc2,$0x0
 mov_i32 loc2,r0
 add_i32 loc3,loc2,r1
 mov_i32 loc2,loc3
 mov_i32 loc4,$0x0
 mov_i32 loc4,r6
 add_i32 loc5,loc4,r7
 mov_i32 loc4,loc5
 mov_i32 r0,loc2
 mov_i32 r6,loc4

AFTER:
 ---- 00000000000200fc
 add_i32 loc2,r0,r1
 mov_i32 r0,loc2
 add_i32 loc3,r6,r7
 mov_i32 r6,loc3

We can also short-circuit packets with .new values by reading from the
real destination instead of the temporary.
0x00020100:  0x78005ff3	{	R19 = #0xff
0x00020104:  0x2002e204		if (cmp.eq(N19.new,R2)) jump:t PC+8 }

BEFORE:
 ---- 0000000000020100
 mov_i32 pc,$0x20108
 mov_i32 loc8,$0x0
 mov_i32 loc8,$0xff
 setcond_i32 loc10,loc8,r2,eq
 mov_i32 loc6,loc10
 mov_i32 r19,loc8
 add_i32 pkt_cnt,pkt_cnt,$0x2
 add_i32 insn_cnt,insn_cnt,$0x4
 brcond_i32 loc6,$0x0,eq,$L1
 goto_tb $0x0
 mov_i32 pc,$0x20108
 exit_tb $0x7fbb54000040
 set_label $L1
 goto_tb $0x1
 exit_tb $0x7fbb54000041
 set_label $L0
 exit_tb $0x7fbb54000043

AFTER:
 ---- 0000000000020100
 mov_i32 pc,$0x20108
 mov_i32 r19,$0xff
 setcond_i32 loc7,r19,r2,eq
 mov_i32 loc4,loc7
 add_i32 pkt_cnt,pkt_cnt,$0x2
 add_i32 insn_cnt,insn_cnt,$0x4
 brcond_i32 loc4,$0x0,eq,$L1
 goto_tb $0x0
 mov_i32 pc,$0x20108
 exit_tb $0x7f9764000040
 set_label $L1
 goto_tb $0x1
 exit_tb $0x7f9764000041
 set_label $L0
 exit_tb $0x7f9764000043

Signed-off-by: Taylor Simpson <ltaylorsimpson@gmail.com>
Reviewed-by: Brian Cain <bcain@quicinc.com>
Message-Id: <20240201103340.119081-3-ltaylorsimpson@gmail.com>
Signed-off-by: Brian Cain <bcain@quicinc.com>