== Summary == - benchmarking coremark with VRP based extension elimination * extension elimination in some cases affecting other optimizations * With this improvements are marginal (details below)
== Plan == - study crc where extension elimination is resulting in bad code - Find a solution
==Details== If an assignment gimple statement has RHS expression value that can fit in LHS type, truncation is redundant. Zero/sign extensions are redundant in this case and rtl statement can be replaced as
from: (insn 12 11 0 (set (reg:SI 110 [ D.4128 ]) (zero_extend:SI (subreg:HI (reg:SI 117) 0))) c5.c:8 -1 (nil)) to: (insn 12 11 0 (set (subreg/s/u:HI (reg:SI 110 [ D.4128 ]) 0) (subreg:HI (reg:SI 117) 0)) c5.c:8 -1 (nil))
With this change, for the following case:
short unPack( unsigned char c ) { /* Only want lower four bit nibble */ c = c & (unsigned char)0x0F ;
if( c > 7 ) { /* Negative nibble */ return( ( short )( c - 16 ) ) ; } else { /* positive nibble */ return( ( short )c ) ; } }
asm without elimination unPack: @ args = 0, pretend = 0, frame = 0 @ frame_needed = 0, uses_anonymous_args = 0 @ link register save eliminated. and r0, r0, #15 cmp r0, #7 subhi r0, r0, #16 uxthhi r0, r0 sxth r0, r0 bx lr .size
asm with elimination unPack: @ args = 0, pretend = 0, frame = 0 @ frame_needed = 0, uses_anonymous_args = 0 @ link register save eliminated. and r0, r0, #15 cmp r0, #7 subhi r0, r0, #16 sxth r0, r0 bx lr
In some cases, changed rtl statement is not eliminated by later passes and is generated as a mov instruction. Worse, it also seems to affect the other optimization passes and resulting in worse code for crc. Not found the cause for it yet.