RFR: 8075443: AARCH64: Missed L2I optimizations in C2 (original) (raw)

Vladimir Kozlov vladimir.kozlov at oracle.com
Thu Mar 19 17:26:34 UTC 2015

Looks good to me.

Thanks, Vladimir

On 3/19/15 9:26 AM, Andrew Haley wrote:

I think I need another reviewer for this.

Thanks, Andrew.

On 03/18/2015 02:54 PM, Andrew Haley wrote: We use iRegIOrL2I erratically in aarch64.ad. We can use it whenever we have a 32-bit input operand, and doing so improves code quality in many cases. We should use iRegIOrL2I instead of iRegI whenever it can be used as an input operand and the patterns allow us to do so.

I also took the opportunity to do a little technical debt reduction. In a few places we were doing int-mode operations in 64-bit-wide instructions. This doesn't really matter, but it's a lot easier to reason about code generation if you have a rule that instructions must always use the correct mode. This is an example of a pattern which has both changes. @@ -8274,14 +8274,14 @@ %} instruct XorIregnotreg(iRegINoSp dst, - iRegI src1, iRegI src2, immIM1 m1, + iRegIorL2I src1, iRegIorL2I src2, immIM1 m1, rFlagsReg cr) %{ match(Set dst (XorI m1 (XorI src2 src1))); inscost(INSNCOST); - format %{ "eon dst,dst, dst,src1, $src2" %} - - insencode %{ - _ eon(asRegister($dst$$reg), + format %{ "eonw dst,dst, dst,src1, $src2" %} + + insencode %{ + _ eonw(asRegister($dst$$reg), asRegister($src1$$reg), asRegister($src2$$reg), Assembler::LSL, 0); http://cr.openjdk.java.net/~aph/8075443/ Andrew.

More information about the hotspot-compiler-dev mailing list