Request for reviews (XL): 7116452: Add support for AVX instructions
tom.rodriguez at oracle.com
Wed Dec 14 13:08:27 PST 2011
>> Reading through the x86 ad files really put me in the mood to eliminate the spurious formatting and naming differences between them. Large portions of them are (or could be) exactly the same.
> I have the same feeling. I want to create third x86.ad file which has common instructions definitions. Currently Xmm registers naming is different (regXD vs regD) but we can fix it.
I actually wasted some time yesterday fixing some naming and formatting. There are a lot of spurious renames like RegMem to reg_mem plus all the formatting stuff. Switching to MacroAssembler in more places would remove a lot of differences too. There's also the eReg vs. rReg naming but I think we should just switch to rReg and there's the upper vs. lower case format output and I think we should switch to lower. Plus things like movq vs movl. It's ugly but we should minimize as much as we can.
>> On Dec 8, 2011, at 3:53 PM, Vladimir Kozlov wrote:
>>> 7116452: Add support for AVX instructions
>>> Initial changes were submitted by Intel. I refactored it to simplify prefix usage in instructions codding (added simd_prefix methods) and VEX encoding was fixed to generate 2bytes prefix when possible. Changes in .ad files were not complete (especially in 32-bit .ad) and were not aggressive as I want. I changed more mach nodes encoding to use macroassembler instructions. Added missing decoding parts in Assembler::locate_operand() and NativeMovRegMem::instruction_start().
>>> Note: no new AVX instructions were added in these changes. And no 3 operands format was added to MacroAssembler. It will be other changes. Destination operand is used as second source in current implementation where applicable.
>>> Float compare implementation in x86_32.ad was replaced with implementation from x86_64.ad. It uses less branches and does not destroy EAX register. Note: ucomiss instruction produces the same result as comiss since we masking numeric exceptions. Also ucomiss could be a little faster since it does not need to check control word for QNaN values.
>>> Vector instructions with VEX prefix use unaligned load for memory operands where with old REX prefix it require 16 bytes alignment. Instructions version with memory operand were added for that but they should be used only with VEX prefix, assert was added. ANDPD and XORPD with memory operand were used before with 16 bytes aligned memory (we have special code to do it). I added assert to check address alignment for these instructions.
>>> As part of these changes REX.W prefix was removed from instructions where it was not needed: MOVDQA, MOVDQU, PCMPESTRI, PSRLQ, PSRLDQ, PTEST.
>>> Tested with UseAVX=1|0, UseSSE=4|2|1|0, CTW, VM regression tests, nsk.
More information about the hotspot-compiler-dev