[X86] Different scheduler information for vex/non-vex vector moves on Sandybridge

[RKSimon]

Bugzilla Link	37286
Resolution	FIXED
Resolved on	May 25, 2018 05:38
Version	trunk
OS	Windows NT
Blocks	#31672
CC	@legrosbuffle,@topperc,@gchatelet
Fixed by commit(s)	333271

Extended Description

Is this correct?

Port015:
MOVDQArr
MOVDQUrr

Port05
VMOVDQA(Y?)rr
VMOVDQU(Y?)rr

[llvmbot]

That's consistent with Agner Fog's results and with instlatx64 (http://users.atw.hu/instlatx64/GenuineIntel00206A7_SandyBridge3_InstLatX64.txt): They report MOVDQA xmm, xmm and VMOVDQA xmm, xmm at 0.33c (and the DQU version), so integer XMM register copies must run on all 3 vector ALU ports like Agner Fog says.

instlatx64 reports VMOVDQA ymm, ymm (and DQU) at 0.50c throughput, so we know it's 2 ports. I checked IACA 2.3 -arch SNB, and it says those instructions can only run on ports 0 and 5, so that confirms your numbers.

Agner Fog's tables don't have numbers for vmovdqa ymm, just xmm for Sandybridge. vmovdqa ymm,ymm is a rarely-used part of AVX1.

Interestingly, vmovaps ymm is port 5 only (like the xmm version). Presumably there's a bypass delay if you use VMOVDQA ymm,ymm between at least some kinds of FP instructions, or else they probably would have decoded vmovaps ymm,ymm to run on p0/p5.

IvB of course introduced mov-elimination, but SnB only eliminates xor-zeroing.

[RKSimon]

Ah, so it should really be this?

Port015:
MOVDQArr xmm
MOVDQUrr xmm
VMOVDQArr xmm
VMOVDQUrr xmm

Port05
VMOVDQArr ymm
VMOVDQUrr ymm

[llvmbot]

llvm-exegesis confirms this:

$ ./bin/llvm-exegesis -mode=uops -opcode-name=...

---

key:
opcode_name: MOVDQArr
mode: uops
config: ''
cpu_name: sandybridge
llvm_triple: x86_64-unknown-linux-gnu
num_repetitions: 10000
measurements:

• { key: '3', value: 0.3605, debug_string: SBPort0 }
• { key: '4', value: 0.3334, debug_string: SBPort1 }
• { key: '5', value: 0.0005, debug_string: SBPort4 }
• { key: '6', value: 0.3162, debug_string: SBPort5 }
• { key: '10', value: 0.0008, debug_string: SBPort23 }
  error: ''
  info: 'instruction has no tied variables picking Uses different from defs
  Snippet:
  MOVDQArr XMM9, XMM2
  '
  ...

---

key:
opcode_name: VMOVDQArr
mode: uops
config: ''
cpu_name: sandybridge
llvm_triple: x86_64-unknown-linux-gnu
num_repetitions: 10000
measurements:

• { key: '3', value: 0.3527, debug_string: SBPort0 }
• { key: '4', value: 0.3418, debug_string: SBPort1 }
• { key: '5', value: 0.0003, debug_string: SBPort4 }
• { key: '6', value: 0.3166, debug_string: SBPort5 }
• { key: '10', value: 0.001, debug_string: SBPort23 }
  error: ''
  info: 'instruction has no tied variables picking Uses different from defs
  Snippet:
  VMOVDQArr XMM4, XMM11
  '
  ...

---

key:
opcode_name: VMOVDQAYrr
mode: uops
config: ''
cpu_name: sandybridge
llvm_triple: x86_64-unknown-linux-gnu
num_repetitions: 10000
measurements:

• { key: '3', value: 0.5019, debug_string: SBPort0 }
• { key: '4', value: 0.0016, debug_string: SBPort1 }
• { key: '5', value: 0.0003, debug_string: SBPort4 }
• { key: '6', value: 0.5027, debug_string: SBPort5 }
• { key: '10', value: 0.0004, debug_string: SBPort23 }
  error: ''
  info: 'instruction has no tied variables picking Uses different from defs
  Snippet:
  VMOVDQAYrr YMM7, YMM0
  '
  ...

[RKSimon]

Thanks, fixed in rL333271