Misc LSRA throughput improvements

[kunalspathak]

While working on consecutive-registers, I realized few things that could help in the throughput:
1. We pass around RegisterType in various methods, but that parameter is only used for TARGET_ARM. So wrap the parameter in `ARM_ARG. Done separately in #86016.

2. We call updateAssignedInterval() frequently, but more than half of the time, we pass interval == nullptr which is essentially clearing the interval. Introduced clearAssignedInterval() for that purpose.

3. Use BitOperations::PopCount() in a method that is used for IsSingleRegister() check. Done separately as part of #85944.

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

While working on consecutive-registers, I realized few things that could help in the throughput:

1. We pass around RegisterType in various methods, but that parameter is only used for TARGET_ARM. So wrap the parameter in `ARM_ARG.
2. We call updateAssignedInterval() frequently, but more than half of the time, we pass interval == nullptr which is essentially clearing the interval. Introduced clearAssignedInterval() for that purpose.
3. Use BitOperations::PopCount() in a method that is used for IsSingleRegister() check.

Author:	kunalspathak
Assignees:	kunalspathak
Labels:	area-CodeGen-coreclr
Milestone:	-

[kunalspathak]

TP regressions is surprising. Probably need to compare the assembly of before vs. after to see which individual change might be causing it.

[kunalspathak]

TP regressions is surprising. Probably need to compare the assembly of before vs. after to see which individual change might be causing it.

Base: 538795279, Diff: 549102807, +1.9131%

?newRefPosition@LinearScan@@AEAAPEAVRefPosition@@PEAVInterval@@IW4RefType@@PEAUGenTree@@_KI@Z : 3693050  : +30.40%  : 22.83% : +0.6854%
?associateRefPosWithInterval@LinearScan@@AEAAXPEAVRefPosition@@@Z                             : 3035189  : +29.77%  : 18.76% : +0.5633%
?updateAssignedInterval@LinearScan@@AEAAXPEAVRegRecord@@PEAVInterval@@@Z                      : 2998888  : NA       : 18.54% : +0.5566%
?applySelection@RegisterSelection@LinearScan@@AEAA_NH_K@Z                                     : 2059105  : NA       : 12.73% : +0.3822%
??$select@$0A@@RegisterSelection@LinearScan@@QEAA_KPEAVInterval@@PEAVRefPosition@@@Z          : 1195021  : +4.50%   : 7.39%  : +0.2218%
?addRefsForPhysRegMask@LinearScan@@AEAAX_KIW4RefType@@_N@Z                                    : 230276   : +3.90%   : 1.42%  : +0.0427%
?buildInternalRegisterUses@LinearScan@@AEAAXXZ                                                : 28726    : +8.16%   : 0.18%  : +0.0053%
?updateAssignedInterval@LinearScan@@AEAAXPEAVRegRecord@@PEAVInterval@@W4var_types@@@Z         : -2913849 : -100.00% : 18.01% : -0.5408%

I didn't realize that we do not use intrinsics for popcount and which is why we are seeing lot of regressions. This is yet another example of why cross compilation comparison for TP might not be always accurate. cc: @jakobbotsch @BruceForstall

runtime/src/coreclr/jit/utils.cpp

Lines 2788 to 2808 in 8854509

	uint32_t BitOperations::PopCount(uint32_t value)
	{
	#if defined(_MSC_VER)
	// Inspired by the Stanford Bit Twiddling Hacks by Sean Eron Anderson:
	// http://graphics.stanford.edu/~seander/bithacks.html
	const uint32_t c1 = 0x55555555u;
	const uint32_t c2 = 0x33333333u;
	const uint32_t c3 = 0x0F0F0F0Fu;
	const uint32_t c4 = 0x01010101u;
	value -= (value >> 1) & c1;
	value = (value & c2) + ((value >> 2) & c2);
	value = (((value + (value >> 4)) & c3) * c4) >> 24;
	return value;
	#else
	int32_t result = __builtin_popcount(value);
	return static_cast<uint32_t>(result);
	#endif
	}

newRefposition:

buildInternalregisterusage

I will revert the popcount change.

[BruceForstall]

I didn't realize that we do not use intrinsics for popcount

Would be worthwhile profiling current function versus popcount -- probably we should switch to popcount.

[jakobbotsch]

I didn't realize that we do not use intrinsics for popcount and which is why we are seeing lot of regressions. This is yet another example of why cross compilation comparison for TP might not be always accurate. cc: @jakobbotsch @BruceForstall

It's a good point, but also a place where we should actively ensure we have parity regardless of the compiler we are using. It seems unfortunate that we are regressing either MSVC produced code or Clang produced code.

[kunalspathak]

Looking at the diffs for minopts benchmarks_run windows-x64 I see slight regression:

Base: 538795279, Diff: 538880143, +0.0158%

?updateAssignedInterval@LinearScan@@AEAAXPEAVRegRecord@@PEAVInterval@@@Z              : 2998888  : NA       : 50.71% : +0.5566%
?updateAssignedInterval@LinearScan@@AEAAXPEAVRegRecord@@PEAVInterval@@W4var_types@@@Z : -2913849 : -100.00% : 49.27% : -0.5408%

But looking at the code, we should still profitable, because we are eliminating a condition:

[kunalspathak]

Would be worthwhile profiling current function versus popcount -- probably we should switch to popcount.

Fixed.

[tannergooding]

It's a good point, but also a place where we should actively ensure we have parity regardless of the compiler we are using. It seems unfortunate that we are regressing either MSVC produced code or Clang produced code.

GCC/Clang generate the exact code that was codified for MSVC, they just do it implicitly via the builtin (and only switch to emitting actual popcnt if the ISA switch is passed in)

This was likely one of the many cases where we "execute more instructions" but the code was actually faster in practice.

[kunalspathak]

if the ISA switch is passed in

you mean when building clrjit using clang, right?

This was likely one of the many cases where we "execute more instructions" but the code was actually faster in practice.

Agree. btw, I do see that with VC++ popcnt removes lot of that code and generate the actual intrinsic which will be faster too. I will send a separate PR to see the effect of that alone rather than mixing up with some of the LSRA improvements I am doing here.

[jakobbotsch]

This was likely one of the many cases where we "execute more instructions" but the code was actually faster in practice.

I'm confused, are you saying the MSVC multiplication code was faster than using a single popcnt instruction?

[kunalspathak]

GCC/Clang generate the exact code that was codified for MSVC

hhm. https://godbolt.org/z/rEndsvhv6

[tannergooding]

I'm confused, are you saying the MSVC multiplication code was faster than using a single popcnt instruction?

@jakobbotsch: No, rather GCC/Clang don't emit popcnt here because our target machine is -msse2. In order for GCC/Clang to emit popcnt the target machine must be at least -msse42. For pre sse4.2, they emit the same logic as the multiplication code.

In order for us to emit popcnt here, we'd need to do it "opportunistically" via a cached CPUID check (much as we do for atomic operations on Arm64):

if (supportsPopcnt)
{
    return __popcnt(value);
}
else
{
    // Bit twiddling logic
}

[kunalspathak]

Latest diffs

[BruceForstall]

Odd it shows x64 as a slight regression.