[llvm] Update default cutoff threshold for machine function splitter.

[snehasish]

Summary

Based on internal testing at Google we found that setting the profile
summary cutoff threshold to 999950 yields the best results in terms of
itlb and icache metrics (as observed on Intel CPUs).

*default* = Split out code if no profile count available for block
*size-%* = The fraction of bytes split out of .text and .text.hot
*itlb* = Misses per kilo instructions (MPKI) for itlb
*icache* = Misses per kilo instructions (MPKI) for L1 icache

Search1

cutoff	size-%	itlb	icache
default	42.5861	0.0822151	2.46363
999999	44.9350	0.0767194	2.44416
999950	50.0660	0.075744	2.4091
999500	56.9158	0.082564	2.4188
995000	63.8625	0.0814927	2.42832
990000	71.7314	0.106906	2.57785

Search2

cutoff	size-%	itlb	icache
default	2.8845	0.626712	4.73245
999999	3.3291	0.602309	4.70045
999950	3.8577	0.587842	4.71632
999500	4.4170	0.63577	4.68351
995000	5.1020	0.657969	4.82272
990000	5.7153	0.719122	5.39496

[echristo]

Arguably since this is going to be very cpu dependent perhaps it should be
part of TTI?

-eric

[snehasish]

Add a FIXME to move defaults to TTI.

[snehasish]

This pass is only enabled on X86 platforms (D87047 for clang options and platform check). Longer term it does make sense to move it to TTI so I've added a FIXME to get this change off the critical path and I'll follow up with a refactoring change.

[MaskRay]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

42

Nit: TTI -> TargetTransformInfo

[snehasish]

Spell out TTI as TargetTransformInfo, rebase.

[snehasish]

PTAL, thanks!

[tmsriram]

LGTM. The concerns raised have been addressed and it will be moved to TTI while porting to other platforms.

[echristo]

Sorry, didn't see this last time. Couple of inline comments, mostly nits.

-eric

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

43

Can you comment the value? What it's based off of, what it means, etc.

[snehasish]

Add comments, rebase.

[snehasish]

Thanks for the review.

[echristo]

Awesome. Thanks for the update :)

-eric

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

 #include "llvm/CodeGen/Passes.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/Module.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Support/CommandLine.h"
 
 using namespace llvm;
 
+// FIXME: This cutoff value is CPU dependent and should be moved to
+// TargetTransformInfo once we consider enabling this on other platforms.
+// The value is expressed as a ProfileSummaryInfo integer percentile cutoff.
+// Defaults to 999950, i.e. all blocks colder than 99.995 percentile are split.
+// The default was empirically determined to be optimal when considering cutoff
+// values between 99%-ile to 100%-ile with respect to iTLB and icache metrics on
+// Intel CPUs.
 static cl::opt<unsigned>
     PercentileCutoff("mfs-psi-cutoff",
                      cl::desc("Percentile profile summary cutoff used to "
                               "determine cold blocks. Unused if set to zero."),
-                     cl::init(0), cl::Hidden);
+                     cl::init(999950), cl::Hidden);
 
 static cl::opt<unsigned> ColdCountThreshold(
     "mfs-count-threshold",
     cl::desc(
         "Minimum number of times a block must be executed to be retained."),
     cl::init(1), cl::Hidden);
 
 namespace {

llvm/test/CodeGen/X86/machine-function-splitter.ll

 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions | FileCheck %s -check-prefix=MFS-DEFAULTS
-; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-count-threshold=2000 | FileCheck %s --dump-input=always -check-prefix=MFS-OPTS1
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-psi-cutoff=0 -mfs-count-threshold=2000 | FileCheck %s --dump-input=always -check-prefix=MFS-OPTS1
 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-psi-cutoff=950000 | FileCheck %s -check-prefix=MFS-OPTS2
 
 define void @foo1(i1 zeroext %0) nounwind !prof !14 !section_prefix !15 {
 ;; Check that cold block is moved to .text.split.
 ; MFS-DEFAULTS-LABEL: foo1
 ; MFS-DEFAULTS:       .section        .text.split.foo1
 ; MFS-DEFAULTS-NEXT:  foo1.cold:
 ; MFS-DEFAULTS-NOT:   callq   bar