[FuncSpec] Update MinFunctionSize logic

[hazzlim]

Always require functions to be larger than MinFunctionSize when SpecializeLiteralConstant is enabled, and increase MinFunctionSize to 500, to prevent excessive triggering of specialisations on small functions.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-function-specialization

Author: Hari Limaye (hazzlim)

Changes

Always require functions to be larger than MinFunctionSize when SpecializeLiteralConstant is enabled, and increase MinFunctionSize to 400, to prevent excessive triggering of specialisations on small functions.

---

Full diff: https://github.com/llvm/llvm-project/pull/112711.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/IPO/FunctionSpecialization.cpp (+10-5)

diff --git a/llvm/lib/Transforms/IPO/FunctionSpecialization.cpp b/llvm/lib/Transforms/IPO/FunctionSpecialization.cpp
index bd0a337e579e48..6d5c8b4229a417 100644
--- a/llvm/lib/Transforms/IPO/FunctionSpecialization.cpp
+++ b/llvm/lib/Transforms/IPO/FunctionSpecialization.cpp
@@ -57,9 +57,9 @@ static cl::opt<unsigned> MaxBlockPredecessors(
     "considered during the estimation of dead code"));
 
 static cl::opt<unsigned> MinFunctionSize(
-    "funcspec-min-function-size", cl::init(300), cl::Hidden, cl::desc(
-    "Don't specialize functions that have less than this number of "
-    "instructions"));
+    "funcspec-min-function-size", cl::init(500), cl::Hidden,
+    cl::desc("Don't specialize functions that have less than this number of "
+             "instructions"));
 
 static cl::opt<unsigned> MaxCodeSizeGrowth(
     "funcspec-max-codesize-growth", cl::init(3), cl::Hidden, cl::desc(
@@ -641,12 +641,17 @@ bool FunctionSpecializer::run() {
         Metrics.analyzeBasicBlock(&BB, GetTTI(F), EphValues);
     }
 
+    // When specializing literal constants is enabled, always require functions
+    // to be larger than MinFunctionSize, to prevent excessive specialization.
+    bool RequireMinSize =
+        !ForceSpecialization &&
+        (SpecializeLiteralConstant || !F.hasFnAttribute(Attribute::NoInline));
+
     // If the code metrics reveal that we shouldn't duplicate the function,
     // or if the code size implies that this function is easy to get inlined,
     // then we shouldn't specialize it.
     if (Metrics.notDuplicatable || !Metrics.NumInsts.isValid() ||
-        (!ForceSpecialization && !F.hasFnAttribute(Attribute::NoInline) &&
-         Metrics.NumInsts < MinFunctionSize))
+        (RequireMinSize && Metrics.NumInsts < MinFunctionSize))
       continue;
 
     // TODO: For now only consider recursive functions when running multiple

[kiranchandramohan]

MinFunctionSize to 400

Nit: It is 500 in the source.

[hazzlim]

Reduces the O3 (non-LTO) compile-time of CTMark, measured by the metric of instruction count, when SpecializeLiteralConstant is enabled.

Benchmark        |  Specs  | Change  |
-----------------+---------+---------+
ClamAV           | 2 -> 2  | -0.01%  |
7zip             | 0 -> 0  |  0.00%  |
kimwitu++        | 0 -> 0  | -0.01%  |
tramp3d-v4       | 0 -> 0  |  0.00%  |
sqlite3          | 0 -> 0  |  0.00%  |
mafft            | 0 -> 0  |  0.00%  |
SPASS            | 1 -> 0  | -0.43%  |
lencod           | 3 -> 0  | -0.39%  |
consumer-typeset | 0 -> 0  | -0.02%  |
Bullet           | 0 -> 0  |  0.00%  |
-----------------+---------+---------+
GEOMEAN          |         | -0.09%  |

It has no impact on Release-LTOg builds (either with or without SpecializeLiteralConstant), as no specializations are performed.

This change is intended to allow us to enable SpecializeLiteralConstant by default, without impacting compile-time.

[momchil-velikov]

LGTM

[momchil-velikov]

May add "const" for a little bit of extra self-documentation. Or not.

[hazzlim]

Good point - done :)

[labrinea]

The test that was added in https://reviews.llvm.org/D135862 will start failing once we enable specilaziation of literal constatnts. You can fix it now or later by adding -funcspec-for-literal-constant=false on the RUN line.

[hazzlim]

The test that was added in https://reviews.llvm.org/D135862 will start failing once we enable specilaziation of literal constatnts. You can fix it now or later by adding -funcspec-for-literal-constant=false on the RUN line.

Good point - I think that it makes sense to fix things / make it explicit in this change. Done :)

[labrinea]

Thanks for the contribution!