[GlobalMerge] Add MinSize feature to the GlobalMerge Pass. (#93686)

michaelmaitland · web-flow · commit 0f669154e114 · 2024-06-03T09:10:56.000-04:00
We add a feature that prevents the GlobalMerge pass from considering
data smaller than a minimum size in bytes for merging.

The MinSize is set in 3 ways:
1. If global-merge-min-data-size is explicitly set, then it uses that
value.
2. If SmallDataLimit is set and non-zero, then SmallDataLimit + 1 is
used.
3. Otherwise, 0 is used, which means all sizes are considered for
merging.

We found that this feature allowed us to see the benefit of the
GlobalMerge pass while eliminating some merging that was not beneficial.
This feature allowed us to enable the GlobalMerge pass on RISC-V in our
downstream by default because it led to improvements on multiple
benchmark suites.

I plan to post a separate patch to propose enabling this by default on
RISC-V. But I do not want that discussion to be part of the discussion
of adding this feature, so I am keeping the patches separate.
diff --git a/llvm/include/llvm/CodeGen/GlobalMerge.h b/llvm/include/llvm/CodeGen/GlobalMerge.h
@@ -21,6 +21,8 @@ struct GlobalMergeOptions {
   // functions), see the code that passes in the offset in the ARM backend
   // for more information.
   unsigned MaxOffset = 0;
+  // The minimum size in bytes of each global that should considered in merging.
+  unsigned MinSize = 0;
   bool GroupByUse = true;
   bool IgnoreSingleUse = true;
   bool MergeConst = false;
diff --git a/llvm/lib/CodeGen/GlobalMerge.cpp b/llvm/lib/CodeGen/GlobalMerge.cpp
@@ -134,6 +134,12 @@ static cl::opt<cl::boolOrDefault>
 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
      cl::desc("Enable global merge pass on external linkage"));
 
+static cl::opt<unsigned>
+    GlobalMergeMinDataSize("global-merge-min-data-size",
+                           cl::desc("The minimum size in bytes of each global "
+                                    "that should considered in merging."),
+                           cl::init(0), cl::Hidden);
+
 STATISTIC(NumMerged, "Number of globals merged");
 
 namespace {
@@ -198,6 +204,19 @@ class GlobalMerge : public FunctionPass {
   }
 
   bool doInitialization(Module &M) override {
+    auto GetSmallDataLimit = [](Module &M) -> std::optional<uint64_t> {
+      Metadata *SDL = M.getModuleFlag("SmallDataLimit");
+      if (!SDL)
+        return std::nullopt;
+      return mdconst::extract<ConstantInt>(SDL)->getZExtValue();
+    };
+    if (GlobalMergeMinDataSize.getNumOccurrences())
+      Opt.MinSize = GlobalMergeMinDataSize;
+    else if (auto SDL = GetSmallDataLimit(M); SDL && *SDL > 0)
+      Opt.MinSize = *SDL + 1;
+    else
+      Opt.MinSize = 0;
+
     GlobalMergeImpl P(TM, Opt);
     return P.run(M);
   }
@@ -670,7 +689,8 @@ bool GlobalMergeImpl::run(Module &M) {
       continue;
 
     Type *Ty = GV.getValueType();
-    if (DL.getTypeAllocSize(Ty) < Opt.MaxOffset) {
+    TypeSize AllocSize = DL.getTypeAllocSize(Ty);
+    if (AllocSize < Opt.MaxOffset && AllocSize >= Opt.MinSize) {
       if (TM &&
           TargetLoweringObjectFile::getKindForGlobal(&GV, *TM).isBSS())
         BSSGlobals[{AddressSpace, Section}].push_back(&GV);
diff --git a/llvm/test/CodeGen/RISCV/global-merge-minsize-smalldata-nonzero.ll b/llvm/test/CodeGen/RISCV/global-merge-minsize-smalldata-nonzero.ll
@@ -0,0 +1,48 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv64 -riscv-enable-global-merge -verify-machineinstrs < %s \
+; RUN:   | FileCheck %s -check-prefix=SMALL-DATA
+; RUN: llc -mtriple=riscv64 -riscv-enable-global-merge -global-merge-min-data-size=0 \
+; RUN:    -verify-machineinstrs < %s | FileCheck %s -check-prefix=MINSIZE
+
+@ig1 = internal global i32 0, align 4
+@ig2 = internal global i32 0, align 4
+
+@eg1 = dso_local global i32 0, align 4
+@eg2 = dso_local global i32 0, align 4
+
+; This test shows that GlobalDataMinSize is set to SmallDataLimit + 1 when
+; SmallDataLimit module flag is set as non-zero, and that global-merge-min-data-size
+; overrides the small data limit.
+
+define void @f1(i32 %a) nounwind {
+; SMALL-DATA-LABEL: f1:
+; SMALL-DATA:       # %bb.0:
+; SMALL-DATA-NEXT:    lui a1, %hi(ig1)
+; SMALL-DATA-NEXT:    sw a0, %lo(ig1)(a1)
+; SMALL-DATA-NEXT:    lui a1, %hi(ig2)
+; SMALL-DATA-NEXT:    sw a0, %lo(ig2)(a1)
+; SMALL-DATA-NEXT:    lui a1, %hi(eg1)
+; SMALL-DATA-NEXT:    sw a0, %lo(eg1)(a1)
+; SMALL-DATA-NEXT:    lui a1, %hi(eg2)
+; SMALL-DATA-NEXT:    sw a0, %lo(eg2)(a1)
+; SMALL-DATA-NEXT:    ret
+;
+; MINSIZE-LABEL: f1:
+; MINSIZE:       # %bb.0:
+; MINSIZE-NEXT:    lui a1, %hi(.L_MergedGlobals)
+; MINSIZE-NEXT:    sw a0, %lo(.L_MergedGlobals)(a1)
+; MINSIZE-NEXT:    addi a1, a1, %lo(.L_MergedGlobals)
+; MINSIZE-NEXT:    sw a0, 4(a1)
+; MINSIZE-NEXT:    sw a0, 8(a1)
+; MINSIZE-NEXT:    sw a0, 12(a1)
+; MINSIZE-NEXT:    ret
+  store i32 %a, ptr @ig1, align 4
+  store i32 %a, ptr @ig2, align 4
+  store i32 %a, ptr @eg1, align 4
+  store i32 %a, ptr @eg2, align 4
+  ret void
+}
+
+
+!llvm.module.flags = !{!0}
+!0 = !{i32 8, !"SmallDataLimit", i32 8}
diff --git a/llvm/test/CodeGen/RISCV/global-merge-minsize-smalldata-zero.ll b/llvm/test/CodeGen/RISCV/global-merge-minsize-smalldata-zero.ll
@@ -0,0 +1,49 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -riscv-enable-global-merge -verify-machineinstrs < %s \
+; RUN:   | FileCheck %s -check-prefix=SMALL-DATA
+; RUN: llc -mtriple=riscv64 -riscv-enable-global-merge -global-merge-min-data-size=5 \
+; RUN:    -verify-machineinstrs < %s | FileCheck %s -check-prefix=MINSIZE
+
+@ig1 = internal global i32 0, align 4
+@ig2 = internal global i32 0, align 4
+
+@eg1 = dso_local global i32 0, align 4
+@eg2 = dso_local global i32 0, align 4
+
+
+; This test shows that GlobalDataMinSize is set to 0 when SmallDataLimit module
+; flag is set to zero, and that the global-merge-min-data-size option overrides
+; the small data limit.
+
+define void @f1(i32 %a) nounwind {
+; SMALL-DATA-LABEL: f1:
+; SMALL-DATA:       # %bb.0:
+; SMALL-DATA-NEXT:    lui a1, %hi(.L_MergedGlobals)
+; SMALL-DATA-NEXT:    sw a0, %lo(.L_MergedGlobals)(a1)
+; SMALL-DATA-NEXT:    addi a1, a1, %lo(.L_MergedGlobals)
+; SMALL-DATA-NEXT:    sw a0, 4(a1)
+; SMALL-DATA-NEXT:    sw a0, 8(a1)
+; SMALL-DATA-NEXT:    sw a0, 12(a1)
+; SMALL-DATA-NEXT:    ret
+;
+; MINSIZE-LABEL: f1:
+; MINSIZE:       # %bb.0:
+; MINSIZE-NEXT:    lui a1, %hi(ig1)
+; MINSIZE-NEXT:    sw a0, %lo(ig1)(a1)
+; MINSIZE-NEXT:    lui a1, %hi(ig2)
+; MINSIZE-NEXT:    sw a0, %lo(ig2)(a1)
+; MINSIZE-NEXT:    lui a1, %hi(eg1)
+; MINSIZE-NEXT:    sw a0, %lo(eg1)(a1)
+; MINSIZE-NEXT:    lui a1, %hi(eg2)
+; MINSIZE-NEXT:    sw a0, %lo(eg2)(a1)
+; MINSIZE-NEXT:    ret
+  store i32 %a, ptr @ig1, align 4
+  store i32 %a, ptr @ig2, align 4
+  store i32 %a, ptr @eg1, align 4
+  store i32 %a, ptr @eg2, align 4
+  ret void
+}
+
+
+!llvm.module.flags = !{!0}
+!0 = !{i32 8, !"SmallDataLimit", i32 0}
diff --git a/llvm/test/CodeGen/RISCV/global-merge-minsize.ll b/llvm/test/CodeGen/RISCV/global-merge-minsize.ll
@@ -0,0 +1,40 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -riscv-enable-global-merge -verify-machineinstrs < %s \
+; RUN:   | FileCheck %s -check-prefix=RV32
+; RUN: llc -mtriple=riscv32 -riscv-enable-global-merge -global-merge-min-data-size=5 \
+; RUN:   -verify-machineinstrs < %s | FileCheck %s -check-prefix=RV32-MINSIZE
+
+@ig1 = internal global i32 0, align 4
+@ig2 = internal global i32 0, align 4
+
+@eg1 = dso_local global i32 0, align 4
+@eg2 = dso_local global i32 0, align 4
+
+define void @f1(i32 %a) nounwind {
+; RV32-LABEL: f1:
+; RV32:       # %bb.0:
+; RV32-NEXT:    lui a1, %hi(.L_MergedGlobals)
+; RV32-NEXT:    sw a0, %lo(.L_MergedGlobals)(a1)
+; RV32-NEXT:    addi a1, a1, %lo(.L_MergedGlobals)
+; RV32-NEXT:    sw a0, 4(a1)
+; RV32-NEXT:    sw a0, 8(a1)
+; RV32-NEXT:    sw a0, 12(a1)
+; RV32-NEXT:    ret
+;
+; RV32-MINSIZE-LABEL: f1:
+; RV32-MINSIZE:       # %bb.0:
+; RV32-MINSIZE-NEXT:    lui a1, %hi(ig1)
+; RV32-MINSIZE-NEXT:    sw a0, %lo(ig1)(a1)
+; RV32-MINSIZE-NEXT:    lui a1, %hi(ig2)
+; RV32-MINSIZE-NEXT:    sw a0, %lo(ig2)(a1)
+; RV32-MINSIZE-NEXT:    lui a1, %hi(eg1)
+; RV32-MINSIZE-NEXT:    sw a0, %lo(eg1)(a1)
+; RV32-MINSIZE-NEXT:    lui a1, %hi(eg2)
+; RV32-MINSIZE-NEXT:    sw a0, %lo(eg2)(a1)
+; RV32-MINSIZE-NEXT:    ret
+  store i32 %a, ptr @ig1, align 4
+  store i32 %a, ptr @ig2, align 4
+  store i32 %a, ptr @eg1, align 4
+  store i32 %a, ptr @eg2, align 4
+  ret void
+}
