[Hexagon] Emit lookup tables in text section based on a flag

[sgundapa]

Summary

The flag "-hexagon-emit-lut-text" (defaulted to false) is added to decide
on where to keep the switch generated lookup table.

[kparzysz]

lib/Target/Hexagon/HexagonTargetObjectFile.cpp

142–145

This could be

if (const Function *Fn = getLutUsedFunction(GO))
  return selectSectionForLookupTable(G0, TM, Fn);

419

Maybe replace it with "if (!I) continue" to reduce indentation.

421

Are there legitimate cases where instruction does not have a parent? If not, you could remove this test (maybe add an assertion instead).

[sgundapa]

lib/Target/Hexagon/HexagonTargetObjectFile.cpp

419

I will work on indentation and push a new one

421

I am not sure about this which is why I am trying to be safe by returning a nullptr. There might be scenarios of dead basic blocks or functions . I probably need to run a big code base to see what might happen.

[sgundapa]

Updated the patch with the indentation edit

lib/Target/Hexagon/HexagonTargetObjectFile.h

public:
 
     bool isGlobalInSmallSection(const GlobalObject *GO,
                                 const TargetMachine &TM) const;
 
     bool isSmallDataEnabled() const;
 
     unsigned getSmallDataSize() const;
 
+    const Function *getLutUsedFunction(const GlobalObject *GO) const;
+
   private:
     MCSectionELF *SmallDataSection;
     MCSectionELF *SmallBSSSection;
 
     unsigned getSmallestAddressableSize(const Type *Ty, const GlobalValue *GV,
         const TargetMachine &TM) const;
 
     MCSection *selectSmallSectionForGlobal(const GlobalObject *GO,
                                            SectionKind Kind,
                                            const TargetMachine &TM) const;
+
+    MCSection *selectSectionForLookupTable(const GlobalObject *GO,
+                                           const TargetMachine &TM,
+                                           const Function *Fn) const;
   };
 
 } // namespace llvm
 
 #endif

lib/Target/Hexagon/HexagonTargetObjectFile.cpp

 static cl::opt<bool> StaticsInSData("hexagon-statics-in-small-data",
   cl::init(false), cl::Hidden, cl::ZeroOrMore,
   cl::desc("Allow static variables in .sdata"));
 
 static cl::opt<bool> TraceGVPlacement("trace-gv-placement",
   cl::Hidden, cl::init(false),
   cl::desc("Trace global value placement"));
 
+static cl::opt<bool>
+    EmitLutInText("hexagon-emit-lut-text", cl::Hidden, cl::init(false),
+                 cl::desc("Emit hexagon lookup tables in function section"));
+
 // TraceGVPlacement controls messages for all builds. For builds with assertions
 // (debug or release), messages are also controlled by the usual debug flags
 // (e.g. -debug and -debug-only=globallayout)
 #define TRACE_TO(s, X) s << X
 #ifdef NDEBUG
 #define TRACE(X)                                                               \
   do {                                                                         \
     if (TraceGVPlacement) {                                                    \
TRACE((GO->hasPrivateLinkage() ? "private_linkage " : "")
          << (GO->hasInternalLinkage() ? "internal " : "")
          << (GO->hasExternalLinkage() ? "external " : "")
          << (GO->hasCommonLinkage() ? "common_linkage " : "")
          << (GO->hasCommonLinkage() ? "common " : "" )
          << (Kind.isCommon() ? "kind_common " : "" )
          << (Kind.isBSS() ? "kind_bss " : "" )
          << (Kind.isBSSLocal() ? "kind_bss_local " : "" ));
 
+  // If the lookup table is used by more than one function, do not place
+  // it in text section.
+  if (EmitLutInText && GO->getName().startswith("switch.table")) {
+    const Function *Fn = getLutUsedFunction(GO);
+    if (Fn) {
+      return selectSectionForLookupTable(GO, TM, Fn);
+    }
+  }
+
   if (isGlobalInSmallSection(GO, TM))
     return selectSmallSectionForGlobal(GO, Kind, TM);
 
   if (Kind.isCommon()) {
     // This is purely for LTO+Linker Script because commons don't really have a
     // section. However, the BitcodeSectionWriter pass will query for the
     // sections of commons (and the linker expects us to know their section) so
     // we'll return one here.
if (Kind.isData()) {
     return getContext().getELFSection(Name.str(), ELF::SHT_PROGBITS,
                 ELF::SHF_WRITE | ELF::SHF_ALLOC | ELF::SHF_HEX_GPREL);
   }
 
   TRACE("default ELF section\n");
   // Otherwise, we work the same as ELF.
   return TargetLoweringObjectFileELF::SelectSectionForGlobal(GO, Kind, TM);
 }
+
+// Return the function that uses the lookup table. If there are more
+// than one live function that uses this look table, bail out and place
+// the lookup table in default section.
+const Function *
+HexagonTargetObjectFile::getLutUsedFunction(const GlobalObject *GO) const {
+  const Function *ReturnFn = nullptr;
+  for (auto U : GO->users()) {
+    // validate each instance of user to be a live function.
+    auto *I = dyn_cast<Instruction>(U);
+    if (I) {
+      auto *Bb = I->getParent();
+      if (Bb) {
+        auto *UserFn = Bb->getParent();
+        if (!ReturnFn) {
+          ReturnFn = UserFn;
+        } else if (ReturnFn != UserFn)
+          return nullptr;
+      }
+    }
+  }
+  return ReturnFn;
+}
+
+MCSection *HexagonTargetObjectFile::selectSectionForLookupTable(
+    const GlobalObject *GO, const TargetMachine &TM, const Function *Fn) const {
+
+  SectionKind Kind = SectionKind::getText();
+  // If the function has explicit section, place the lookup table in this
+  // explicit section.
+  if (Fn->hasSection())
+    return getExplicitSectionGlobal(Fn, Kind, TM);
+
+  const auto *FuncObj = dyn_cast<GlobalObject>(Fn);
+  return SelectSectionForGlobal(FuncObj, Kind, TM);
+}

test/CodeGen/Hexagon/switch-lut-explicit-section.ll

+;RUN: llc -O2 -hexagon-emit-lut-text=true < %s | FileCheck --check-prefix=FUNCTEXT %s
+;RUN: llc -O2 -hexagon-emit-lut-text=true -function-sections < %s | FileCheck --check-prefix=FUNCTEXT %s
+
+;This test checks the placement of lookup table in explicit section from the attribute set.
+target datalayout = "e-m:e-p:32:32:32-a:0-n16:32-i64:64:64-i32:32:32-i16:16:16-i1:8:8-f32:32:32-f64:64:64-v32:32:32-v64:64:64-v512:512:512-v1024:1024:1024-v2048:2048:2048"
+target triple = "hexagon-unknown--elf"
+
+;FUNCTEXT: .text
+;FUNCTEXT: .section{{.*}}tcm.hexagon,
+;FUNCTEXT-NOT: .section{{.*}}.rodata
+;FUNCTEXT-NOT: .text
+;FUNCTEXT: .Lswitch.table:
+;FUNCTEXT-NEXT: .word
+
[e-mail] = private unnamed_addr constant [9 x i32] [i32 9, i32 20, i32 14, i32 22, i32 12, i32 5, i32 98, i32 8, i32 11] #0
+
+; Function Attrs: norecurse nounwind readnone
+define i32 @foo(i32 %x) local_unnamed_addr #0 section "tcm.hexagon" {
+entry:
+  %0 = icmp ult i32 %x, 9
+  br i1 %0, label %switch.lookup, label %return
+
+switch.lookup:                                    ; preds = %entry
+  %switch.gep = getelementptr inbounds [9 x i32], [9 x i32]* @switch.table, i32 0, i32 %x
+  %switch.load = load i32, i32* %switch.gep, align 4
+  ret i32 %switch.load
+
+return:                                           ; preds = %entry
+  ret i32 19
+}
+
+attributes #0 = { norecurse nounwind readnone "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="hexagonv60" "target-features"="-hvx-double,-long-calls" "unsafe-fp-math"="false" "use-soft-float"="false" }

test/CodeGen/Hexagon/switch-lut-function-section.ll

+;RUN: llc -O2 -hexagon-emit-lut-text=true -function-sections < %s | FileCheck --check-prefix=FUNCTEXT %s
+
+;This test checks the placement of lookup table in function's text section.
+target datalayout = "e-m:e-p:32:32:32-a:0-n16:32-i64:64:64-i32:32:32-i16:16:16-i1:8:8-f32:32:32-f64:64:64-v32:32:32-v64:64:64-v512:512:512-v1024:1024:1024-v2048:2048:2048"
+target triple = "hexagon-unknown--elf"
+
+;FUNCTEXT: .text
+;FUNCTEXT: .section{{.*}}text.foo,
+;FUNCTEXT-NOT: .section{{.*}}.rodata
+;FUNCTEXT: .Lswitch.table:
+;FUNCTEXT-NEXT: .word
+
[e-mail] = private unnamed_addr constant [9 x i32] [i32 9, i32 20, i32 14, i32 22, i32 12, i32 5, i32 98, i32 8, i32 11] #0
+
+; Function Attrs: norecurse nounwind readnone
+define i32 @foo(i32 %x) local_unnamed_addr #0 {
+entry:
+  %0 = icmp ult i32 %x, 9
+  br i1 %0, label %switch.lookup, label %return
+
+switch.lookup:                                    ; preds = %entry
+  %switch.gep = getelementptr inbounds [9 x i32], [9 x i32]* @switch.table, i32 0, i32 %x
+  %switch.load = load i32, i32* %switch.gep, align 4
+  ret i32 %switch.load
+
+return:                                           ; preds = %entry
+  ret i32 19
+}
+
+attributes #0 = { norecurse nounwind readnone "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="hexagonv60" "target-features"="-hvx-double,-long-calls" "unsafe-fp-math"="false" "use-soft-float"="false" }

test/CodeGen/Hexagon/switch-lut-multiple-functions.ll

+;RUN: llc -O2 -hexagon-emit-lut-text=true < %s | FileCheck --check-prefix=TEXT %s
+;If the look up table is used by more than one function, we should ignore the
+;flag and place it the rodata.
+target datalayout = "e-m:e-p:32:32:32-a:0-n16:32-i64:64:64-i32:32:32-i16:16:16-i1:8:8-f32:32:32-f64:64:64-v32:32:32-v64:64:64-v512:512:512-v1024:1024:1024-v2048:2048:2048"
+target triple = "hexagon-unknown--elf"
+
+;TEXT: .text
+;TEXT: .section{{.*}}.rodata
+;TEXT: .Lswitch.table:
+;TEXT-NEXT: .word
[e-mail] = private unnamed_addr constant [9 x i32] [i32 9, i32 20, i32 14, i32 22, i32 12, i32 5, i32 98, i32 8, i32 11]
+
+; Function Attrs: norecurse nounwind readnone
+define i32 @foo(i32 %x) local_unnamed_addr #0 {
+entry:
+  %0 = icmp ult i32 %x, 9
+  br i1 %0, label %switch.lookup, label %return
+
+switch.lookup:                                    ; preds = %entry
+  %switch.gep = getelementptr inbounds [9 x i32], [9 x i32]* @switch.table, i32 0, i32 %x
+  %switch.load = load i32, i32* %switch.gep, align 4
+  ret i32 %switch.load
+
+return:                                           ; preds = %entry
+  ret i32 19
+}
+
+define i32 @goo(i32 %x) local_unnamed_addr #0 {
+entry:
+  %0 = icmp ult i32 %x, 9
+  br i1 %0, label %switch.lookup, label %return
+
+switch.lookup:                                    ; preds = %entry
+  %switch.gep = getelementptr inbounds [9 x i32], [9 x i32]* @switch.table, i32 0, i32 %x
+  %switch.load = load i32, i32* %switch.gep, align 4
+  ret i32 %switch.load
+
+return:                                           ; preds = %entry
+  ret i32 19
+}
+
+attributes #0 = { norecurse nounwind readnone "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="hexagonv60" "target-features"="-hvx-double,-long-calls" "unsafe-fp-math"="false" "use-soft-float"="false" }

test/CodeGen/Hexagon/switch-lut-text-section.ll

+;RUN: llc -O2 -hexagon-emit-lut-text=true < %s | FileCheck --check-prefix=TEXT %s
+;This test checks the placement of lookup table in text section.
+target datalayout = "e-m:e-p:32:32:32-a:0-n16:32-i64:64:64-i32:32:32-i16:16:16-i1:8:8-f32:32:32-f64:64:64-v32:32:32-v64:64:64-v512:512:512-v1024:1024:1024-v2048:2048:2048"
+target triple = "hexagon-unknown--elf"
+
+;TEXT: .text
+;TEXT-NOT: .section{{.*}}.rodata
+;TEXT: .Lswitch.table:
+;TEXT-NEXT: .word
[e-mail] = private unnamed_addr constant [9 x i32] [i32 9, i32 20, i32 14, i32 22, i32 12, i32 5, i32 98, i32 8, i32 11]
+
+; Function Attrs: norecurse nounwind readnone
+define i32 @foo(i32 %x) local_unnamed_addr #0 {
+entry:
+  %0 = icmp ult i32 %x, 9
+  br i1 %0, label %switch.lookup, label %return
+
+switch.lookup:                                    ; preds = %entry
+  %switch.gep = getelementptr inbounds [9 x i32], [9 x i32]* @switch.table, i32 0, i32 %x
+  %switch.load = load i32, i32* %switch.gep, align 4
+  ret i32 %switch.load
+
+return:                                           ; preds = %entry
+  ret i32 19
+}
+
+attributes #0 = { norecurse nounwind readnone "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="hexagonv60" "target-features"="-hvx-double,-long-calls" "unsafe-fp-math"="false" "use-soft-float"="false" }