Optimize stack usage

Author: LucasSte

llvm/lib/Target/SBF/SBFFrameLowering.cpp

@@ -37,7 +37,7 @@ void SBFFrameLowering::emitPrologue(MachineFunction &MF,
 
     if (Subtarget.isDynamicFramesV1())
       NumBytes = -NumBytes;
-    else if (NumBytes <= FrameSize)
+    else if (NumBytes <= FrameSize && !Subtarget.getOptimizeStackSpace())
       // In V3, we don't bump if the number of bytes is less than the default
       // frame size.
       return;

llvm/lib/Target/SBF/SBFRegisterInfo.cpp

@@ -179,6 +179,9 @@ int SBFRegisterInfo::resolveInternalFrameIndex(
     if (SubTarget.isDynamicFramesV1())
       return Offset + static_cast<int>(StackSize);
 
+    if (SubTarget.getOptimizeStackSpace())
+      return -(Offset + static_cast<int>(StackSize));
+
     return -(Offset + std::max(static_cast<int>(StackSize), static_cast<int>(FrameLength)));
   }
 

llvm/lib/Target/SBF/SBFSubtarget.h

@@ -95,6 +95,9 @@ class SBFSubtarget : public SBFGenSubtargetInfo {
   // Whether we are dealing with dynamic stack frames in SBPFv3
   bool HasDynamicFramesV3;
 
+  // Whether we should bump down the frame pointer in SBPFv3
+  bool OptimizeStackSpace;
+
   std::unique_ptr<CallLowering> CallLoweringInfo;
   std::unique_ptr<InstructionSelector> InstSelector;
   std::unique_ptr<LegalizerInfo> Legalizer;
@@ -130,6 +133,7 @@ class SBFSubtarget : public SBFGenSubtargetInfo {
   bool getNewMemEncoding() const { return NewMemEncoding; }
   bool getHasStaticSyscalls() const { return HasStaticSyscalls; }
   bool getHasJmp32() const { return HasJmp32; }
+  bool getOptimizeStackSpace() const { return OptimizeStackSpace; }
   const SBFInstrInfo *getInstrInfo() const override { return &InstrInfo; }
   const SBFFrameLowering *getFrameLowering() const override {
     return &FrameLowering;

llvm/lib/Target/SBF/SBFTargetFeatures.td

@@ -58,6 +58,9 @@ def FeatureJmp32 : SubtargetFeature<"jmp-32", "HasJmp32", "true",
 def FeatureDynamicFramesV3 : SubtargetFeature<"dynamic-frames-v3", "HasDynamicFramesV3", "true",
                                             "Enable dynamic frames in SBPFv3">;
 
+def OptimizeStackSpace : SubtargetFeature<"optimize-stack-space", "OptimizeStackSpace", "true",
+                                                                       "Optimize stack space usage in SBPFv3">;
+
 class Proc<string Name, list<SubtargetFeature> Features>
  : Processor<Name, NoItineraries, Features>;
 

llvm/test/CodeGen/SBF/dynamic_stack_frame_add_and_sub.ll

@@ -1,5 +1,6 @@
 ; RUN: llc < %s -march=sbf -mattr=+dynamic-frames | FileCheck %s
 ; RUN: llc -march=sbf -mattr=+dynamic-frames-v3,+alu32 < %s | FileCheck --check-prefix=CHECK-V3 %s
+; RUN: llc -march=sbf -mattr=+dynamic-frames-v3,+optimize-stack-space < %s | FileCheck --check-prefix=CHECK-OPT %s
 ;
 ; Source:
 ; int test_func(int * vec, int idx) {
@@ -15,6 +16,7 @@ define i32 @test_func(ptr noundef %vec, i32 noundef %idx) #0 {
 ; CHECK-LABEL: test_func:
 ; CHECK: add64 r10, -128
 ; CHECK-V3-NOT: add64 r10, 128
+; CHECK-OPT: add64 r10, -3968
 entry:
   %vec.addr = alloca ptr, align 8
   %idx.addr = alloca i512, align 4
@@ -36,6 +38,7 @@ declare i64 @read_ptr(ptr %a);
 define i64 @test_func_4096(i64 %idx) {
 ; CHECK-LABEL: test_func_4096
 ; CHECK-V3-NOT: add64 r10, 4096
+; CHECK-OPT-NOT: add64 r10, 4096
 entry:
     %large_var = alloca [4096 x i8], align 8
     %val = call i64 @read_ptr(ptr %large_var)
@@ -45,6 +48,7 @@ entry:
 define i64 @test_func_4128(i64 %idx) {
 ; CHECK-LABEL: test_func_4128
 ; CHECK-V3: add64 r10, 64
+; CHECK-OPT: add64 r10, 64
 ; The stack is aligned at 64, so we bump 64 to have a stack size of 4096+64=4160,
 ; so we can fit the 4128 bytes of the array.
 entry: