From f3a778d75759ebb5c388ae93522a4cfabaebbd13 Mon Sep 17 00:00:00 2001
From: Krzysztof Parzyszek <kparzysz@codeaurora.org>
Date: Thu, 20 Jul 2017 19:43:19 +0000
Subject: [PATCH] Implement LaneBitmask::getNumLanes and
 LaneBitmask::getHighestLane

This should eliminate most uses of countPopulation and Log2_32 on
the lane mask values.

llvm-svn: 308658
---
 llvm/include/llvm/MC/LaneBitmask.h        | 7 +++++++
 llvm/lib/CodeGen/SplitKit.cpp             | 6 +++---
 llvm/lib/Target/AMDGPU/GCNRegPressure.cpp | 2 +-
 llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp | 3 +--
 llvm/utils/TableGen/CodeGenRegisters.cpp  | 8 ++------
 5 files changed, 14 insertions(+), 12 deletions(-)

diff --git a/llvm/include/llvm/MC/LaneBitmask.h b/llvm/include/llvm/MC/LaneBitmask.h
index 73b987b074db..35f472d817a0 100644
--- a/llvm/include/llvm/MC/LaneBitmask.h
+++ b/llvm/include/llvm/MC/LaneBitmask.h
@@ -73,6 +73,13 @@ namespace llvm {
 
     constexpr Type getAsInteger() const { return Mask; }
 
+    unsigned getNumLanes() const {
+      return countPopulation(Mask);
+    }
+    unsigned getHighestLane() const {
+      return Log2_32(Mask);
+    }
+
     static LaneBitmask getNone() { return LaneBitmask(0); }
     static LaneBitmask getAll()  { return ~LaneBitmask(0); }
     static LaneBitmask getLane(unsigned Lane) {
diff --git a/llvm/lib/CodeGen/SplitKit.cpp b/llvm/lib/CodeGen/SplitKit.cpp
index 323045fd2aaa..aaaa342b520b 100644
--- a/llvm/lib/CodeGen/SplitKit.cpp
+++ b/llvm/lib/CodeGen/SplitKit.cpp
@@ -552,7 +552,7 @@ SlotIndex SplitEditor::buildCopy(unsigned FromReg, unsigned ToReg,
     if ((SubRegMask & ~LaneMask).any())
       continue;
 
-    unsigned PopCount = countPopulation(SubRegMask.getAsInteger());
+    unsigned PopCount = SubRegMask.getNumLanes();
     PossibleIndexes.push_back(Idx);
     if (PopCount > BestCover) {
       BestCover = PopCount;
@@ -583,8 +583,8 @@ SlotIndex SplitEditor::buildCopy(unsigned FromReg, unsigned ToReg,
 
       // Try to cover as much of the remaining lanes as possible but
       // as few of the already covered lanes as possible.
-      int Cover = countPopulation((SubRegMask & LanesLeft).getAsInteger())
-                - countPopulation((SubRegMask & ~LanesLeft).getAsInteger());
+      int Cover = (SubRegMask & LanesLeft).getNumLanes()
+                - (SubRegMask & ~LanesLeft).getNumLanes();
       if (Cover > BestCover) {
         BestCover = Cover;
         BestIdx = Idx;
diff --git a/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp b/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
index 09cac8c2c8f1..0384340174fe 100644
--- a/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
@@ -107,7 +107,7 @@ void GCNRegPressure::inc(unsigned Reg,
     assert(PrevMask < NewMask);
 
     Value[Kind == SGPR_TUPLE ? SGPR32 : VGPR32] +=
-      Sign * countPopulation((~PrevMask & NewMask).getAsInteger());
+      Sign * (~PrevMask & NewMask).getNumLanes();
 
     if (PrevMask.none()) {
       assert(NewMask.any());
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
index 4a3fbb4593bb..d9a5ce338bfe 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -1275,8 +1275,7 @@ const TargetRegisterClass *SIRegisterInfo::getSubRegClass(
     return RC;
 
   // We can assume that each lane corresponds to one 32-bit register.
-  LaneBitmask::Type Mask = getSubRegIndexLaneMask(SubIdx).getAsInteger();
-  unsigned Count = countPopulation(Mask);
+  unsigned Count = getSubRegIndexLaneMask(SubIdx).getNumLanes();
   if (isSGPRClass(RC)) {
     switch (Count) {
     case 1:
diff --git a/llvm/utils/TableGen/CodeGenRegisters.cpp b/llvm/utils/TableGen/CodeGenRegisters.cpp
index 6399fb5ec1dd..77450aef9a5c 100644
--- a/llvm/utils/TableGen/CodeGenRegisters.cpp
+++ b/llvm/utils/TableGen/CodeGenRegisters.cpp
@@ -1295,9 +1295,7 @@ void CodeGenRegBank::computeSubRegLaneMasks() {
       // Moving from a class with no subregisters we just had a single lane:
       // The subregister must be a leaf subregister and only occupies 1 bit.
       // Move the bit from the class without subregisters into that position.
-      static_assert(sizeof(Idx.LaneMask.getAsInteger()) == 4,
-                    "Change Log2_32 to a proper one");
-      unsigned DstBit = Log2_32(Idx.LaneMask.getAsInteger());
+      unsigned DstBit = Idx.LaneMask.getHighestLane();
       assert(Idx.LaneMask == LaneBitmask::getLane(DstBit) &&
              "Must be a leaf subregister");
       MaskRolPair MaskRol = { LaneBitmask::getLane(0), (uint8_t)DstBit };
@@ -1328,9 +1326,7 @@ void CodeGenRegBank::computeSubRegLaneMasks() {
         assert(Composite->getComposites().empty());
 
         // Create Mask+Rotate operation and merge with existing ops if possible.
-        static_assert(sizeof(Composite->LaneMask.getAsInteger()) == 4,
-                      "Change Log2_32 to a proper one");
-        unsigned DstBit = Log2_32(Composite->LaneMask.getAsInteger());
+        unsigned DstBit = Composite->LaneMask.getHighestLane();
         int Shift = DstBit - SrcBit;
         uint8_t RotateLeft = Shift >= 0 ? (uint8_t)Shift
                                         : LaneBitmask::BitWidth + Shift;