[ARM] Add new target feature to fuse literal generation

This feature enables the fusion of such operations on Cortex A57 and Cortex
A72, as recommended in their Software Optimisation Guides, sections 4.14 and
4.11, respectively.

Differential revision: https://reviews.llvm.org/D49563

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@338147 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARM.td b/lib/Target/ARM/ARM.td
index 742b355188948becdb82bebc87905a582a53b96b..2e62a07904188004b718afea8d1acdeaaf173183 100644 (file)
--- a/lib/Target/ARM/ARM.td
+++ b/lib/Target/ARM/ARM.td
@@ -141,6 +141,10 @@ def FeatureFPAO           : SubtargetFeature<"fpao", "HasFPAO", "true",
 def FeatureFuseAES        : SubtargetFeature<"fuse-aes", "HasFuseAES", "true",
                                              "CPU fuses AES crypto operations">;
 
+// Fast execution of bottom and top halves of literal generation
+def FeatureFuseLiterals   : SubtargetFeature<"fuse-literals", "HasFuseLiterals", "true",
+                                             "CPU fuses literal generation operations">;
+
 // The way of reading thread pointer                                             
 def FeatureReadTp :  SubtargetFeature<"read-tp-hard", "ReadTPHard", "true",
                                       "Reading thread pointer from register">;

diff --git a/lib/Target/ARM/ARMMacroFusion.cpp b/lib/Target/ARM/ARMMacroFusion.cpp
index f2dc650a6f3ea94d8724423d3f366f2b5286890a..d11fe9d5c502dccfc575c0a00432010eb0072cb8 100644 (file)
--- a/lib/Target/ARM/ARMMacroFusion.cpp
+++ b/lib/Target/ARM/ARMMacroFusion.cpp
@@ -19,6 +19,47 @@
 
 namespace llvm {
 
+// Fuse AES crypto encoding or decoding.
+static bool isAESPair(const MachineInstr *FirstMI,
+                      const MachineInstr &SecondMI) {
+  // Assume the 1st instr to be a wildcard if it is unspecified.
+  unsigned FirstOpcode =
+      FirstMI ? FirstMI->getOpcode()
+              : static_cast<unsigned>(ARM::INSTRUCTION_LIST_END);
+  unsigned SecondOpcode = SecondMI.getOpcode();
+
+  switch(SecondOpcode) {
+  // AES encode.
+  case ARM::AESMC :
+    return FirstOpcode == ARM::AESE ||
+           FirstOpcode == ARM::INSTRUCTION_LIST_END;
+  // AES decode.
+  case ARM::AESIMC:
+    return FirstOpcode == ARM::AESD ||
+           FirstOpcode == ARM::INSTRUCTION_LIST_END;
+  }
+
+  return false;
+}
+
+// Fuse literal generation.
+static bool isLiteralsPair(const MachineInstr *FirstMI,
+                           const MachineInstr &SecondMI) {
+  // Assume the 1st instr to be a wildcard if it is unspecified.
+  unsigned FirstOpcode =
+      FirstMI ? FirstMI->getOpcode()
+              : static_cast<unsigned>(ARM::INSTRUCTION_LIST_END);
+  unsigned SecondOpcode = SecondMI.getOpcode();
+
+  // 32 bit immediate.
+  if ((FirstOpcode == ARM::INSTRUCTION_LIST_END ||
+       FirstOpcode == ARM::MOVi16) &&
+      SecondOpcode == ARM::MOVTi16)
+    return true;
+
+  return false;
+}
+
 /// Check if the instr pair, FirstMI and SecondMI, should be fused
 /// together. Given SecondMI, when FirstMI is unspecified, then check if
 /// SecondMI may be part of a fused pair at all.
@@ -28,24 +69,10 @@ static bool shouldScheduleAdjacent(const TargetInstrInfo &TII,
                                    const MachineInstr &SecondMI) {
   const ARMSubtarget &ST = static_cast<const ARMSubtarget&>(TSI);
 
-  // Assume wildcards for unspecified instrs.
-  unsigned FirstOpcode =
-    FirstMI ? FirstMI->getOpcode()
-            : static_cast<unsigned>(ARM::INSTRUCTION_LIST_END);
-  unsigned SecondOpcode = SecondMI.getOpcode();
-
-  if (ST.hasFuseAES())
-    // Fuse AES crypto operations.
-    switch(SecondOpcode) {
-    // AES encode.
-    case ARM::AESMC :
-      return FirstOpcode == ARM::AESE ||
-             FirstOpcode == ARM::INSTRUCTION_LIST_END;
-    // AES decode.
-    case ARM::AESIMC:
-      return FirstOpcode == ARM::AESD ||
-             FirstOpcode == ARM::INSTRUCTION_LIST_END;
-    }
+  if (ST.hasFuseAES() && isAESPair(FirstMI, SecondMI))
+    return true;
+  if (ST.hasFuseLiterals() && isLiteralsPair(FirstMI, SecondMI))
+    return true;
 
   return false;
 }

diff --git a/lib/Target/ARM/ARMSubtarget.h b/lib/Target/ARM/ARMSubtarget.h
index 165077926c88bda521b3764cceed2a74084ef1d0..74aee9a8ed38bb2af8eb34c4956a4cec85e3ca5c 100644 (file)
--- a/lib/Target/ARM/ARMSubtarget.h
+++ b/lib/Target/ARM/ARMSubtarget.h
@@ -327,6 +327,10 @@ protected:
   /// pairs faster.
   bool HasFuseAES = false;
 
+  /// HasFuseLiterals - if true, processor executes back to back
+  /// bottom and top halves of literal generation faster.
+  bool HasFuseLiterals = false;
+
   /// If true, if conversion may decide to leave some instructions unpredicated.
   bool IsProfitableToUnpredicate = false;
 
@@ -616,8 +620,9 @@ public:
   bool hasFullFP16() const { return HasFullFP16; }
 
   bool hasFuseAES() const { return HasFuseAES; }
+  bool hasFuseLiterals() const { return HasFuseLiterals; }
   /// Return true if the CPU supports any kind of instruction fusion.
-  bool hasFusion() const { return hasFuseAES(); }
+  bool hasFusion() const { return hasFuseAES() || hasFuseLiterals(); }
 
   const Triple &getTargetTriple() const { return TargetTriple; }
 

diff --git a/test/CodeGen/ARM/misched-fusion-lit.ll b/test/CodeGen/ARM/misched-fusion-lit.ll
new file mode 100644 (file)
index 0000000..cfc4356
--- /dev/null
+++ b/test/CodeGen/ARM/misched-fusion-lit.ll
@@ -0,0 +1,39 @@
+; RUN: llc %s -o - -mtriple=armv8-unknown -mattr=-fuse-literals,+use-misched | FileCheck %s --check-prefix=CHECK --check-prefix=CHECKDONT
+; RUN: llc %s -o - -mtriple=armv8-unknown -mattr=+fuse-literals,+use-misched | FileCheck %s --check-prefix=CHECK --check-prefix=CHECKFUSE
+
+@g = common global i32* zeroinitializer
+
+define i32* @litp(i32 %a, i32 %b) {
+entry:
+  %add = add nsw i32 %b, %a
+  %ptr = getelementptr i32, i32* bitcast (i32* (i32, i32)* @litp to i32*), i32 %add
+  %res = getelementptr i32, i32* bitcast (i32** @g to i32*), i32 %add
+  store i32* %ptr, i32** @g, align 4
+  ret i32* %res
+
+; CHECK-LABEL: litp:
+; CHECK:          movw [[R:r[0-9]+]], :lower16:litp
+; CHECKDONT-NEXT: movw [[S:r[0-9]+]], :lower16:g
+; CHECKFUSE-NEXT: movt [[R]], :upper16:litp
+; CHECKFUSE-NEXT: movw [[S:r[0-9]+]], :lower16:g
+; CHECKFUSE-NEXT: movt [[S]], :upper16:g
+}
+
+define i32 @liti(i32 %a, i32 %b) {
+entry:
+  %adda = add i32 %a, -262095121
+  %add1 = add i32 %adda, %b
+  %addb = add i32 %b, 121110837
+  %add2 = add i32 %addb, %a
+  store i32 %add1, i32* bitcast (i32** @g to i32*), align 4
+  ret i32 %add2
+
+; CHECK-LABEL: liti:
+; CHECK:          movw [[R:r[0-9]+]], #309
+; CHECKDONT-NEXT: add {{r[0-9]+}}, {{r[0-9]+}}, {{r[0-9]+}}
+; CHECKFUSE-NEXT: movt [[R]], #1848
+; CHECKFUSE:      movw [[S:r[0-9]+]], :lower16:g
+; CHECKFUSE-NEXT: movt [[S]], :upper16:g
+; CHECKFUSE-NEXT: movw [[T:r[0-9]+]], #48879
+; CHECKFUSE-NEXT: movt [[T]], #61536
+}