Enabling the generation of dependency breakers for partial updates on Cortex-A15. Also fixing a small bug in getting the update clearence for VLD1LNd32.

llvm-svn: 178134

lib/Target/ARM/ARMBaseInstrInfo.cpp

@@ -3734,9 +3734,9 @@ ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
   if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
     return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
 
-  // A9-like cores are particularly picky about mixing the two and want these
+  // CortexA9 is particularly picky about mixing the two and wants these
   // converted.
-  if (Subtarget.isLikeA9() && !isPredicated(MI) &&
+  if (Subtarget.isCortexA9() && !isPredicated(MI) &&
       (MI->getOpcode() == ARM::VMOVRS ||
        MI->getOpcode() == ARM::VMOVSR ||
        MI->getOpcode() == ARM::VMOVS))
@@ -4023,14 +4023,12 @@ ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
 // VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
 //
 // FCONSTD can be used as a dependency-breaking instruction.
-
-
 unsigned ARMBaseInstrInfo::
 getPartialRegUpdateClearance(const MachineInstr *MI,
                              unsigned OpNum,
                              const TargetRegisterInfo *TRI) const {
-  // Only Swift has partial register update problems.
-  if (!SwiftPartialUpdateClearance || !Subtarget.isSwift())
+  if (!SwiftPartialUpdateClearance ||
+      !(Subtarget.isSwift() || Subtarget.isCortexA15()))
     return 0;
 
   assert(TRI && "Need TRI instance");
@@ -4056,7 +4054,7 @@ getPartialRegUpdateClearance(const MachineInstr *MI,
 
     // Explicitly reads the dependency.
   case ARM::VLD1LNd32:
-    UseOp = 1;
+    UseOp = 3;
     break;
   default:
     return 0;

lib/Target/ARM/ARMTargetMachine.cpp

@@ -185,8 +185,7 @@ bool ARMPassConfig::addPreSched2() {
       addPass(createARMLoadStoreOptimizationPass());
       printAndVerify("After ARM load / store optimizer");
     }
-    if ((DisableA15SDOptimization || !getARMSubtarget().isCortexA15()) &&
-      getARMSubtarget().hasNEON())
+    if (getARMSubtarget().hasNEON())
       addPass(createExecutionDependencyFixPass(&ARM::DPRRegClass));
   }
 

test/CodeGen/ARM/a15-SD-dep.ll

@@ -5,7 +5,7 @@
 ; CHECK-DISABLED: t1:
 define <2 x float> @t1(float %f) {
   ; CHECK-ENABLED: vdup.32 d{{[0-9]*}}, d0[0]
-  ; CHECK-DISABLED: vmov.32 d0[1], r{{.}}
+  ; CHECK-DISABLED-NOT: vdup.32 d{{[0-9]*}}, d0[0]
   %i1 = insertelement <2 x float> undef, float %f, i32 1
   %i2 = fadd <2 x float> %i1, %i1
   ret <2 x float> %i2
@@ -15,7 +15,7 @@ define <2 x float> @t1(float %f) {
 ; CHECK-DISABLED: t2:
 define <4 x float> @t2(float %g, float %f) {
   ; CHECK-ENABLED: vdup.32 q{{[0-9]*}}, d0[0]
-  ; CHECK-DISABLED: vmov.32 d0[1], r{{.}}
+  ; CHECK-DISABLED-NOT: vdup.32 d{{[0-9]*}}, d0[0]
   %i1 = insertelement <4 x float> undef, float %f, i32 1
   %i2 = fadd <4 x float> %i1, %i1
   ret <4 x float> %i2
@@ -25,6 +25,7 @@ define <4 x float> @t2(float %g, float %f) {
 ; CHECK-DISABLED: t3:
 define arm_aapcs_vfpcc <2 x float> @t3(float %f) {
   ; CHECK-ENABLED: vdup.32 d{{[0-9]*}}, d0[0] 
+  ; CHECK-DISABLED-NOT: vdup.32 d{{[0-9]*}}, d0[0]
   %i1 = insertelement <2 x float> undef, float %f, i32 1
   %i2 = fadd <2 x float> %i1, %i1
   ret <2 x float> %i2

test/CodeGen/ARM/a15-partial-update.ll

@@ -0,0 +1,38 @@
+; RUN: llc -O1 -mcpu=cortex-a15 -mtriple=armv7-linux-gnueabi -verify-machineinstrs < %s  | FileCheck %s
+
+; CHECK: t1:
+define <2 x float> @t1(float* %A, <2 x float> %B) {
+; The generated code for this test uses a vld1.32 instruction
+; to write the lane 1 of a D register containing the value of
+; <2 x float> %B. Since the D register is defined, it would
+; be incorrect to fully write it (with a vmov.f64) before the
+; vld1.32 instruction. The test checks that a vmov.f64 was not
+; generated.
+
+; CHECK-NOT: vmov.{{.*}} d{{[0-9]+}},
+  %tmp2 = load float* %A, align 4
+  %tmp3 = insertelement <2 x float> %B, float %tmp2, i32 1
+  ret <2 x float> %tmp3
+}
+
+; CHECK: t2:
+define void @t2(<4 x i8> *%in, <4 x i8> *%out, i32 %n) {
+entry:
+  br label %loop
+loop:
+; The code generated by this test uses a vld1.32 instruction.
+; We check that a dependency breaking vmov* instruction was
+; generated.
+
+; CHECK: vmov.{{.*}} d{{[0-9]+}},
+  %oldcount = phi i32 [0, %entry], [%newcount, %loop]
+  %newcount = add i32 %oldcount, 1
+  %p1 = getelementptr <4 x i8> *%in, i32 %newcount
+  %p2 = getelementptr <4 x i8> *%out, i32 %newcount
+  %tmp1 = load <4 x i8> *%p1, align 4
+  store <4 x i8> %tmp1, <4 x i8> *%p2
+  %cmp = icmp eq i32 %newcount, %n
+  br i1 %cmp, label %loop, label %ret
+ret:
+  ret void
+}