[ARM]Fix an assertion failure in A15SDOptimizer about DPair reg class by treating...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204304 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/A15SDOptimizer.cpp b/lib/Target/ARM/A15SDOptimizer.cpp
index 8edf00a70a60097f23518d126e330e36b3475357..37bf9033a058438667b219fd3c6af8887424eddd 100644 (file)
--- a/lib/Target/ARM/A15SDOptimizer.cpp
+++ b/lib/Target/ARM/A15SDOptimizer.cpp
@@ -416,7 +416,8 @@ SmallVector<unsigned, 8> A15SDOptimizer::getReadDPRs(MachineInstr *MI) {
     if (!MO.isReg() || !MO.isUse())
       continue;
     if (!usesRegClass(MO, &ARM::DPRRegClass) &&
-        !usesRegClass(MO, &ARM::QPRRegClass))
+        !usesRegClass(MO, &ARM::QPRRegClass) &&
+        !usesRegClass(MO, &ARM::DPairRegClass)) // Treat DPair as QPR
       continue;
 
     Defs.push_back(MO.getReg());
@@ -536,7 +537,10 @@ A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) {
   InsertPt++;
   unsigned Out;
 
-  if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::QPRRegClass)) {
+  // DPair has the same length as QPR and also has two DPRs as subreg.
+  // Treat DPair as QPR.
+  if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::QPRRegClass) ||
+      MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::DPairRegClass)) {
     unsigned DSub0 = createExtractSubreg(MBB, InsertPt, DL, Reg,
                                          ARM::dsub_0, &ARM::DPRRegClass);
     unsigned DSub1 = createExtractSubreg(MBB, InsertPt, DL, Reg,
@@ -569,7 +573,9 @@ A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) {
       default: llvm_unreachable("Unknown preferred lane!");
     }
 
-    bool UsesQPR = usesRegClass(MI->getOperand(0), &ARM::QPRRegClass);
+    // Treat DPair as QPR
+    bool UsesQPR = usesRegClass(MI->getOperand(0), &ARM::QPRRegClass) ||
+                   usesRegClass(MI->getOperand(0), &ARM::DPairRegClass);
 
     Out = createImplicitDef(MBB, InsertPt, DL);
     Out = createInsertSubreg(MBB, InsertPt, DL, Out, PrefLane, Reg);

diff --git a/test/CodeGen/ARM/a15-SD-dep.ll b/test/CodeGen/ARM/a15-SD-dep.ll
index 019ff6129b0072550ec8626f5893247fd015d0bc..5e5ca4b873f3872fec8c0b3f4701ca02f3ff09b4 100644 (file)
--- a/test/CodeGen/ARM/a15-SD-dep.ll
+++ b/test/CodeGen/ARM/a15-SD-dep.ll
@@ -56,3 +56,62 @@ define arm_aapcs_vfpcc <4 x float> @t5(<4 x float> %q, float %f) {
   %i2 = fadd <4 x float> %i1, %i1
   ret <4 x float> %i2
 }
+
+; Test that DPair can be successfully passed as QPR.
+; CHECK-ENABLED-LABEL: test_DPair1:
+; CHECK-DISABLED-LABEL: test_DPair1:
+define void @test_DPair1(i32 %vsout, i8* nocapture %out, float %x, float %y) {
+entry:
+  %0 = insertelement <4 x float> undef, float %x, i32 1
+  %1 = insertelement <4 x float> %0, float %y, i32 0
+  ; CHECK-ENABLED: vdup.32 d{{[0-9]*}}, d{{[0-9]*}}[0]
+  ; CHECK-ENABLED: vdup.32 d{{[0-9]*}}, d{{[0-9]*}}[1]
+  ; CHECK-ENABLED: vdup.32 d{{[0-9]*}}, d{{[0-9]*}}[0]
+  ; CHECK-ENABLED: vdup.32 d{{[0-9]*}}, d{{[0-9]*}}[1]
+  ; CHECK-DISABLED-NOT: vdup
+  switch i32 %vsout, label %sw.epilog [
+    i32 1, label %sw.bb
+    i32 0, label %sw.bb6
+  ]
+
+sw.bb:                                            ; preds = %entry
+  %2 = insertelement <4 x float> %1, float 0.000000e+00, i32 0
+  br label %sw.bb6
+
+sw.bb6:                                           ; preds = %sw.bb, %entry
+  %sum.0 = phi <4 x float> [ %1, %entry ], [ %2, %sw.bb ]
+  %3 = extractelement <4 x float> %sum.0, i32 0
+  %conv = fptoui float %3 to i8
+  store i8 %conv, i8* %out, align 1
+  ret void
+
+sw.epilog:                                        ; preds = %entry
+  ret void
+}
+
+; CHECK-ENABLED-LABEL: test_DPair2:
+; CHECK-DISABLED-LABEL: test_DPair2:
+define void @test_DPair2(i32 %vsout, i8* nocapture %out, float %x) {
+entry:
+  %0 = insertelement <4 x float> undef, float %x, i32 0
+  ; CHECK-ENABLED: vdup.32 q{{[0-9]*}}, d{{[0-9]*}}[0]
+  ; CHECK-DISABLED-NOT: vdup
+  switch i32 %vsout, label %sw.epilog [
+    i32 1, label %sw.bb
+    i32 0, label %sw.bb1
+  ]
+
+sw.bb:                                            ; preds = %entry
+  %1 = insertelement <4 x float> %0, float 0.000000e+00, i32 0
+  br label %sw.bb1
+
+sw.bb1:                                           ; preds = %entry, %sw.bb
+  %sum.0 = phi <4 x float> [ %0, %entry ], [ %1, %sw.bb ]
+  %2 = extractelement <4 x float> %sum.0, i32 0
+  %conv = fptoui float %2 to i8
+  store i8 %conv, i8* %out, align 1
+  br label %sw.epilog
+
+sw.epilog:                                        ; preds = %entry, %sw.bb1
+  ret void
+}
\ No newline at end of file