[MachineInstrBuilder] Support for adding a ConstantPoolIndex MO with an additional...

MachineInstrBuilder::addDisp can already add an immediate or global address MO with an adjusted offset, this patch adds support for constant pool indices as well.

All remaining MO types still assert - there are a number of other types that could support adjusted offsets but I have no test cases at this time.

Required to fix a regression in D13988 found by Mikael Holmén during stress testing (test case attached).

Differential Revision: http://reviews.llvm.org/D14867

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253795 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/CodeGen/MachineInstrBuilder.h b/include/llvm/CodeGen/MachineInstrBuilder.h
index 947fcc70442dfc5624190347f11172a8bc4eb913..4afee44c4b31d4344e8b4e54e478fa3a2aafe7e1 100644 (file)
--- a/include/llvm/CodeGen/MachineInstrBuilder.h
+++ b/include/llvm/CodeGen/MachineInstrBuilder.h
@@ -200,22 +200,24 @@ public:
   // Add a displacement from an existing MachineOperand with an added offset.
   const MachineInstrBuilder &addDisp(const MachineOperand &Disp, int64_t off,
                                      unsigned char TargetFlags = 0) const {
+    // If caller specifies new TargetFlags then use it, otherwise the
+    // default behavior is to copy the target flags from the existing
+    // MachineOperand. This means if the caller wants to clear the
+    // target flags it needs to do so explicitly.
+    if (0 == TargetFlags)
+      TargetFlags = Disp.getTargetFlags();
+
     switch (Disp.getType()) {
       default:
         llvm_unreachable("Unhandled operand type in addDisp()");
       case MachineOperand::MO_Immediate:
         return addImm(Disp.getImm() + off);
-      case MachineOperand::MO_GlobalAddress: {
-        // If caller specifies new TargetFlags then use it, otherwise the
-        // default behavior is to copy the target flags from the existing
-        // MachineOperand. This means if the caller wants to clear the
-        // target flags it needs to do so explicitly.
-        if (TargetFlags)
-          return addGlobalAddress(Disp.getGlobal(), Disp.getOffset() + off,
-                                  TargetFlags);
+      case MachineOperand::MO_ConstantPoolIndex:
+        return addConstantPoolIndex(Disp.getIndex(), Disp.getOffset() + off,
+                                    TargetFlags);
+      case MachineOperand::MO_GlobalAddress:
         return addGlobalAddress(Disp.getGlobal(), Disp.getOffset() + off,
-                                Disp.getTargetFlags());
-      }
+                                TargetFlags);
     }
   }
 

diff --git a/test/CodeGen/X86/insertps-from-constantpool.ll b/test/CodeGen/X86/insertps-from-constantpool.ll
new file mode 100644 (file)
index 0000000..cfcfeac
--- /dev/null
+++ b/test/CodeGen/X86/insertps-from-constantpool.ll
@@ -0,0 +1,20 @@
+; RUN: llc -mtriple=i686-unknown-unknown -mattr=+sse4.1 < %s | FileCheck %s --check-prefix=X32
+; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 < %s | FileCheck %s --check-prefix=X64
+
+; Test for case where insertps folds the load of an insertion element from a constant pool.
+
+define <4 x float> @fold_from_constantpool(<4 x float> %a) {
+; X32-LABEL: fold_from_constantpool:
+; X32:       # BB#0:
+; X32-NEXT:    insertps {{.*#+}} xmm0 = mem[0],xmm0[1,2,3]
+; X32-NEXT:    retl
+;
+; X64-LABEL: fold_from_constantpool:
+; X64:       # BB#0:
+; X64-NEXT:    insertps {{.*#+}} xmm0 = mem[0],xmm0[1,2,3]
+; X64-NEXT:    retq
+  %1 = call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> <float 0.0, float 1.0, float 0.0, float 0.0>, i8 64)
+  ret <4 x float> %1
+}
+
+declare <4 x float> @llvm.x86.sse41.insertps(<4 x float>, <4 x float>, i8) nounwind readnone