Add trunc (select c, a, b) -> select c (trunc a), (trunc b) combine.

Do this if the truncate is free and the select is legal.

llvm-svn: 212640

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 7497feb..ab07954 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -6013,6 +6013,20 @@ SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
     }
   }
 
+  // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
+  if (N0.getOpcode() == ISD::SELECT) {
+    EVT SrcVT = N0.getValueType();
+    if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
+        TLI.isTruncateFree(SrcVT, VT)) {
+      SDLoc SL(N0);
+      SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
+      SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
+      EVT SetCCVT = getSetCCResultType(VT);
+      SDValue Cond = DAG.getSExtOrTrunc(N0.getOperand(0), SL, SetCCVT);
+      return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
+    }
+  }
+
   // Fold a series of buildvector, bitcast, and truncate if possible.
   // For example fold
   //   (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to

diff --git a/llvm/test/CodeGen/R600/select64.ll b/llvm/test/CodeGen/R600/select64.ll
index 6b87d98..dba25e3 100644 (file)
--- a/llvm/test/CodeGen/R600/select64.ll
+++ b/llvm/test/CodeGen/R600/select64.ll
@@ -13,3 +13,38 @@ entry:
   store i64 %1, i64 addrspace(1)* %out
   ret void
 }
+
+; CHECK-LABEL: @select_trunc_i64
+; CHECK: V_CNDMASK_B32
+; CHECK-NOT: V_CNDMASK_B32
+define void @select_trunc_i64(i32 addrspace(1)* %out, i32 %cond, i64 %in) nounwind {
+  %cmp = icmp ugt i32 %cond, 5
+  %sel = select i1 %cmp, i64 0, i64 %in
+  %trunc = trunc i64 %sel to i32
+  store i32 %trunc, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; CHECK-LABEL: @select_trunc_i64_2
+; CHECK: V_CNDMASK_B32
+; CHECK-NOT: V_CNDMASK_B32
+define void @select_trunc_i64_2(i32 addrspace(1)* %out, i32 %cond, i64 %a, i64 %b) nounwind {
+  %cmp = icmp ugt i32 %cond, 5
+  %sel = select i1 %cmp, i64 %a, i64 %b
+  %trunc = trunc i64 %sel to i32
+  store i32 %trunc, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; CHECK-LABEL: @v_select_trunc_i64_2
+; CHECK: V_CNDMASK_B32
+; CHECK-NOT: V_CNDMASK_B32
+define void @v_select_trunc_i64_2(i32 addrspace(1)* %out, i32 %cond, i64 addrspace(1)* %aptr, i64 addrspace(1)* %bptr) nounwind {
+  %cmp = icmp ugt i32 %cond, 5
+  %a = load i64 addrspace(1)* %aptr, align 8
+  %b = load i64 addrspace(1)* %bptr, align 8
+  %sel = select i1 %cmp, i64 %a, i64 %b
+  %trunc = trunc i64 %sel to i32
+  store i32 %trunc, i32 addrspace(1)* %out, align 4
+  ret void
+}

diff --git a/llvm/test/CodeGen/X86/shift-parts.ll b/llvm/test/CodeGen/X86/shift-parts.ll
index ce4f538..763da63 100644 (file)
--- a/llvm/test/CodeGen/X86/shift-parts.ll
+++ b/llvm/test/CodeGen/X86/shift-parts.ll
@@ -1,17 +1,19 @@
-; RUN: llc < %s -march=x86-64 | grep shrdq
+; RUN: llc -march=x86-64 < %s | FileCheck %s
 ; PR4736
 
 %0 = type { i32, i8, [35 x i8] }
 
 @g_144 = external global %0, align 8              ; <%0*> [#uses=1]
 
-define i32 @int87(i32 %uint64p_8) nounwind {
+; CHECK: shrdq
+
+define i32 @int87(i32 %uint64p_8, i1 %cond) nounwind {
 entry:
   %srcval4 = load i320* bitcast (%0* @g_144 to i320*), align 8 ; <i320> [#uses=1]
   br label %for.cond
 
 for.cond:                                         ; preds = %for.cond, %entry
-  %call3.in.in.in.v = select i1 undef, i320 192, i320 128 ; <i320> [#uses=1]
+  %call3.in.in.in.v = select i1 %cond, i320 192, i320 128 ; <i320> [#uses=1]
   %call3.in.in.in = lshr i320 %srcval4, %call3.in.in.in.v ; <i320> [#uses=1]
   %call3.in = trunc i320 %call3.in.in.in to i32   ; <i32> [#uses=1]
   %tobool = icmp eq i32 %call3.in, 0              ; <i1> [#uses=1]