[llvm-commits] [llvm] r155397 - in /llvm/trunk: lib/Target/X86/X86ISelLowering.cpp test/CodeGen/X86/vec_cast2.ll

[Nadav Rotem]

Author: nadav
Date: Mon Apr 23 16:53:37 2012
New Revision: 155397

URL: http://llvm.org/viewvc/llvm-project?rev=155397&view=rev
Log:
Optimize the vector UINT_TO_FP, SINT_TO_FP and FP_TO_SINT operations where the integer type is i8 (commonly used in graphics).


Added:
    llvm/trunk/test/CodeGen/X86/vec_cast2.ll
Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=155397&r1=155396&r2=155397&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Mon Apr 23 16:53:37 2012
@@ -1221,7 +1221,9 @@
   setTargetDAGCombine(ISD::ANY_EXTEND);
   setTargetDAGCombine(ISD::SIGN_EXTEND);
   setTargetDAGCombine(ISD::TRUNCATE);
+  setTargetDAGCombine(ISD::UINT_TO_FP);
   setTargetDAGCombine(ISD::SINT_TO_FP);
+  setTargetDAGCombine(ISD::FP_TO_SINT);
   if (Subtarget->is64Bit())
     setTargetDAGCombine(ISD::MUL);
   if (Subtarget->hasBMI())
@@ -14985,9 +14987,43 @@
   return SDValue();
 }
 
+static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
+                                        const X86TargetLowering *XTLI) {
+  SDValue Op0 = N->getOperand(0);
+  EVT InVT = Op0->getValueType(0);
+  if (!InVT.isSimple())
+    return SDValue();
+
+  // UINT_TO_FP(v4i8) -> SINT_TO_FP(ZEXT(v4i8 to v4i32))
+  MVT SrcVT = InVT.getSimpleVT();
+  if (SrcVT == MVT::v8i8 || SrcVT == MVT::v4i8) {
+    DebugLoc dl = N->getDebugLoc();
+    MVT DstVT = (SrcVT.getVectorNumElements() == 4 ? MVT::v4i32 : MVT::v8i32);
+    SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
+    // Notice that we use SINT_TO_FP because we know that the high bits
+    // are zero and SINT_TO_FP is better supported by the hardware.
+    return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
+  }
+
+  return SDValue();
+}
+
 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
                                         const X86TargetLowering *XTLI) {
   SDValue Op0 = N->getOperand(0);
+  EVT InVT = Op0->getValueType(0);
+  if (!InVT.isSimple())
+    return SDValue();
+
+  // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
+  MVT SrcVT = InVT.getSimpleVT();
+  if (SrcVT == MVT::v8i8 || SrcVT == MVT::v4i8) {
+    DebugLoc dl = N->getDebugLoc();
+    MVT DstVT = (SrcVT.getVectorNumElements() == 4 ? MVT::v4i32 : MVT::v8i32);
+    SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
+    return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
+  }
+
   // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
   // a 32-bit target where SSE doesn't support i64->FP operations.
   if (Op0.getOpcode() == ISD::LOAD) {
@@ -15006,6 +15042,24 @@
   return SDValue();
 }
 
+static SDValue PerformFP_TO_SINTCombine(SDNode *N, SelectionDAG &DAG,
+                                        const X86TargetLowering *XTLI) {
+  EVT InVT = N->getValueType(0);
+  if (!InVT.isSimple())
+    return SDValue();
+
+  // v4i8 = FP_TO_SINT() -> v4i8 = TRUNCATE (V4i32 = FP_TO_SINT()
+  MVT VT = InVT.getSimpleVT();
+  if (VT == MVT::v8i8 || VT == MVT::v4i8) {
+    DebugLoc dl = N->getDebugLoc();
+    MVT DstVT = (VT.getVectorNumElements() == 4 ? MVT::v4i32 : MVT::v8i32);
+    SDValue I = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, N->getOperand(0));
+    return DAG.getNode(ISD::TRUNCATE, dl, VT, I);
+  }
+
+  return SDValue();
+}
+
 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
                                  X86TargetLowering::DAGCombinerInfo &DCI) {
@@ -15142,7 +15196,9 @@
   case ISD::XOR:            return PerformXorCombine(N, DAG, DCI, Subtarget);
   case ISD::LOAD:           return PerformLOADCombine(N, DAG, Subtarget);
   case ISD::STORE:          return PerformSTORECombine(N, DAG, Subtarget);
+  case ISD::UINT_TO_FP:     return PerformUINT_TO_FPCombine(N, DAG, this);
   case ISD::SINT_TO_FP:     return PerformSINT_TO_FPCombine(N, DAG, this);
+  case ISD::FP_TO_SINT:     return PerformFP_TO_SINTCombine(N, DAG, this);
   case ISD::FADD:           return PerformFADDCombine(N, DAG, Subtarget);
   case ISD::FSUB:           return PerformFSUBCombine(N, DAG, Subtarget);
   case X86ISD::FXOR:

Added: llvm/trunk/test/CodeGen/X86/vec_cast2.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vec_cast2.ll?rev=155397&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/vec_cast2.ll (added)
+++ llvm/trunk/test/CodeGen/X86/vec_cast2.ll Mon Apr 23 16:53:37 2012
@@ -0,0 +1,49 @@
+; RUN: llc < %s -mtriple=i386-apple-darwin10 -mcpu=corei7-avx -mattr=+avx | FileCheck %s
+
+;CHECK: foo1_8
+;CHECK: vcvtdq2ps
+;CHECK: ret
+define <8 x float> @foo1_8(<8 x i8> %src) {
+  %res = sitofp <8 x i8> %src to <8 x float>
+  ret <8 x float> %res
+}
+
+;CHECK: foo1_4
+;CHECK: vcvtdq2ps
+;CHECK: ret
+define <4 x float> @foo1_4(<4 x i8> %src) {
+  %res = sitofp <4 x i8> %src to <4 x float>
+  ret <4 x float> %res
+}
+
+;CHECK: foo2_8
+;CHECK: vcvtdq2ps
+;CHECK: ret
+define <8 x float> @foo2_8(<8 x i8> %src) {
+  %res = uitofp <8 x i8> %src to <8 x float>
+  ret <8 x float> %res
+}
+
+;CHECK: foo2_4
+;CHECK: vcvtdq2ps
+;CHECK: ret
+define <4 x float> @foo2_4(<4 x i8> %src) {
+  %res = uitofp <4 x i8> %src to <4 x float>
+  ret <4 x float> %res
+}
+
+;CHECK: foo3_8
+;CHECK: vcvttps2dq
+;CHECK: ret
+define <8 x i8> @foo3_8(<8 x float> %src) {
+  %res = fptosi <8 x float> %src to <8 x i8>
+  ret <8 x i8> %res
+}
+;CHECK: foo3_4
+;CHECK: vcvttps2dq
+;CHECK: ret
+define <4 x i8> @foo3_4(<4 x float> %src) {
+  %res = fptosi <4 x float> %src to <4 x i8>
+  ret <4 x i8> %res
+}
+