[llvm] r191131 - [X86] Emulate AVX 256bit MIN/MAX support by splitting the vector.
Juergen Ributzka
juergen at apple.com
Fri Sep 20 21:55:22 PDT 2013
Author: ributzka
Date: Fri Sep 20 23:55:22 2013
New Revision: 191131
URL: http://llvm.org/viewvc/llvm-project?rev=191131&view=rev
Log:
[X86] Emulate AVX 256bit MIN/MAX support by splitting the vector.
In AVX 256bit vectors are valid vectors and therefore the Type Legalizer doesn't
split the VSELECT and SETCC nodes. AVX only supports MIN/MAX on 128bit vectors
and this fix enables vector splitting for this special case in the X86 DAG
Combiner.
This fix is related to PR16695, PR17002, and <rdar://problem/14594431>.
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
llvm/trunk/test/CodeGen/X86/vec_split.ll
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=191131&r1=191130&r2=191131&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Fri Sep 20 23:55:22 2013
@@ -16371,24 +16371,28 @@ static SDValue PerformEXTRACT_VECTOR_ELT
}
/// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
-static unsigned matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS,
- SDValue RHS, SelectionDAG &DAG,
- const X86Subtarget *Subtarget) {
+static std::pair<unsigned, bool>
+matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
+ SelectionDAG &DAG, const X86Subtarget *Subtarget) {
if (!VT.isVector())
- return 0;
+ return std::make_pair(0, false);
+ bool NeedSplit = false;
switch (VT.getSimpleVT().SimpleTy) {
- default: return 0;
+ default: return std::make_pair(0, false);
case MVT::v32i8:
case MVT::v16i16:
case MVT::v8i32:
if (!Subtarget->hasAVX2())
- return 0;
+ NeedSplit = true;
+ if (!Subtarget->hasAVX())
+ return std::make_pair(0, false);
+ break;
case MVT::v16i8:
case MVT::v8i16:
case MVT::v4i32:
if (!Subtarget->hasSSE2())
- return 0;
+ return std::make_pair(0, false);
}
// SSE2 has only a small subset of the operations.
@@ -16399,6 +16403,7 @@ static unsigned matchIntegerMINMAX(SDVal
ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+ unsigned Opc = 0;
// Check for x CC y ? x : y.
if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
DAG.isEqualTo(RHS, Cond.getOperand(1))) {
@@ -16406,16 +16411,16 @@ static unsigned matchIntegerMINMAX(SDVal
default: break;
case ISD::SETULT:
case ISD::SETULE:
- return hasUnsigned ? X86ISD::UMIN : 0;
+ Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
case ISD::SETUGT:
case ISD::SETUGE:
- return hasUnsigned ? X86ISD::UMAX : 0;
+ Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
case ISD::SETLT:
case ISD::SETLE:
- return hasSigned ? X86ISD::SMIN : 0;
+ Opc = hasSigned ? X86ISD::SMIN : 0; break;
case ISD::SETGT:
case ISD::SETGE:
- return hasSigned ? X86ISD::SMAX : 0;
+ Opc = hasSigned ? X86ISD::SMAX : 0; break;
}
// Check for x CC y ? y : x -- a min/max with reversed arms.
} else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
@@ -16424,20 +16429,20 @@ static unsigned matchIntegerMINMAX(SDVal
default: break;
case ISD::SETULT:
case ISD::SETULE:
- return hasUnsigned ? X86ISD::UMAX : 0;
+ Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
case ISD::SETUGT:
case ISD::SETUGE:
- return hasUnsigned ? X86ISD::UMIN : 0;
+ Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
case ISD::SETLT:
case ISD::SETLE:
- return hasSigned ? X86ISD::SMAX : 0;
+ Opc = hasSigned ? X86ISD::SMAX : 0; break;
case ISD::SETGT:
case ISD::SETGE:
- return hasSigned ? X86ISD::SMIN : 0;
+ Opc = hasSigned ? X86ISD::SMIN : 0; break;
}
}
- return 0;
+ return std::make_pair(Opc, NeedSplit);
}
/// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
@@ -16795,9 +16800,30 @@ static SDValue PerformSELECTCombine(SDNo
}
// Try to match a min/max vector operation.
- if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC)
- if (unsigned Op = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget))
- return DAG.getNode(Op, DL, N->getValueType(0), LHS, RHS);
+ if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
+ unsigned Opc;
+ bool NeedSplit;
+ std::tie(Opc, NeedSplit) = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
+
+ if (Opc && NeedSplit) {
+ unsigned NumElems = VT.getVectorNumElements();
+ // Extract the LHS vectors
+ SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
+ SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
+
+ // Extract the RHS vectors
+ SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
+ SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
+
+ // Create min/max for each subvector
+ LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
+ RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
+
+ // Merge the result
+ return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
+ } else if (Opc)
+ return DAG.getNode(Opc, DL, VT, LHS, RHS);
+ }
// Simplify vector selection if the selector will be produced by CMPP*/PCMP*.
if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
Modified: llvm/trunk/test/CodeGen/X86/vec_split.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vec_split.ll?rev=191131&r1=191130&r2=191131&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/vec_split.ll (original)
+++ llvm/trunk/test/CodeGen/X86/vec_split.ll Fri Sep 20 23:55:22 2013
@@ -1,10 +1,14 @@
; RUN: llc -march=x86-64 -mcpu=corei7 < %s | FileCheck %s -check-prefix=SSE4
+; RUN: llc -march=x86-64 -mcpu=corei7-avx < %s | FileCheck %s -check-prefix=AVX1
; RUN: llc -march=x86-64 -mcpu=core-avx2 < %s | FileCheck %s -check-prefix=AVX2
define <16 x i16> @split16(<16 x i16> %a, <16 x i16> %b, <16 x i8> %__mask) {
; SSE4-LABEL: split16:
; SSE4: pminuw
; SSE4: pminuw
+; AVX1-LABEL: split16:
+; AVX1: vpminuw
+; AVX1: vpminuw
; AVX2-LABEL: split16:
; AVX2: vpminuw
; AVX2: ret
@@ -19,6 +23,11 @@ define <32 x i16> @split32(<32 x i16> %a
; SSE4: pminuw
; SSE4: pminuw
; SSE4: pminuw
+; AVX1-LABEL: split32:
+; AVX1: vpminuw
+; AVX1: vpminuw
+; AVX1: vpminuw
+; AVX1: vpminuw
; AVX2-LABEL: split32:
; AVX2: vpminuw
; AVX2: vpminuw
More information about the llvm-commits
mailing list