[x86, SSE] AVX1 PR28129 (256-bit all-ones rematerialization)

Further perf tests on Jaguar indicate that:

vxorps  %ymm0, %ymm0, %ymm0
vcmpps  $15, %ymm0, %ymm0, %ymm0

is consistently faster (by about 9%) than:

vpcmpeqd  %xmm0, %xmm0, %xmm0
vinsertf128  $1, %xmm0, %ymm0, %ymm0

Testing equivalent code on a SandyBridge (E5-2640) puts it slightly (~3%) faster as well.

Committed on behalf of @dtemirbulatov

Differential Revision: https://reviews.llvm.org/D32416

llvm-svn: 302989

2 files changed, 22 insertions, 10 deletions

diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index 33c9fcf839e..092ceb207ad 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -7627,6 +7627,13 @@ bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
   case X86::AVX2_SETALLONES:
     return Expand2AddrUndef(MIB, get(X86::VPCMPEQDYrr));
+  case X86::AVX1_SETALLONES: {
+    unsigned Reg = MIB->getOperand(0).getReg();
+    // VCMPPSYrri with an immediate 0xf should produce VCMPTRUEPS.
+    MIB->setDesc(get(X86::VCMPPSYrri));
+    MIB.addReg(Reg, RegState::Undef).addReg(Reg, RegState::Undef).addImm(0xf);
+    return true;
+  }
   case X86::AVX512_512_SETALLONES: {
     unsigned Reg = MIB->getOperand(0).getReg();
     MIB->setDesc(get(X86::VPTERNLOGDZrri));
@@ -8515,6 +8522,7 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
       Alignment = 64;
       break;
     case X86::AVX2_SETALLONES:
+    case X86::AVX1_SETALLONES:
     case X86::AVX_SET0:
     case X86::AVX512_256_SET0:
       Alignment = 32;
@@ -8560,6 +8568,7 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
   case X86::V_SET0:
   case X86::V_SETALLONES:
   case X86::AVX2_SETALLONES:
+  case X86::AVX1_SETALLONES:
   case X86::AVX_SET0:
   case X86::AVX512_128_SET0:
   case X86::AVX512_256_SET0:
@@ -8601,13 +8610,14 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
     else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES)
       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()),16);
     else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 ||
-             Opc == X86::AVX512_256_SET0)
+             Opc == X86::AVX512_256_SET0 || Opc == X86::AVX1_SETALLONES)
       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 8);
     else
       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction()->getContext()), 4);
 
     bool IsAllOnes = (Opc == X86::V_SETALLONES || Opc == X86::AVX2_SETALLONES ||
-                      Opc == X86::AVX512_512_SETALLONES);
+                      Opc == X86::AVX512_512_SETALLONES ||
+                      Opc == X86::AVX1_SETALLONES);
     const Constant *C = IsAllOnes ? Constant::getAllOnesValue(Ty) :
                                     Constant::getNullValue(Ty);
     unsigned CPI = MCP.getConstantPoolIndex(C, Alignment);
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index 48da2fa607a..f73d85e7e01 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -486,6 +486,10 @@ let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
     isPseudo = 1, SchedRW = [WriteZero] in {
   def V_SETALLONES : I<0, Pseudo, (outs VR128:$dst), (ins), "",
                        [(set VR128:$dst, (v4i32 immAllOnesV))]>;
+  let Predicates = [HasAVX1Only, OptForMinSize] in {
+  def AVX1_SETALLONES: I<0, Pseudo, (outs VR256:$dst), (ins), "",
+                          [(set VR256:$dst, (v8i32 immAllOnesV))]>;
+  }
   let Predicates = [HasAVX2] in
   def AVX2_SETALLONES : I<0, Pseudo, (outs VR256:$dst), (ins), "",
                           [(set VR256:$dst, (v8i32 immAllOnesV))]>;
@@ -7755,14 +7759,12 @@ def VINSERTF128rm : AVXAIi8<0x18, MRMSrcMem, (outs VR256:$dst),
           []>, Sched<[WriteFShuffleLd, ReadAfterLd]>, VEX_4V, VEX_L;
 }
 
-
-// Without AVX2 we need to concat two v4i32 V_SETALLONES to create a 256-bit
-// all ones value.
-let Predicates = [HasAVX1Only] in
-def : Pat<(v8i32 immAllOnesV),
-          (VINSERTF128rr
-           (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), (V_SETALLONES), sub_xmm),
-           (V_SETALLONES), 1)>;
+// To create a 256-bit all ones value, we should produce VCMPTRUEPS
+// with YMM register containing zero.
+// FIXME: Avoid producing vxorps to clear the fake inputs.
+let Predicates = [HasAVX1Only] in {
+def : Pat<(v8i32 immAllOnesV), (VCMPPSYrri (AVX_SET0), (AVX_SET0), 0xf)>;
+}
 
 multiclass vinsert_lowering<string InstrStr, ValueType From, ValueType To,
                             PatFrag memop_frag> {