[AVX-512] Replace V_SET0 in AVX-512 patterns with AVX512_128_SET0. Enhance AVX512_128...

We'll now expand AVX512_128_SET0 to an EVEX VXORD if VLX available. Or if its not, but register allocation has selected a non-extended register we will use VEX VXORPS. And if its an extended register without VLX we'll use a 512-bit XOR. Do the same for AVX512_FsFLD0SS/SD.

This makes it possible for the register allocator to have all 32 registers available to work with.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@292004 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index a473073ead2ae2ca643503816ed1412eff67e9b2..5ef943c2c76ae1ac59427c058a8270c54bb3914b 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -460,7 +460,7 @@ def AVX512_512_SEXT_MASK_64 : I<0, Pseudo, (outs VR512:$dst),
 }
 
 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
-    isPseudo = 1, Predicates = [HasVLX], SchedRW = [WriteZero] in {
+    isPseudo = 1, Predicates = [HasAVX512], SchedRW = [WriteZero] in {
 def AVX512_128_SET0 : I<0, Pseudo, (outs VR128X:$dst), (ins), "",
                [(set VR128X:$dst, (v4i32 immAllZerosV))]>;
 def AVX512_256_SET0 : I<0, Pseudo, (outs VR256X:$dst), (ins), "",
@@ -470,7 +470,7 @@ def AVX512_256_SET0 : I<0, Pseudo, (outs VR256X:$dst), (ins), "",
 // Alias instructions that map fld0 to xorps for sse or vxorps for avx.
 // This is expanded by ExpandPostRAPseudos.
 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
-    isPseudo = 1, SchedRW = [WriteZero], Predicates = [HasVLX, HasDQI] in {
+    isPseudo = 1, SchedRW = [WriteZero], Predicates = [HasAVX512] in {
   def AVX512_FsFLD0SS : I<0, Pseudo, (outs FR32X:$dst), (ins), "",
                           [(set FR32X:$dst, fp32imm0)]>;
   def AVX512_FsFLD0SD : I<0, Pseudo, (outs FR64X:$dst), (ins), "",
@@ -3439,31 +3439,31 @@ let Predicates = [HasAVX512] in {
   // Move scalar to XMM zero-extended, zeroing a VR128X then do a
   // MOVS{S,D} to the lower bits.
   def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32X:$src)))),
-            (VMOVSSZrr (v4f32 (V_SET0)), FR32X:$src)>;
+            (VMOVSSZrr (v4f32 (AVX512_128_SET0)), FR32X:$src)>;
   def : Pat<(v4f32 (X86vzmovl (v4f32 VR128X:$src))),
-            (VMOVSSZrr (v4f32 (V_SET0)), (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
+            (VMOVSSZrr (v4f32 (AVX512_128_SET0)), (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
   def : Pat<(v4i32 (X86vzmovl (v4i32 VR128X:$src))),
-            (VMOVSSZrr (v4i32 (V_SET0)), (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
+            (VMOVSSZrr (v4i32 (AVX512_128_SET0)), (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
   def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64X:$src)))),
-            (VMOVSDZrr (v2f64 (V_SET0)), FR64X:$src)>;
+            (VMOVSDZrr (v2f64 (AVX512_128_SET0)), FR64X:$src)>;
   }
 
   // Move low f32 and clear high bits.
   def : Pat<(v8f32 (X86vzmovl (v8f32 VR256X:$src))),
             (SUBREG_TO_REG (i32 0),
-             (VMOVSSZrr (v4f32 (V_SET0)),
+             (VMOVSSZrr (v4f32 (AVX512_128_SET0)),
               (EXTRACT_SUBREG (v8f32 VR256X:$src), sub_xmm)), sub_xmm)>;
   def : Pat<(v8i32 (X86vzmovl (v8i32 VR256X:$src))),
             (SUBREG_TO_REG (i32 0),
-             (VMOVSSZrr (v4i32 (V_SET0)),
+             (VMOVSSZrr (v4i32 (AVX512_128_SET0)),
               (EXTRACT_SUBREG (v8i32 VR256X:$src), sub_xmm)), sub_xmm)>;
   def : Pat<(v16f32 (X86vzmovl (v16f32 VR512:$src))),
             (SUBREG_TO_REG (i32 0),
-             (VMOVSSZrr (v4f32 (V_SET0)),
+             (VMOVSSZrr (v4f32 (AVX512_128_SET0)),
               (EXTRACT_SUBREG (v16f32 VR512:$src), sub_xmm)), sub_xmm)>;
   def : Pat<(v16i32 (X86vzmovl (v16i32 VR512:$src))),
             (SUBREG_TO_REG (i32 0),
-             (VMOVSSZrr (v4i32 (V_SET0)),
+             (VMOVSSZrr (v4i32 (AVX512_128_SET0)),
               (EXTRACT_SUBREG (v16i32 VR512:$src), sub_xmm)), sub_xmm)>;
 
   let AddedComplexity = 20 in {
@@ -3525,11 +3525,11 @@ let Predicates = [HasAVX512] in {
   }
   def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
                    (v4f32 (scalar_to_vector FR32X:$src)), (iPTR 0)))),
-            (SUBREG_TO_REG (i32 0), (v4f32 (VMOVSSZrr (v4f32 (V_SET0)),
+            (SUBREG_TO_REG (i32 0), (v4f32 (VMOVSSZrr (v4f32 (AVX512_128_SET0)),
                                             FR32X:$src)), sub_xmm)>;
   def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
                    (v2f64 (scalar_to_vector FR64X:$src)), (iPTR 0)))),
-            (SUBREG_TO_REG (i64 0), (v2f64 (VMOVSDZrr (v2f64 (V_SET0)),
+            (SUBREG_TO_REG (i64 0), (v2f64 (VMOVSDZrr (v2f64 (AVX512_128_SET0)),
                                      FR64X:$src)), sub_xmm)>;
   def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
                    (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
@@ -3538,18 +3538,18 @@ let Predicates = [HasAVX512] in {
   // Move low f64 and clear high bits.
   def : Pat<(v4f64 (X86vzmovl (v4f64 VR256X:$src))),
             (SUBREG_TO_REG (i32 0),
-             (VMOVSDZrr (v2f64 (V_SET0)),
+             (VMOVSDZrr (v2f64 (AVX512_128_SET0)),
                        (EXTRACT_SUBREG (v4f64 VR256X:$src), sub_xmm)), sub_xmm)>;
   def : Pat<(v8f64 (X86vzmovl (v8f64 VR512:$src))),
             (SUBREG_TO_REG (i32 0),
-             (VMOVSDZrr (v2f64 (V_SET0)),
+             (VMOVSDZrr (v2f64 (AVX512_128_SET0)),
                        (EXTRACT_SUBREG (v8f64 VR512:$src), sub_xmm)), sub_xmm)>;
 
   def : Pat<(v4i64 (X86vzmovl (v4i64 VR256X:$src))),
-            (SUBREG_TO_REG (i32 0), (VMOVSDZrr (v2i64 (V_SET0)),
+            (SUBREG_TO_REG (i32 0), (VMOVSDZrr (v2i64 (AVX512_128_SET0)),
                        (EXTRACT_SUBREG (v4i64 VR256X:$src), sub_xmm)), sub_xmm)>;
   def : Pat<(v8i64 (X86vzmovl (v8i64 VR512:$src))),
-            (SUBREG_TO_REG (i32 0), (VMOVSDZrr (v2i64 (V_SET0)),
+            (SUBREG_TO_REG (i32 0), (VMOVSDZrr (v2i64 (AVX512_128_SET0)),
                        (EXTRACT_SUBREG (v8i64 VR512:$src), sub_xmm)), sub_xmm)>;
 
   // Extract and store.

diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index 8c5dbd51866e204b50b0b6fca91447f20ca3751e..d30a0683cf65248cb4896feaae15520e57e9a0ca 100644 (file)
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -6831,14 +6831,33 @@ bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
     assert(HasAVX && "AVX not supported");
     return Expand2AddrUndef(MIB, get(X86::VXORPSYrr));
   case X86::AVX512_128_SET0:
-    return Expand2AddrUndef(MIB, get(X86::VPXORDZ128rr));
-  case X86::AVX512_256_SET0:
-    return Expand2AddrUndef(MIB, get(X86::VPXORDZ256rr));
+  case X86::AVX512_FsFLD0SS:
+  case X86::AVX512_FsFLD0SD: {
+    bool HasVLX = Subtarget.hasVLX();
+    unsigned SrcReg = MIB->getOperand(0).getReg();
+    const TargetRegisterInfo *TRI = &getRegisterInfo();
+    if (HasVLX || TRI->getEncodingValue(SrcReg) < 16)
+      return Expand2AddrUndef(MIB,
+                              get(HasVLX ? X86::VPXORDZ128rr : X86::VXORPSrr));
+    // Extended register without VLX. Use a larger XOR.
+    SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_xmm, &X86::VR512RegClass);
+    MIB->getOperand(0).setReg(SrcReg);
+    return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
+  }
+  case X86::AVX512_256_SET0: {
+    bool HasVLX = Subtarget.hasVLX();
+    unsigned SrcReg = MIB->getOperand(0).getReg();
+    const TargetRegisterInfo *TRI = &getRegisterInfo();
+    if (HasVLX || TRI->getEncodingValue(SrcReg) < 16)
+      return Expand2AddrUndef(MIB,
+                              get(HasVLX ? X86::VPXORDZ256rr : X86::VXORPSYrr));
+    // Extended register without VLX. Use a larger XOR.
+    SrcReg = TRI->getMatchingSuperReg(SrcReg, X86::sub_ymm, &X86::VR512RegClass);
+    MIB->getOperand(0).setReg(SrcReg);
+    return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
+  }
   case X86::AVX512_512_SET0:
     return Expand2AddrUndef(MIB, get(X86::VPXORDZrr));
-  case X86::AVX512_FsFLD0SS:
-  case X86::AVX512_FsFLD0SD:
-    return Expand2AddrUndef(MIB, get(X86::VXORPSZ128rr));
   case X86::V_SETALLONES:
     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VPCMPEQDrr : X86::PCMPEQDrr));
   case X86::AVX2_SETALLONES:

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 1812d01711d167df42b495dc41d054c06b2146dd..e301d0ab7acfa67709be0c0aac8ed794a9cb62fb 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -446,9 +446,9 @@ def : Pat<(v4f64  (bitconvert (v8f32  VR256:$src))), (v4f64  VR256:$src)>;
 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
     isPseudo = 1, SchedRW = [WriteZero] in {
   def FsFLD0SS : I<0, Pseudo, (outs FR32:$dst), (ins), "",
-                   [(set FR32:$dst, fp32imm0)]>, Requires<[HasSSE1, NoVLX_Or_NoDQI]>;
+                   [(set FR32:$dst, fp32imm0)]>, Requires<[HasSSE1, NoAVX512]>;
   def FsFLD0SD : I<0, Pseudo, (outs FR64:$dst), (ins), "",
-                   [(set FR64:$dst, fpimm0)]>, Requires<[HasSSE2, NoVLX_Or_NoDQI]>;
+                   [(set FR64:$dst, fpimm0)]>, Requires<[HasSSE2, NoAVX512]>;
 }
 
 //===----------------------------------------------------------------------===//
@@ -461,12 +461,12 @@ let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
 // We set canFoldAsLoad because this can be converted to a constant-pool
 // load of an all-zeros value if folding it would be beneficial.
 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
-    isPseudo = 1, Predicates = [NoVLX], SchedRW = [WriteZero] in {
+    isPseudo = 1, SchedRW = [WriteZero] in {
 def V_SET0 : I<0, Pseudo, (outs VR128:$dst), (ins), "",
                [(set VR128:$dst, (v4f32 immAllZerosV))]>;
 }
 
-let Predicates = [NoVLX] in
+let Predicates = [NoAVX512] in
 def : Pat<(v4i32 immAllZerosV), (V_SET0)>;
 
 
@@ -475,7 +475,7 @@ def : Pat<(v4i32 immAllZerosV), (V_SET0)>;
 // at the rename stage without using any execution unit, so SET0PSY
 // and SET0PDY can be used for vector int instructions without penalty
 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
-    isPseudo = 1, Predicates = [HasAVX, NoVLX], SchedRW = [WriteZero] in {
+    isPseudo = 1, Predicates = [NoAVX512], SchedRW = [WriteZero] in {
 def AVX_SET0 : I<0, Pseudo, (outs VR256:$dst), (ins), "",
                  [(set VR256:$dst, (v8i32 immAllZerosV))]>;
 }

diff --git a/test/CodeGen/X86/vector-shuffle-256-v4.ll b/test/CodeGen/X86/vector-shuffle-256-v4.ll
index b318288e9516f6535e2e50ab74c1cbfd127a76bb..1198514be5f16b1587b79dbeca1631a96752075f 100644 (file)
--- a/test/CodeGen/X86/vector-shuffle-256-v4.ll
+++ b/test/CodeGen/X86/vector-shuffle-256-v4.ll
@@ -1229,7 +1229,7 @@ define <4 x double> @insert_reg_and_zero_v4f64(double %a) {
 ;
 ; AVX512VL-LABEL: insert_reg_and_zero_v4f64:
 ; AVX512VL:       # BB#0:
-; AVX512VL-NEXT:    vxorpd %xmm1, %xmm1, %xmm1
+; AVX512VL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
 ; AVX512VL-NEXT:    vmovsd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
 ; AVX512VL-NEXT:    retq
   %v = insertelement <4 x double> undef, double %a, i32 0