[AVX-512] Add more patterns to fold masked VPTERNLOG with load when the passthru...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295640 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index b3b8452067cef7fd3cfbc6da75bb640372a38bd2..5fda485dc8e08a278da39b1549a37681b29ac42b 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -8916,6 +8916,32 @@ def VPTERNLOG132_imm8 : SDNodeXForm<imm, [{
   if (Imm & 0x40) NewImm |= 0x20;
   return getI8Imm(NewImm, SDLoc(N));
 }]>;
+def VPTERNLOG231_imm8 : SDNodeXForm<imm, [{
+  // Convert a VPTERNLOG immediate by moving operand 1 to the end.
+  uint8_t Imm = N->getZExtValue();
+  // Move bits 1->2, 2->4, 3->6, 4->1, 5->3, 6->5
+  uint8_t NewImm = Imm & 0x81;
+  if (Imm & 0x02) NewImm |= 0x04;
+  if (Imm & 0x04) NewImm |= 0x10;
+  if (Imm & 0x08) NewImm |= 0x40;
+  if (Imm & 0x10) NewImm |= 0x02;
+  if (Imm & 0x20) NewImm |= 0x08;
+  if (Imm & 0x40) NewImm |= 0x20;
+  return getI8Imm(NewImm, SDLoc(N));
+}]>;
+def VPTERNLOG312_imm8 : SDNodeXForm<imm, [{
+  // Convert a VPTERNLOG immediate by moving operand 2 to the beginning.
+  uint8_t Imm = N->getZExtValue();
+  // Move bits 1->4, 2->1, 3->5, 4->2, 5->6, 6->3
+  uint8_t NewImm = Imm & 0x81;
+  if (Imm & 0x02) NewImm |= 0x10;
+  if (Imm & 0x04) NewImm |= 0x02;
+  if (Imm & 0x08) NewImm |= 0x20;
+  if (Imm & 0x10) NewImm |= 0x04;
+  if (Imm & 0x20) NewImm |= 0x40;
+  if (Imm & 0x40) NewImm |= 0x08;
+  return getI8Imm(NewImm, SDLoc(N));
+}]>;
 
 multiclass avx512_ternlog<bits<8> opc, string OpcodeStr, SDNode OpNode,
                           X86VectorVTInfo _>{
@@ -8992,6 +9018,30 @@ multiclass avx512_ternlog<bits<8> opc, string OpcodeStr, SDNode OpNode,
                    _.RC:$src1)),
             (!cast<Instruction>(NAME#_.ZSuffix#rmik) _.RC:$src1, _.KRCWM:$mask,
              _.RC:$src2, addr:$src3, (VPTERNLOG132_imm8 imm:$src4))>;
+  def : Pat<(_.VT (vselect _.KRCWM:$mask,
+                   (OpNode (bitconvert (_.LdFrag addr:$src3)),
+                    _.RC:$src2, _.RC:$src1, (i8 imm:$src4)),
+                   _.RC:$src1)),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmik) _.RC:$src1, _.KRCWM:$mask,
+             _.RC:$src2, addr:$src3, (VPTERNLOG321_imm8 imm:$src4))>;
+  def : Pat<(_.VT (vselect _.KRCWM:$mask,
+                   (OpNode _.RC:$src2, _.RC:$src1,
+                    (bitconvert (_.LdFrag addr:$src3)), (i8 imm:$src4)),
+                   _.RC:$src1)),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmik) _.RC:$src1, _.KRCWM:$mask,
+             _.RC:$src2, addr:$src3, (VPTERNLOG213_imm8 imm:$src4))>;
+  def : Pat<(_.VT (vselect _.KRCWM:$mask,
+                   (OpNode _.RC:$src2, (bitconvert (_.LdFrag addr:$src3)),
+                    _.RC:$src1, (i8 imm:$src4)),
+                   _.RC:$src1)),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmik) _.RC:$src1, _.KRCWM:$mask,
+             _.RC:$src2, addr:$src3, (VPTERNLOG231_imm8 imm:$src4))>;
+  def : Pat<(_.VT (vselect _.KRCWM:$mask,
+                   (OpNode (bitconvert (_.LdFrag addr:$src3)),
+                    _.RC:$src1, _.RC:$src2, (i8 imm:$src4)),
+                   _.RC:$src1)),
+            (!cast<Instruction>(NAME#_.ZSuffix#rmik) _.RC:$src1, _.KRCWM:$mask,
+             _.RC:$src2, addr:$src3, (VPTERNLOG312_imm8 imm:$src4))>;
 
   // Additional patterns for matching broadcasts in other positions.
   def : Pat<(_.VT (OpNode (X86VBroadcast (_.ScalarLdFrag addr:$src3)),

diff --git a/test/CodeGen/X86/avx512-vpternlog-commute.ll b/test/CodeGen/X86/avx512-vpternlog-commute.ll
index 684d2c9cb367721d2fefbb53e415aac4c1bc29fa..afb223a1793a373e905e275d0186107fac42bd84 100644 (file)
--- a/test/CodeGen/X86/avx512-vpternlog-commute.ll
+++ b/test/CodeGen/X86/avx512-vpternlog-commute.ll
@@ -228,9 +228,8 @@ define <16 x i32> @vpternlog_v16i32_012_load0_mask(<16 x i32>* %x0ptr, <16 x i32
 define <16 x i32> @vpternlog_v16i32_012_load0_mask1(<16 x i32>* %x0ptr, <16 x i32> %x1, <16 x i32> %x2, i16 %mask) {
 ; CHECK-LABEL: vpternlog_v16i32_012_load0_mask1:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vmovdqa32 (%rdi), %zmm2
 ; CHECK-NEXT:    kmovw %esi, %k1
-; CHECK-NEXT:    vpternlogd $9, %zmm1, %zmm2, %zmm0 {%k1}
+; CHECK-NEXT:    vpternlogd $65, (%rdi), %zmm1, %zmm0 {%k1}
 ; CHECK-NEXT:    retq
   %x0 = load <16 x i32>, <16 x i32>* %x0ptr
   %res = call <16 x i32> @llvm.x86.avx512.mask.pternlog.d.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i32 33, i16 -1)
@@ -242,9 +241,8 @@ define <16 x i32> @vpternlog_v16i32_012_load0_mask1(<16 x i32>* %x0ptr, <16 x i3
 define <16 x i32> @vpternlog_v16i32_012_load0_mask2(<16 x i32>* %x0ptr, <16 x i32> %x1, <16 x i32> %x2, i16 %mask) {
 ; CHECK-LABEL: vpternlog_v16i32_012_load0_mask2:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vmovdqa32 (%rdi), %zmm2
 ; CHECK-NEXT:    kmovw %esi, %k1
-; CHECK-NEXT:    vpternlogd $33, %zmm2, %zmm0, %zmm1 {%k1}
+; CHECK-NEXT:    vpternlogd $33, (%rdi), %zmm0, %zmm1 {%k1}
 ; CHECK-NEXT:    vmovdqa64 %zmm1, %zmm0
 ; CHECK-NEXT:    retq
   %x0 = load <16 x i32>, <16 x i32>* %x0ptr
@@ -268,9 +266,8 @@ define <16 x i32> @vpternlog_v16i32_012_load1_mask(<16 x i32> %x0, <16 x i32>* %
 define <16 x i32> @vpternlog_v16i32_012_load1_mask2(<16 x i32> %x0, <16 x i32>* %x1ptr, <16 x i32> %x2, i16 %mask) {
 ; CHECK-LABEL: vpternlog_v16i32_012_load1_mask2:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vmovdqa32 (%rdi), %zmm2
 ; CHECK-NEXT:    kmovw %esi, %k1
-; CHECK-NEXT:    vpternlogd $33, %zmm0, %zmm2, %zmm1 {%k1}
+; CHECK-NEXT:    vpternlogd $9, (%rdi), %zmm0, %zmm1 {%k1}
 ; CHECK-NEXT:    vmovdqa64 %zmm1, %zmm0
 ; CHECK-NEXT:    retq
   %x1 = load <16 x i32>, <16 x i32>* %x1ptr
@@ -294,9 +291,8 @@ define <16 x i32> @vpternlog_v16i32_012_load2_mask(<16 x i32> %x0, <16 x i32> %x
 define <16 x i32> @vpternlog_v16i32_012_load2_mask1(<16 x i32> %x0, <16 x i32> %x1, <16 x i32>* %x2ptr, i16 %mask) {
 ; CHECK-LABEL: vpternlog_v16i32_012_load2_mask1:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vmovdqa32 (%rdi), %zmm2
 ; CHECK-NEXT:    kmovw %esi, %k1
-; CHECK-NEXT:    vpternlogd $9, %zmm2, %zmm0, %zmm1 {%k1}
+; CHECK-NEXT:    vpternlogd $9, (%rdi), %zmm0, %zmm1 {%k1}
 ; CHECK-NEXT:    vmovdqa64 %zmm1, %zmm0
 ; CHECK-NEXT:    retq
   %x2 = load <16 x i32>, <16 x i32>* %x2ptr