[X86] Change the masked FPCLASS implementation to use AND instead of OR to combine...

While the description for the instruction does mention OR, its talking about how the individual classification test results are ORed together.

The incoming mask is used as a zeroing write mask. If the bit is 1 the classification is written to the output. The bit is 0 the output is 0. This equivalent to an AND.

Here is pseudocode from the intrinsics guide

FOR j := 0 to 1
        i := j*64
        IF k1[j]
                k[j] := CheckFPClass_FP64(a[i+63:i], imm8[7:0])
        ELSE
                k[j] := 0
        FI
ENDFOR
k[MAX:2] := 0

llvm-svn: 326306

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 5449ef7..9225881 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -19911,9 +19911,8 @@ static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
   case X86ISD::CMPM_RND:
   case X86ISD::CMPMU:
   case X86ISD::VPSHUFBITQMB:
-    return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
   case X86ISD::VFPCLASS:
-    return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
+    return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
   case ISD::TRUNCATE:
   case X86ISD::VTRUNC:
   case X86ISD::VTRUNCS:
@@ -19951,10 +19950,9 @@ static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
 
   SDValue IMask = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i1, Mask);
   if (Op.getOpcode() == X86ISD::FSETCCM ||
-      Op.getOpcode() == X86ISD::FSETCCM_RND)
+      Op.getOpcode() == X86ISD::FSETCCM_RND ||
+      Op.getOpcode() == X86ISD::VFPCLASSS)
     return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
-  if (Op.getOpcode() == X86ISD::VFPCLASSS)
-    return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
 
   if (PreservedSrc.isUndef())
     PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 9fc9a7d..e8ebbb3 100644 (file)
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -2601,7 +2601,7 @@ multiclass avx512_scalar_fpclass<bits<8> opc, string OpcodeStr, SDNode OpNode,
                       (ins _.KRCWM:$mask, _.RC:$src1, i32u8imm:$src2),
                       OpcodeStr##_.Suffix#
                       "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}",
-                      [(set _.KRC:$dst,(or _.KRCWM:$mask,
+                      [(set _.KRC:$dst,(and _.KRCWM:$mask,
                                       (OpNode (_.VT _.RC:$src1),
                                       (i32 imm:$src2))))], itins.rr>,
                       EVEX_K, Sched<[itins.Sched]>;
@@ -2617,7 +2617,7 @@ multiclass avx512_scalar_fpclass<bits<8> opc, string OpcodeStr, SDNode OpNode,
                     (ins _.KRCWM:$mask, _.IntScalarMemOp:$src1, i32u8imm:$src2),
                     OpcodeStr##_.Suffix##
                     "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}",
-                    [(set _.KRC:$dst,(or _.KRCWM:$mask,
+                    [(set _.KRC:$dst,(and _.KRCWM:$mask,
                         (OpNode _.ScalarIntMemCPat:$src1,
                             (i32 imm:$src2))))], itins.rm>,
                     EVEX_K, Sched<[itins.Sched.Folded, ReadAfterLd]>;
@@ -2641,7 +2641,7 @@ multiclass avx512_vector_fpclass<bits<8> opc, string OpcodeStr, SDNode OpNode,
                       (ins _.KRCWM:$mask, _.RC:$src1, i32u8imm:$src2),
                       OpcodeStr##_.Suffix#
                       "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}",
-                      [(set _.KRC:$dst,(or _.KRCWM:$mask,
+                      [(set _.KRC:$dst,(and _.KRCWM:$mask,
                                        (OpNode (_.VT _.RC:$src1),
                                        (i32 imm:$src2))))], itins.rr>,
                       EVEX_K, Sched<[itins.Sched]>;
@@ -2657,7 +2657,7 @@ multiclass avx512_vector_fpclass<bits<8> opc, string OpcodeStr, SDNode OpNode,
                     (ins _.KRCWM:$mask, _.MemOp:$src1, i32u8imm:$src2),
                     OpcodeStr##_.Suffix##mem#
                     "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}",
-                    [(set _.KRC:$dst, (or _.KRCWM:$mask, (OpNode
+                    [(set _.KRC:$dst, (and _.KRCWM:$mask, (OpNode
                                   (_.VT (bitconvert (_.LdFrag addr:$src1))),
                                   (i32 imm:$src2))))], itins.rm>,
                     EVEX_K, Sched<[itins.Sched.Folded, ReadAfterLd]>;
@@ -2676,7 +2676,7 @@ multiclass avx512_vector_fpclass<bits<8> opc, string OpcodeStr, SDNode OpNode,
                     OpcodeStr##_.Suffix##broadcast##"\t{$src2, ${src1}"##
                           _.BroadcastStr##", $dst {${mask}}|$dst {${mask}}, ${src1}"##
                                                    _.BroadcastStr##", $src2}",
-                    [(set _.KRC:$dst,(or _.KRCWM:$mask, (OpNode
+                    [(set _.KRC:$dst,(and _.KRCWM:$mask, (OpNode
                                      (_.VT (X86VBroadcast
                                            (_.ScalarLdFrag addr:$src1))),
                                      (i32 imm:$src2))))], itins.rm>,