[PowerPC] Exploit xxeval instruction for operations of the form ternary(A,X,B) and ternary(A,X,C). (#152956)

tonykuttai · Tony Varghese · web-flow · commit 2e7ea9c94504 · 2025-09-01T10:13:54.000+05:30
Adds support for ternary equivalent operations of the form `ternary(A, X, B)` and `ternary(A, X, C)` where `X=[and(B,C)| nor(B,C)| eqv(B,C)| nand(B,C)]`. The following are the patterns involved and the imm values: | **Operation** | **Immediate Value** | |----------------------------|---------------------| | ternary(A, and(B,C), B) | 49 | | ternary(A, nor(B,C), B) | 56 | | ternary(A, eqv(B,C), B) | 57 | | ternary(A, nand(B,C), B) | 62 | | | | | ternary(A, and(B,C), C) | 81 | | ternary(A, nor(B,C), C) | 88 | | ternary(A, eqv(B,C), C) | 89 | | ternary(A, nand(B,C), C) | 94 | eg. `xxeval XT, XA, XB, XC, 49` - performs `XA ? and(XB, XC) : B`and places the result in `XT`. This is the continuation of [[PowerPC] Exploit xxeval instruction for ternary patterns - ternary(A, X, and(B,C))](#141733 (comment)). --------- Co-authored-by: Tony Varghese <tony.varghese@ibm.com>
diff --git a/llvm/lib/Target/PowerPC/PPCInstrP10.td b/llvm/lib/Target/PowerPC/PPCInstrP10.td
@@ -2230,6 +2230,13 @@ def VEqv
                                             (v4i32(bitconvert node:$a)),
                                             (v4i32(bitconvert node:$b)))))]>;
 
+// Vector NAND operation (not(and))
+def VNand
+    : PatFrags<(ops node:$a, node:$b), [(vnot(and node:$a, node:$b)),
+                                        (bitconvert(vnot(and
+                                            (v4i32(bitconvert node:$a)),
+                                            (v4i32(bitconvert node:$b)))))]>;
+
 // =============================================================================
 // XXEVAL Ternary Pattern Multiclass: XXEvalTernarySelectAnd
 // This class matches the equivalent Ternary Operation: A ? f(B,C) : AND(B,C)
@@ -2265,6 +2272,56 @@ multiclass XXEvalTernarySelectAnd<ValueType Vt> {
             Vt, (vselect Vt:$vA, (VNot Vt:$vB), (VAnd Vt:$vB, Vt:$vC)), 28>;
 }
 
+// =============================================================================
+// XXEVAL Ternary Pattern Multiclass: XXEvalTernarySelectB
+// This class matches the equivalent Ternary Operation: A ? f(B,C) : B
+// and emit the corresponding xxeval instruction with the imm value.
+//
+// The patterns implement xxeval vector select operations where:
+// - A is the selector vector
+// - f(B,C) is the "true" case op on vectors B and C (AND, NOR, EQV, NAND)
+// - B is the "false" case operand (vector B)
+//
+// Note: Patterns (A? C : B) and (A? not(C) : B) are not considered
+// for XXEVAL instruction (4 Cycle) as XXSEL (3 cycle) instruction performs 
+// better.
+// =============================================================================
+multiclass XXEvalTernarySelectB<ValueType Vt>{
+  // Pattern: (A ? AND(B,C) : B) XXEVAL immediate value: 49
+  def : XXEvalPattern<Vt, (vselect Vt:$vA, (VAnd Vt:$vB, Vt:$vC), Vt:$vB), 49>;
+  // Pattern: (A ? NOR(B,C) : B) XXEVAL immediate value: 56
+  def : XXEvalPattern<Vt, (vselect Vt:$vA, (VNor Vt:$vB, Vt:$vC), Vt:$vB), 56>;
+  // Pattern: (A ? EQV(B,C) : B) XXEVAL immediate value: 57
+  def : XXEvalPattern<Vt, (vselect Vt:$vA, (VEqv Vt:$vB, Vt:$vC), Vt:$vB), 57>;
+  // Pattern: (A ? NAND(B,C) : B) XXEVAL immediate value: 62
+  def : XXEvalPattern<Vt, (vselect Vt:$vA, (VNand Vt:$vB, Vt:$vC), Vt:$vB), 62>;
+}
+
+// =============================================================================
+// XXEVAL Ternary Pattern Multiclass: XXEvalTernarySelectC
+// This class matches the equivalent Ternary Operation: A ? f(B,C) : C
+// and emit the corresponding xxeval instruction with the imm value.
+//
+// The patterns implement xxeval vector select operations where:
+// - A is the selector vector
+// - f(B,C) is the "true" case op on vectors B and C (AND, NOR, EQV, NAND)
+// - C is the "false" case operand (vector C)
+//
+// Note: Patterns (A? B : C) and (A? not(B) : C) are not considered
+// for XXEVAL instruction (4 Cycle) as XXSEL (3 cycle) instruction performs 
+// better.
+// =============================================================================
+multiclass XXEvalTernarySelectC<ValueType Vt>{
+  // Pattern: (A ? AND(B,C) : C) XXEVAL immediate value: 81
+  def : XXEvalPattern<Vt, (vselect Vt:$vA, (VAnd Vt:$vB, Vt:$vC), Vt:$vC), 81>;
+  // Pattern: (A ? NOR(B,C) : C) XXEVAL immediate value: 88
+  def : XXEvalPattern<Vt, (vselect Vt:$vA, (VNor Vt:$vB, Vt:$vC), Vt:$vC), 88>;
+  // Pattern: (A ? EQV(B,C) : C) XXEVAL immediate value: 89
+  def : XXEvalPattern<Vt, (vselect Vt:$vA, (VEqv Vt:$vB, Vt:$vC), Vt:$vC), 89>;
+  // Pattern: (A ? NAND(B,C) : C) XXEVAL immediate value: 94
+  def : XXEvalPattern<Vt, (vselect Vt:$vA, (VNand Vt:$vB, Vt:$vC), Vt:$vC), 94>;
+}
+
 let Predicates = [PrefixInstrs, HasP10Vector] in {
   let AddedComplexity = 400 in {
     def : Pat<(v4i32 (build_vector i32immNonAllOneNonZero:$A,
@@ -2376,6 +2433,8 @@ let Predicates = [PrefixInstrs, HasP10Vector] in {
     // XXEval Patterns for ternary Operations.
     foreach Ty = [v4i32, v2i64, v8i16, v16i8] in {
         defm : XXEvalTernarySelectAnd<Ty>;
+        defm : XXEvalTernarySelectB<Ty>;
+        defm : XXEvalTernarySelectC<Ty>; 
     }
 
     // Anonymous patterns to select prefixed VSX loads and stores.
diff --git a/llvm/test/CodeGen/PowerPC/xxeval-vselect-x-b.ll b/llvm/test/CodeGen/PowerPC/xxeval-vselect-x-b.ll
@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-; Test file to verify the emission of Vector Selection instructions when ternary operators are used.
+; Test file to verify the emission of Vector Evaluation instructions when ternary operators are used.
 
 ; RUN: llc -verify-machineinstrs -mcpu=pwr10 -mtriple=powerpc64le-unknown-unknown \
 ; RUN:   -ppc-asm-full-reg-names --ppc-vsr-nums-as-vr < %s | FileCheck %s
@@ -15,10 +15,9 @@ define <4 x i32> @ternary_A_and_BC_B_4x32(<4 x i1> %A, <4 x i32> %B, <4 x i32> %
 ; CHECK-LABEL: ternary_A_and_BC_B_4x32:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxleqv v5, v5, v5
-; CHECK-NEXT:    xxland vs0, v3, v4
 ; CHECK-NEXT:    vslw v2, v2, v5
 ; CHECK-NEXT:    vsraw v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 49
 ; CHECK-NEXT:    blr
 entry:
   %and = and <4 x i32> %B, %C
@@ -31,11 +30,10 @@ define <2 x i64> @ternary_A_and_BC_B_2x64(<2 x i1> %A, <2 x i64> %B, <2 x i64> %
 ; CHECK-LABEL: ternary_A_and_BC_B_2x64:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxlxor v5, v5, v5
-; CHECK-NEXT:    xxland vs0, v3, v4
 ; CHECK-NEXT:    xxsplti32dx v5, 1, 63
 ; CHECK-NEXT:    vsld v2, v2, v5
 ; CHECK-NEXT:    vsrad v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 49
 ; CHECK-NEXT:    blr
 entry:
   %and = and <2 x i64> %B, %C
@@ -48,10 +46,9 @@ define <16 x i8> @ternary_A_and_BC_B_16x8(<16 x i1> %A, <16 x i8> %B, <16 x i8>
 ; CHECK-LABEL: ternary_A_and_BC_B_16x8:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxspltib v5, 7
-; CHECK-NEXT:    xxland vs0, v3, v4
 ; CHECK-NEXT:    vslb v2, v2, v5
 ; CHECK-NEXT:    vsrab v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 49
 ; CHECK-NEXT:    blr
 entry:
   %and = and <16 x i8> %B, %C
@@ -64,10 +61,9 @@ define <8 x i16> @ternary_A_and_BC_B_8x16(<8 x i1> %A, <8 x i16> %B, <8 x i16> %
 ; CHECK-LABEL: ternary_A_and_BC_B_8x16:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxspltiw v5, 983055
-; CHECK-NEXT:    xxland vs0, v3, v4
 ; CHECK-NEXT:    vslh v2, v2, v5
 ; CHECK-NEXT:    vsrah v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 49
 ; CHECK-NEXT:    blr
 entry:
   %and = and <8 x i16> %B, %C
@@ -80,10 +76,9 @@ define <4 x i32> @ternary_A_nor_BC_B_4x32(<4 x i1> %A, <4 x i32> %B, <4 x i32> %
 ; CHECK-LABEL: ternary_A_nor_BC_B_4x32:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxleqv v5, v5, v5
-; CHECK-NEXT:    xxlnor vs0, v3, v4
 ; CHECK-NEXT:    vslw v2, v2, v5
 ; CHECK-NEXT:    vsraw v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 56
 ; CHECK-NEXT:    blr
 entry:
   %or = or <4 x i32> %B, %C
@@ -97,11 +92,10 @@ define <2 x i64> @ternary_A_nor_BC_B_2x64(<2 x i1> %A, <2 x i64> %B, <2 x i64> %
 ; CHECK-LABEL: ternary_A_nor_BC_B_2x64:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxlxor v5, v5, v5
-; CHECK-NEXT:    xxlnor vs0, v3, v4
 ; CHECK-NEXT:    xxsplti32dx v5, 1, 63
 ; CHECK-NEXT:    vsld v2, v2, v5
 ; CHECK-NEXT:    vsrad v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 56
 ; CHECK-NEXT:    blr
 entry:
   %or = or <2 x i64> %B, %C
@@ -115,10 +109,9 @@ define <16 x i8> @ternary_A_nor_BC_B_16x8(<16 x i1> %A, <16 x i8> %B, <16 x i8>
 ; CHECK-LABEL: ternary_A_nor_BC_B_16x8:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxspltib v5, 7
-; CHECK-NEXT:    xxlnor vs0, v3, v4
 ; CHECK-NEXT:    vslb v2, v2, v5
 ; CHECK-NEXT:    vsrab v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 56
 ; CHECK-NEXT:    blr
 entry:
   %or = or <16 x i8> %B, %C
@@ -132,10 +125,9 @@ define <8 x i16> @ternary_A_nor_BC_B_8x16(<8 x i1> %A, <8 x i16> %B, <8 x i16> %
 ; CHECK-LABEL: ternary_A_nor_BC_B_8x16:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxspltiw v5, 983055
-; CHECK-NEXT:    xxlnor vs0, v3, v4
 ; CHECK-NEXT:    vslh v2, v2, v5
 ; CHECK-NEXT:    vsrah v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 56
 ; CHECK-NEXT:    blr
 entry:
   %or = or <8 x i16> %B, %C
@@ -149,10 +141,9 @@ define <4 x i32> @ternary_A_eqv_BC_B_4x32(<4 x i1> %A, <4 x i32> %B, <4 x i32> %
 ; CHECK-LABEL: ternary_A_eqv_BC_B_4x32:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxleqv v5, v5, v5
-; CHECK-NEXT:    xxleqv vs0, v3, v4
 ; CHECK-NEXT:    vslw v2, v2, v5
 ; CHECK-NEXT:    vsraw v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 57
 ; CHECK-NEXT:    blr
 entry:
   %xor = xor <4 x i32> %B, %C
@@ -166,11 +157,10 @@ define <2 x i64> @ternary_A_eqv_BC_B_2x64(<2 x i1> %A, <2 x i64> %B, <2 x i64> %
 ; CHECK-LABEL: ternary_A_eqv_BC_B_2x64:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxlxor v5, v5, v5
-; CHECK-NEXT:    xxleqv vs0, v3, v4
 ; CHECK-NEXT:    xxsplti32dx v5, 1, 63
 ; CHECK-NEXT:    vsld v2, v2, v5
 ; CHECK-NEXT:    vsrad v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 57
 ; CHECK-NEXT:    blr
 entry:
   %xor = xor <2 x i64> %B, %C
@@ -184,10 +174,9 @@ define <16 x i8> @ternary_A_eqv_BC_B_16x8(<16 x i1> %A, <16 x i8> %B, <16 x i8>
 ; CHECK-LABEL: ternary_A_eqv_BC_B_16x8:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxspltib v5, 7
-; CHECK-NEXT:    xxleqv vs0, v3, v4
 ; CHECK-NEXT:    vslb v2, v2, v5
 ; CHECK-NEXT:    vsrab v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 57
 ; CHECK-NEXT:    blr
 entry:
   %xor = xor <16 x i8> %B, %C
@@ -201,10 +190,9 @@ define <8 x i16> @ternary_A_eqv_BC_B_8x16(<8 x i1> %A, <8 x i16> %B, <8 x i16> %
 ; CHECK-LABEL: ternary_A_eqv_BC_B_8x16:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxspltiw v5, 983055
-; CHECK-NEXT:    xxleqv vs0, v3, v4
 ; CHECK-NEXT:    vslh v2, v2, v5
 ; CHECK-NEXT:    vsrah v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 57
 ; CHECK-NEXT:    blr
 entry:
   %xor = xor <8 x i16> %B, %C
@@ -218,10 +206,9 @@ define <4 x i32> @ternary_A_nand_BC_B_4x32(<4 x i1> %A, <4 x i32> %B, <4 x i32>
 ; CHECK-LABEL: ternary_A_nand_BC_B_4x32:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxleqv v5, v5, v5
-; CHECK-NEXT:    xxlnand vs0, v3, v4
 ; CHECK-NEXT:    vslw v2, v2, v5
 ; CHECK-NEXT:    vsraw v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 62
 ; CHECK-NEXT:    blr
 entry:
   %and = and <4 x i32> %B, %C
@@ -235,11 +222,10 @@ define <2 x i64> @ternary_A_nand_BC_B_2x64(<2 x i1> %A, <2 x i64> %B, <2 x i64>
 ; CHECK-LABEL: ternary_A_nand_BC_B_2x64:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxlxor v5, v5, v5
-; CHECK-NEXT:    xxlnand vs0, v3, v4
 ; CHECK-NEXT:    xxsplti32dx v5, 1, 63
 ; CHECK-NEXT:    vsld v2, v2, v5
 ; CHECK-NEXT:    vsrad v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 62
 ; CHECK-NEXT:    blr
 entry:
   %and = and <2 x i64> %B, %C
@@ -253,10 +239,9 @@ define <16 x i8> @ternary_A_nand_BC_B_16x8(<16 x i1> %A, <16 x i8> %B, <16 x i8>
 ; CHECK-LABEL: ternary_A_nand_BC_B_16x8:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxspltib v5, 7
-; CHECK-NEXT:    xxlnand vs0, v3, v4
 ; CHECK-NEXT:    vslb v2, v2, v5
 ; CHECK-NEXT:    vsrab v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 62
 ; CHECK-NEXT:    blr
 entry:
   %and = and <16 x i8> %B, %C
@@ -270,10 +255,9 @@ define <8 x i16> @ternary_A_nand_BC_B_8x16(<8 x i1> %A, <8 x i16> %B, <8 x i16>
 ; CHECK-LABEL: ternary_A_nand_BC_B_8x16:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xxspltiw v5, 983055
-; CHECK-NEXT:    xxlnand vs0, v3, v4
 ; CHECK-NEXT:    vslh v2, v2, v5
 ; CHECK-NEXT:    vsrah v2, v2, v5
-; CHECK-NEXT:    xxsel v2, v3, vs0, v2
+; CHECK-NEXT:    xxeval v2, v2, v3, v4, 62
 ; CHECK-NEXT:    blr
 entry:
   %and = and <8 x i16> %B, %C
diff --git a/llvm/test/CodeGen/PowerPC/xxeval-vselect-x-c.ll b/llvm/test/CodeGen/PowerPC/xxeval-vselect-x-c.ll
