[llvm] [VectorCombine] foldShuffleOfBinops - extend to handle icmp/fcmp ops as well (PR #120075)

[Alexey Bataev via llvm-commits]

================
@@ -27,13 +43,37 @@ define <4 x i32> @shuf_icmp_eq_v4i32(<4 x i32> %x, <4 x i32> %y, <4 x i32> %z, <
 ; icmp - eq v2i64 is only cheap on SSE4+ targets with PCMPEQQ
 
 define <2 x i64> @shuf_icmp_eq_v2i64(<2 x i64> %x, <2 x i64> %y, <2 x i64> %z, <2 x i64> %w) {
-; CHECK-LABEL: define <2 x i64> @shuf_icmp_eq_v2i64(
-; CHECK-SAME: <2 x i64> [[X:%.*]], <2 x i64> [[Y:%.*]], <2 x i64> [[Z:%.*]], <2 x i64> [[W:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:    [[C0:%.*]] = icmp eq <2 x i64> [[X]], [[Y]]
-; CHECK-NEXT:    [[C1:%.*]] = icmp eq <2 x i64> [[Z]], [[W]]
-; CHECK-NEXT:    [[S:%.*]] = shufflevector <2 x i1> [[C0]], <2 x i1> [[C1]], <2 x i32> <i32 1, i32 3>
-; CHECK-NEXT:    [[R:%.*]] = sext <2 x i1> [[S]] to <2 x i64>
-; CHECK-NEXT:    ret <2 x i64> [[R]]
+; SSE2-LABEL: define <2 x i64> @shuf_icmp_eq_v2i64(
+; SSE2-SAME: <2 x i64> [[X:%.*]], <2 x i64> [[Y:%.*]], <2 x i64> [[Z:%.*]], <2 x i64> [[W:%.*]]) #[[ATTR0]] {
+; SSE2-NEXT:    [[TMP1:%.*]] = shufflevector <2 x i64> [[X]], <2 x i64> [[Z]], <2 x i32> <i32 1, i32 3>
+; SSE2-NEXT:    [[TMP2:%.*]] = shufflevector <2 x i64> [[Y]], <2 x i64> [[W]], <2 x i32> <i32 1, i32 3>
+; SSE2-NEXT:    [[S:%.*]] = icmp eq <2 x i64> [[TMP1]], [[TMP2]]
+; SSE2-NEXT:    [[R:%.*]] = sext <2 x i1> [[S]] to <2 x i64>
+; SSE2-NEXT:    ret <2 x i64> [[R]]
----------------
alexey-bataev wrote:

Is this really cheaper?

https://github.com/llvm/llvm-project/pull/120075