Fix "Cannot select" crash on bitcast between f64 and int vector types

llvm/lib/Target/Hexagon/HexagonPatterns.td

@@ -530,17 +530,22 @@ def: Pat<(i64 (bitconvert F64:$v)), (I64:$v)>;
 def: Pat<(f64 (bitconvert I64:$v)), (F64:$v)>;
 
 // Bit convert 32- and 64-bit types.
-// All of these are bitcastable to one another: i32, v2i16, v4i8.
+// All of these are bitcastable to one another: i32, f32, v2i16, v4i8.
 defm: NopCast_pat<i32,   v2i16, IntRegs>;
 defm: NopCast_pat<i32,    v4i8, IntRegs>;
 defm: NopCast_pat<v2i16,  v4i8, IntRegs>;
-// All of these are bitcastable to one another: i64, v2i32, v4i16, v8i8.
+defm: NopCast_pat<f32,   v2i16, IntRegs>;
+defm: NopCast_pat<f32,    v4i8, IntRegs>;
+// All of these are bitcastable to one another: i64, f64, v2i32, v4i16, v8i8.
 defm: NopCast_pat<i64,   v2i32, DoubleRegs>;
 defm: NopCast_pat<i64,   v4i16, DoubleRegs>;
 defm: NopCast_pat<i64,    v8i8, DoubleRegs>;
 defm: NopCast_pat<v2i32, v4i16, DoubleRegs>;
 defm: NopCast_pat<v2i32,  v8i8, DoubleRegs>;
 defm: NopCast_pat<v4i16,  v8i8, DoubleRegs>;
+defm: NopCast_pat<f64,   v2i32, DoubleRegs>;
+defm: NopCast_pat<f64,   v4i16, DoubleRegs>;
+defm: NopCast_pat<f64,    v8i8, DoubleRegs>;
 
 
 // --(3) Extend/truncate/saturate ----------------------------------------

llvm/test/CodeGen/Hexagon/bitcast-f64-vector.ll

@@ -0,0 +1,63 @@
+; RUN: llc -mtriple=hexagon -mcpu=hexagonv68 -O3 < %s | FileCheck %s
+
+; Verify that bitcasts between f64 and the integer vector types that share the
+; DoubleRegs register class (v2i32, v4i16, v8i8) are treated as no-ops by the
+; instruction selector and do not cause a "Cannot select" crash.
+;
+; All of i64, f64, v2i32, v4i16, v8i8 live in DoubleRegs (64-bit register
+; pairs). A bitcast between any two of them is a pure reinterpretation of the
+; same 64 bits. Therefore no instruction is emitted.
+;
+; Regression test for: llvm.org/PR195495
+;   llc -mtriple=hexagon -mcpu=hexagonv68 -O3 crashed with
+;   "Cannot select: f64 = bitcast v2i32" when compiling Eigen's packetmath.
+
+; CHECK-LABEL: test_v2i32_to_f64:
+; CHECK: dfcmp
+; CHECK: jumpr r31
+define i1 @test_v2i32_to_f64(<2 x i32> %a) {
+  %bc = bitcast <2 x i32> %a to double
+  %cmp = fcmp une double %bc, 0.0
+  ret i1 %cmp
+}
+
+; f64->v2i32 is a no-op: the argument is already in a DoubleReg pair.
+; CHECK-LABEL: test_f64_to_v2i32:
+; CHECK-NOT: combine
+; CHECK: jumpr r31
+define <2 x i32> @test_f64_to_v2i32(double %a) {
+  %bc = bitcast double %a to <2 x i32>
+  ret <2 x i32> %bc
+}
+
+; CHECK-LABEL: test_v4i16_to_f64:
+; CHECK: dfcmp
+; CHECK: jumpr r31
+define i1 @test_v4i16_to_f64(<4 x i16> %a) {
+  %bc = bitcast <4 x i16> %a to double
+  %cmp = fcmp une double %bc, 0.0
+  ret i1 %cmp
+}
+
+; CHECK-LABEL: test_v8i8_to_f64:
+; CHECK: dfcmp
+; CHECK: jumpr r31
+define i1 @test_v8i8_to_f64(<8 x i8> %a) {
+  %bc = bitcast <8 x i8> %a to double
+  %cmp = fcmp une double %bc, 0.0
+  ret i1 %cmp
+}
+
+; Regression test: the original crash.
+; <4 x i32> is passed in two v2i32 DoubleReg pairs; after type-legalizing
+; <2 x f64> setcc into two scalar f64 setcc ops, each f64 operand is produced
+; by a "f64 = bitcast v2i32" node that previously had no matching pattern.
+; CHECK-LABEL: test_packetmath_reduced:
+; CHECK: dfcmp
+; CHECK: jumpr r31
+define <2 x i1> @test_packetmath_reduced(<4 x i32> %arg) {
+entry:
+  %bc = bitcast <4 x i32> %arg to <2 x double>
+  %cmp = fcmp une <2 x double> %bc, zeroinitializer
+  ret <2 x i1> %cmp
+}