[X86] Add test cases for missed vector rotate matching due to SimplifyDemandedBits...

As requested in D47116

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@332869 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/test/CodeGen/X86/combine-rotates.ll b/test/CodeGen/X86/combine-rotates.ll
index e75973a..77371ee 100644 (file)
--- a/test/CodeGen/X86/combine-rotates.ll
+++ b/test/CodeGen/X86/combine-rotates.ll
@@ -57,3 +57,117 @@ define <4 x i32> @combine_vec_rot_rot_splat_zero(<4 x i32> %x) {
   %6 = or <4 x i32> %4, %5
   ret <4 x i32> %6
 }
+
+define <4 x i32> @rotate_demanded_bits(<4 x i32>, <4 x i32>) {
+; CHECK-LABEL: rotate_demanded_bits:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    andb $30, %sil
+; CHECK-NEXT:    movl %esi, %ecx
+; CHECK-NEXT:    roll %cl, %edi
+; CHECK-NEXT:    movl %edi, %eax
+; CHECK-NEXT:    retq
+; XOP-LABEL: rotate_demanded_bits:
+; XOP:       # %bb.0:
+; XOP-NEXT:    vmovdqa {{.*#+}} xmm2 = [30,30,30,30]
+; XOP-NEXT:    vpand %xmm2, %xmm1, %xmm1
+; XOP-NEXT:    vpshld %xmm1, %xmm0, %xmm3
+; XOP-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; XOP-NEXT:    vpsubd %xmm1, %xmm4, %xmm1
+; XOP-NEXT:    vpand %xmm2, %xmm1, %xmm1
+; XOP-NEXT:    vpsubd %xmm1, %xmm4, %xmm1
+; XOP-NEXT:    vpshld %xmm1, %xmm0, %xmm0
+; XOP-NEXT:    vpor %xmm3, %xmm0, %xmm0
+; XOP-NEXT:    retq
+;
+; AVX512-LABEL: rotate_demanded_bits:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [30,30,30,30]
+; AVX512-NEXT:    vpand %xmm2, %xmm1, %xmm1
+; AVX512-NEXT:    vpsllvd %xmm1, %xmm0, %xmm3
+; AVX512-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; AVX512-NEXT:    vpsubd %xmm1, %xmm4, %xmm1
+; AVX512-NEXT:    vpand %xmm2, %xmm1, %xmm1
+; AVX512-NEXT:    vpsrlvd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpor %xmm3, %xmm0, %xmm0
+; AVX512-NEXT:    retq
+  %3 = and <4 x i32> %1, <i32 30, i32 30, i32 30, i32 30>
+  %4 = shl <4 x i32> %0, %3
+  %5 = sub nsw <4 x i32> zeroinitializer, %3
+  %6 = and <4 x i32> %5, <i32 30, i32 30, i32 30, i32 30>
+  %7 = lshr <4 x i32> %0, %6
+  %8 = or <4 x i32> %7, %4
+  ret <4 x i32> %8
+}
+
+define <4 x i32> @rotate_demanded_bits_2(<4 x i32>, <4 x i32>) {
+; CHECK-LABEL: rotate_demanded_bits_2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    andb $23, %sil
+; CHECK-NEXT:    movl %esi, %ecx
+; CHECK-NEXT:    roll %cl, %edi
+; CHECK-NEXT:    movl %edi, %eax
+; CHECK-NEXT:    retq
+; XOP-LABEL: rotate_demanded_bits_2:
+; XOP:       # %bb.0:
+; XOP-NEXT:    vpand {{.*}}(%rip), %xmm1, %xmm1
+; XOP-NEXT:    vprotd %xmm1, %xmm0, %xmm0
+; XOP-NEXT:    retq
+;
+; AVX512-LABEL: rotate_demanded_bits_2:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpandd {{.*}}(%rip){1to4}, %xmm1, %xmm1
+; AVX512-NEXT:    vprolvd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    retq
+  %3 = and <4 x i32> %1, <i32 23, i32 23, i32 23, i32 23>
+  %4 = shl <4 x i32> %0, %3
+  %5 = sub nsw <4 x i32> zeroinitializer, %3
+  %6 = and <4 x i32> %5, <i32 31, i32 31, i32 31, i32 31>
+  %7 = lshr <4 x i32> %0, %6
+  %8 = or <4 x i32> %7, %4
+  ret <4 x i32> %8
+}
+
+define <4 x i32> @rotate_demanded_bits_3(<4 x i32>, <4 x i32>) {
+; CHECK-LABEL: rotate_demanded_bits_3:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    addb %sil, %sil
+; CHECK-NEXT:    andb $30, %sil
+; CHECK-NEXT:    movl %esi, %ecx
+; CHECK-NEXT:    roll %cl, %edi
+; CHECK-NEXT:    movl %edi, %eax
+; CHECK-NEXT:    retq
+; XOP-LABEL: rotate_demanded_bits_3:
+; XOP:       # %bb.0:
+; XOP-NEXT:    vpaddd %xmm1, %xmm1, %xmm1
+; XOP-NEXT:    vmovdqa {{.*#+}} xmm2 = [30,30,30,30]
+; XOP-NEXT:    vpand %xmm2, %xmm1, %xmm3
+; XOP-NEXT:    vpshld %xmm3, %xmm0, %xmm3
+; XOP-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; XOP-NEXT:    vpsubd %xmm1, %xmm4, %xmm1
+; XOP-NEXT:    vpand %xmm2, %xmm1, %xmm1
+; XOP-NEXT:    vpsubd %xmm1, %xmm4, %xmm1
+; XOP-NEXT:    vpshld %xmm1, %xmm0, %xmm0
+; XOP-NEXT:    vpor %xmm0, %xmm3, %xmm0
+; XOP-NEXT:    retq
+;
+; AVX512-LABEL: rotate_demanded_bits_3:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpaddd %xmm1, %xmm1, %xmm1
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [30,30,30,30]
+; AVX512-NEXT:    vpand %xmm2, %xmm1, %xmm3
+; AVX512-NEXT:    vpsllvd %xmm3, %xmm0, %xmm3
+; AVX512-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; AVX512-NEXT:    vpsubd %xmm1, %xmm4, %xmm1
+; AVX512-NEXT:    vpand %xmm2, %xmm1, %xmm1
+; AVX512-NEXT:    vpsrlvd %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vpor %xmm0, %xmm3, %xmm0
+; AVX512-NEXT:    retq
+  %3 = shl <4 x i32> %1, <i32 1, i32 1, i32 1, i32 1>
+  %4 = and <4 x i32> %3, <i32 30, i32 30, i32 30, i32 30>
+  %5 = shl <4 x i32> %0, %4
+  %6 = sub <4 x i32> zeroinitializer, %3
+  %7 = and <4 x i32> %6, <i32 30, i32 30, i32 30, i32 30>
+  %8 = lshr <4 x i32> %0, %7
+  %9 = or <4 x i32> %5, %8
+  ret <4 x i32> %9
+}