155 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=aarch64-- | FileCheck %s3 4declare i16 @llvm.ushl.sat.i16(i16, i16)5declare <2 x i16> @llvm.ushl.sat.v2i16(<2 x i16>, <2 x i16>)6 7; fold (shlsat undef, x) -> 08define i16 @combine_shl_undef(i16 %x, i16 %y) nounwind {9; CHECK-LABEL: combine_shl_undef:10; CHECK: // %bb.0:11; CHECK-NEXT: mov w0, wzr12; CHECK-NEXT: ret13 %tmp = call i16 @llvm.ushl.sat.i16(i16 undef, i16 %y)14 ret i16 %tmp15}16 17; fold (shlsat x, undef) -> undef18define i16 @combine_shl_by_undef(i16 %x, i16 %y) nounwind {19; CHECK-LABEL: combine_shl_by_undef:20; CHECK: // %bb.0:21; CHECK-NEXT: ret22 %tmp = call i16 @llvm.ushl.sat.i16(i16 %x, i16 undef)23 ret i16 %tmp24}25 26; fold (shlsat poison, x) -> 027define i16 @combine_shl_poison(i16 %x, i16 %y) nounwind {28; CHECK-LABEL: combine_shl_poison:29; CHECK: // %bb.0:30; CHECK-NEXT: mov w0, wzr31; CHECK-NEXT: ret32 %tmp = call i16 @llvm.ushl.sat.i16(i16 poison, i16 %y)33 ret i16 %tmp34}35 36; fold (shlsat x, poison) -> undef37define i16 @combine_shl_by_poison(i16 %x, i16 %y) nounwind {38; CHECK-LABEL: combine_shl_by_poison:39; CHECK: // %bb.0:40; CHECK-NEXT: ret41 %tmp = call i16 @llvm.ushl.sat.i16(i16 %x, i16 poison)42 ret i16 %tmp43}44 45; fold (shlsat x, bitwidth) -> undef46define i16 @combine_shl_by_bitwidth(i16 %x, i16 %y) nounwind {47; CHECK-LABEL: combine_shl_by_bitwidth:48; CHECK: // %bb.0:49; CHECK-NEXT: ret50 %tmp = call i16 @llvm.ushl.sat.i16(i16 %x, i16 16)51 ret i16 %tmp52}53 54; fold (ushlsat 0, x) -> 055define i16 @combine_shl_zero(i16 %x, i16 %y) nounwind {56; CHECK-LABEL: combine_shl_zero:57; CHECK: // %bb.0:58; CHECK-NEXT: mov w0, wzr59; CHECK-NEXT: ret60 %tmp = call i16 @llvm.ushl.sat.i16(i16 0, i16 %y)61 ret i16 %tmp62}63 64; fold (ushlsat x, 0) -> x65define i16 @combine_shlsat_by_zero(i16 %x, i16 %y) nounwind {66; CHECK-LABEL: combine_shlsat_by_zero:67; CHECK: // %bb.0:68; CHECK-NEXT: ret69 %tmp = call i16 @llvm.ushl.sat.i16(i16 %x, i16 0)70 ret i16 %tmp71}72 73; fold (ushlsat c1, c2) -> c374define i16 @combine_shlsat_constfold(i16 %x, i16 %y) nounwind {75; CHECK-LABEL: combine_shlsat_constfold:76; CHECK: // %bb.0:77; CHECK-NEXT: mov w0, #32 // =0x2078; CHECK-NEXT: ret79 %tmp = call i16 @llvm.ushl.sat.i16(i16 8, i16 2)80 ret i16 %tmp81}82 83; fold (ushlsat c1, c2) -> sat max84define i16 @combine_shlsat_satmax(i16 %x, i16 %y) nounwind {85; CHECK-LABEL: combine_shlsat_satmax:86; CHECK: // %bb.0:87; CHECK-NEXT: mov w0, #65535 // =0xffff88; CHECK-NEXT: ret89 %tmp = call i16 @llvm.ushl.sat.i16(i16 8, i16 15)90 ret i16 %tmp91}92 93 94declare void @sink2xi16(i16, i16)95 96; fold (ushlsat c1, c2) -> c3 , c1/c2/c3 being vectors97define void @combine_shlsat_vector() nounwind {98; CHECK-LABEL: combine_shlsat_vector:99; CHECK: // %bb.0:100; CHECK-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill101; CHECK-NEXT: mov w0, #32 // =0x20102; CHECK-NEXT: mov w1, #65535 // =0xffff103; CHECK-NEXT: bl sink2xi16104; CHECK-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload105; CHECK-NEXT: ret106 %tmp = call <2 x i16> @llvm.ushl.sat.v2i16(<2 x i16><i16 8, i16 8>,107 <2 x i16><i16 2, i16 15>)108 ; Pass elements as arguments in a call to get CHECK statements that verify109 ; the constant folding.110 %e0 = extractelement <2 x i16> %tmp, i16 0111 %e1 = extractelement <2 x i16> %tmp, i16 1112 call void @sink2xi16(i16 %e0, i16 %e1)113 ret void114}115 116; Fold shlsat -> shl, if known not to saturate.117define i16 @combine_shlsat_to_shl(i16 %x) nounwind {118; CHECK-LABEL: combine_shlsat_to_shl:119; CHECK: // %bb.0:120; CHECK-NEXT: and w0, w0, #0xfffffffc121; CHECK-NEXT: ret122 %x2 = lshr i16 %x, 2123 %tmp = call i16 @llvm.ushl.sat.i16(i16 %x2, i16 2)124 ret i16 %tmp125}126 127; Do not fold shlsat -> shl.128define i16 @combine_shlsat_to_shl_no_fold(i16 %x) nounwind {129; CHECK-LABEL: combine_shlsat_to_shl_no_fold:130; CHECK: // %bb.0:131; CHECK-NEXT: lsl w8, w0, #14132; CHECK-NEXT: and w8, w8, #0x3fff0000133; CHECK-NEXT: lsl w9, w8, #3134; CHECK-NEXT: cmp w8, w9, lsr #3135; CHECK-NEXT: csinv w8, w9, wzr, eq136; CHECK-NEXT: lsr w0, w8, #16137; CHECK-NEXT: ret138 %x2 = lshr i16 %x, 2139 %tmp = call i16 @llvm.ushl.sat.i16(i16 %x2, i16 3)140 ret i16 %tmp141}142 143; Fold shlsat -> shl, if known not to saturate.144define <2 x i16> @combine_shlsat_to_shl_vec(<2 x i8> %a) nounwind {145; CHECK-LABEL: combine_shlsat_to_shl_vec:146; CHECK: // %bb.0:147; CHECK-NEXT: shl v0.2s, v0.2s, #8148; CHECK-NEXT: ret149 %ext = zext <2 x i8> %a to <2 x i16>150 %tmp = call <2 x i16> @llvm.ushl.sat.v2i16(151 <2 x i16> %ext,152 <2 x i16> <i16 8, i16 8>)153 ret <2 x i16> %tmp154}155