225 lines · plain
1; Test 8-bit atomic min/max operations.2;3; RUN: llc < %s -mtriple=s390x-linux-gnu -disable-block-placement | FileCheck %s4; RUN: llc < %s -mtriple=s390x-linux-gnu -disable-block-placement | FileCheck %s -check-prefix=CHECK-SHIFT15; RUN: llc < %s -mtriple=s390x-linux-gnu -disable-block-placement | FileCheck %s -check-prefix=CHECK-SHIFT26 7; Check signed minimum.8; - CHECK is for the main loop.9; - CHECK-SHIFT1 makes sure that the negated shift count used by the second10; RLL is set up correctly. The negation is independent of the NILL and L11; tested in CHECK.12; - CHECK-SHIFT2 makes sure that %b is shifted into the high part of the word13; before being used, and that the low bits are set to 1. This sequence is14; independent of the other loop prologue instructions.15define i8 @f1(ptr %src, i8 %b) {16; CHECK-LABEL: f1:17; CHECK: risbg [[RISBG:%r[1-9]+]], %r2, 0, 189, 0{{$}}18; CHECK-DAG: sll %r2, 319; CHECK-DAG: l [[OLD:%r[0-9]+]], 0([[RISBG]])20; CHECK: [[LOOP:\.[^:]*]]:21; CHECK: rll [[ROT:%r[0-9]+]], [[OLD]], 0(%r2)22; CHECK: crjle [[ROT]], %r3, [[KEEP:\..*]]23; CHECK: risbg [[ROT]], %r3, 32, 39, 024; CHECK: [[KEEP]]:25; CHECK: rll [[NEW:%r[0-9]+]], [[ROT]], 0({{%r[1-9]+}})26; CHECK: cs [[OLD]], [[NEW]], 0([[RISBG]])27; CHECK: jl [[LOOP]]28; CHECK: rll %r2, [[OLD]], 8(%r2)29; CHECK: br %r1430;31; CHECK-SHIFT1-LABEL: f1:32; CHECK-SHIFT1: sll %r2, 333; CHECK-SHIFT1: lcr [[NEGSHIFT:%r[1-9]+]], %r234; CHECK-SHIFT1: rll35; CHECK-SHIFT1: rll {{%r[0-9]+}}, {{%r[0-9]+}}, 0([[NEGSHIFT]])36; CHECK-SHIFT1: rll37; CHECK-SHIFT1: br %r1438;39; CHECK-SHIFT2-LABEL: f1:40; CHECK-SHIFT2: sll %r3, 2441; CHECK-SHIFT2: rll42; CHECK-SHIFT2: crjle {{%r[0-9]+}}, %r343; CHECK-SHIFT2: rll44; CHECK-SHIFT2: rll45; CHECK-SHIFT2: br %r1446 %res = atomicrmw min ptr %src, i8 %b seq_cst47 ret i8 %res48}49 50; Check signed maximum.51define i8 @f2(ptr %src, i8 %b) {52; CHECK-LABEL: f2:53; CHECK: risbg [[RISBG:%r[1-9]+]], %r2, 0, 189, 0{{$}}54; CHECK-DAG: sll %r2, 355; CHECK-DAG: l [[OLD:%r[0-9]+]], 0([[RISBG]])56; CHECK: [[LOOP:\.[^:]*]]:57; CHECK: rll [[ROT:%r[0-9]+]], [[OLD]], 0(%r2)58; CHECK: crjhe [[ROT]], %r3, [[KEEP:\..*]]59; CHECK: risbg [[ROT]], %r3, 32, 39, 060; CHECK: [[KEEP]]:61; CHECK: rll [[NEW:%r[0-9]+]], [[ROT]], 0({{%r[1-9]+}})62; CHECK: cs [[OLD]], [[NEW]], 0([[RISBG]])63; CHECK: jl [[LOOP]]64; CHECK: rll %r2, [[OLD]], 8(%r2)65; CHECK: br %r1466;67; CHECK-SHIFT1-LABEL: f2:68; CHECK-SHIFT1: sll %r2, 369; CHECK-SHIFT1: lcr [[NEGSHIFT:%r[1-9]+]], %r270; CHECK-SHIFT1: rll71; CHECK-SHIFT1: rll {{%r[0-9]+}}, {{%r[0-9]+}}, 0([[NEGSHIFT]])72; CHECK-SHIFT1: rll73; CHECK-SHIFT1: br %r1474;75; CHECK-SHIFT2-LABEL: f2:76; CHECK-SHIFT2: sll %r3, 2477; CHECK-SHIFT2: rll78; CHECK-SHIFT2: crjhe {{%r[0-9]+}}, %r379; CHECK-SHIFT2: rll80; CHECK-SHIFT2: rll81; CHECK-SHIFT2: br %r1482 %res = atomicrmw max ptr %src, i8 %b seq_cst83 ret i8 %res84}85 86; Check unsigned minimum.87define i8 @f3(ptr %src, i8 %b) {88; CHECK-LABEL: f3:89; CHECK: risbg [[RISBG:%r[1-9]+]], %r2, 0, 189, 0{{$}}90; CHECK-DAG: sll %r2, 391; CHECK-DAG: l [[OLD:%r[0-9]+]], 0([[RISBG]])92; CHECK: [[LOOP:\.[^:]*]]:93; CHECK: rll [[ROT:%r[0-9]+]], [[OLD]], 0(%r2)94; CHECK: clrjle [[ROT]], %r3, [[KEEP:\..*]]95; CHECK: risbg [[ROT]], %r3, 32, 39, 096; CHECK: [[KEEP]]:97; CHECK: rll [[NEW:%r[0-9]+]], [[ROT]], 0({{%r[1-9]+}})98; CHECK: cs [[OLD]], [[NEW]], 0([[RISBG]])99; CHECK: jl [[LOOP]]100; CHECK: rll %r2, [[OLD]], 8(%r2)101; CHECK: br %r14102;103; CHECK-SHIFT1-LABEL: f3:104; CHECK-SHIFT1: sll %r2, 3105; CHECK-SHIFT1: lcr [[NEGSHIFT:%r[1-9]+]], %r2106; CHECK-SHIFT1: rll107; CHECK-SHIFT1: rll {{%r[0-9]+}}, {{%r[0-9]+}}, 0([[NEGSHIFT]])108; CHECK-SHIFT1: rll109; CHECK-SHIFT1: br %r14110;111; CHECK-SHIFT2-LABEL: f3:112; CHECK-SHIFT2: sll %r3, 24113; CHECK-SHIFT2: rll114; CHECK-SHIFT2: clrjle {{%r[0-9]+}}, %r3,115; CHECK-SHIFT2: rll116; CHECK-SHIFT2: rll117; CHECK-SHIFT2: br %r14118 %res = atomicrmw umin ptr %src, i8 %b seq_cst119 ret i8 %res120}121 122; Check unsigned maximum.123define i8 @f4(ptr %src, i8 %b) {124; CHECK-LABEL: f4:125; CHECK: risbg [[RISBG:%r[1-9]+]], %r2, 0, 189, 0{{$}}126; CHECK-DAG: sll %r2, 3127; CHECK-DAG: l [[OLD:%r[0-9]+]], 0([[RISBG]])128; CHECK: [[LOOP:\.[^:]*]]:129; CHECK: rll [[ROT:%r[0-9]+]], [[OLD]], 0(%r2)130; CHECK: clrjhe [[ROT]], %r3, [[KEEP:\..*]]131; CHECK: risbg [[ROT]], %r3, 32, 39, 0132; CHECK: [[KEEP]]:133; CHECK: rll [[NEW:%r[0-9]+]], [[ROT]], 0({{%r[1-9]+}})134; CHECK: cs [[OLD]], [[NEW]], 0([[RISBG]])135; CHECK: jl [[LOOP]]136; CHECK: rll %r2, [[OLD]], 8(%r2)137; CHECK: br %r14138;139; CHECK-SHIFT1-LABEL: f4:140; CHECK-SHIFT1: sll %r2, 3141; CHECK-SHIFT1: lcr [[NEGSHIFT:%r[1-9]+]], %r2142; CHECK-SHIFT1: rll143; CHECK-SHIFT1: rll {{%r[0-9]+}}, {{%r[0-9]+}}, 0([[NEGSHIFT]])144; CHECK-SHIFT1: rll145; CHECK-SHIFT1: br %r14146;147; CHECK-SHIFT2-LABEL: f4:148; CHECK-SHIFT2: sll %r3, 24149; CHECK-SHIFT2: rll150; CHECK-SHIFT2: clrjhe {{%r[0-9]+}}, %r3,151; CHECK-SHIFT2: rll152; CHECK-SHIFT2: rll153; CHECK-SHIFT2: br %r14154 %res = atomicrmw umax ptr %src, i8 %b seq_cst155 ret i8 %res156}157 158; Check the lowest useful signed minimum value. We need to load 0x81000000159; into the source register.160define i8 @f5(ptr %src) {161; CHECK-LABEL: f5:162; CHECK: llilh [[SRC2:%r[0-9]+]], 33024163; CHECK: crjle [[ROT:%r[0-9]+]], [[SRC2]]164; CHECK: risbg [[ROT]], [[SRC2]], 32, 39, 0165; CHECK: br %r14166;167; CHECK-SHIFT1-LABEL: f5:168; CHECK-SHIFT1: br %r14169; CHECK-SHIFT2-LABEL: f5:170; CHECK-SHIFT2: br %r14171 %res = atomicrmw min ptr %src, i8 -127 seq_cst172 ret i8 %res173}174 175; Check the highest useful signed maximum value. We need to load 0x7e000000176; into the source register.177define i8 @f6(ptr %src) {178; CHECK-LABEL: f6:179; CHECK: llilh [[SRC2:%r[0-9]+]], 32256180; CHECK: crjhe [[ROT:%r[0-9]+]], [[SRC2]]181; CHECK: risbg [[ROT]], [[SRC2]], 32, 39, 0182; CHECK: br %r14183;184; CHECK-SHIFT1-LABEL: f6:185; CHECK-SHIFT1: br %r14186; CHECK-SHIFT2-LABEL: f6:187; CHECK-SHIFT2: br %r14188 %res = atomicrmw max ptr %src, i8 126 seq_cst189 ret i8 %res190}191 192; Check the lowest useful unsigned minimum value. We need to load 0x01000000193; into the source register.194define i8 @f7(ptr %src) {195; CHECK-LABEL: f7:196; CHECK: llilh [[SRC2:%r[0-9]+]], 256197; CHECK: clrjle [[ROT:%r[0-9]+]], [[SRC2]],198; CHECK: risbg [[ROT]], [[SRC2]], 32, 39, 0199; CHECK: br %r14200;201; CHECK-SHIFT1-LABEL: f7:202; CHECK-SHIFT1: br %r14203; CHECK-SHIFT2-LABEL: f7:204; CHECK-SHIFT2: br %r14205 %res = atomicrmw umin ptr %src, i8 1 seq_cst206 ret i8 %res207}208 209; Check the highest useful unsigned maximum value. We need to load 0xfe000000210; into the source register.211define i8 @f8(ptr %src) {212; CHECK-LABEL: f8:213; CHECK: llilh [[SRC2:%r[0-9]+]], 65024214; CHECK: clrjhe [[ROT:%r[0-9]+]], [[SRC2]],215; CHECK: risbg [[ROT]], [[SRC2]], 32, 39, 0216; CHECK: br %r14217;218; CHECK-SHIFT1-LABEL: f8:219; CHECK-SHIFT1: br %r14220; CHECK-SHIFT2-LABEL: f8:221; CHECK-SHIFT2: br %r14222 %res = atomicrmw umax ptr %src, i8 254 seq_cst223 ret i8 %res224}225