Skip to content

Commit 2eb357a

Browse files
anjenneranjenner
committed
[AMDGPU] Feedback from pull request.
1 parent 37a1e1c commit 2eb357a

File tree

3 files changed

+142
-185
lines changed

3 files changed

+142
-185
lines changed

Diff for: ‎llvm/test/CodeGen/AMDGPU/private-memory-atomics.ll

+9-14
Original file line numberDiff line numberDiff line change
@@ -634,12 +634,11 @@ define i32 @atomicrmw_dec_private_i32(ptr addrspace(5) %ptr) {
634634
}
635635

636636
define i32 @atomicrmw_usub_cond_private_i32(ptr addrspace(5) %ptr) {
637-
; IR-LABEL: define i32 @atomicrmw_usub_cond_private_i32(
638-
; IR-SAME: ptr addrspace(5) [[PTR:%.*]]) #[[ATTR0]] {
639-
; IR-NEXT: [[TMP1:%.*]] = load i32, ptr addrspace(5) [[PTR]], align 4
637+
; IR-LABEL: @atomicrmw_usub_cond_private_i32(
638+
; IR-NEXT: [[TMP1:%.*]] = load i32, ptr addrspace(5) [[PTR:%.*]], align 4
640639
; IR-NEXT: [[TMP2:%.*]] = icmp uge i32 [[TMP1]], 4
641640
; IR-NEXT: [[TMP3:%.*]] = sub i32 [[TMP1]], 4
642-
; IR-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i32 [[TMP3]], i32 4
641+
; IR-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i32 [[TMP3]], i32 [[TMP1]]
643642
; IR-NEXT: store i32 [[NEW]], ptr addrspace(5) [[PTR]], align 4
644643
; IR-NEXT: ret i32 [[TMP1]]
645644
;
@@ -650,7 +649,7 @@ define i32 @atomicrmw_usub_cond_private_i32(ptr addrspace(5) %ptr) {
650649
; GCN-NEXT: s_waitcnt vmcnt(0)
651650
; GCN-NEXT: v_add_i32_e32 v2, vcc, -4, v1
652651
; GCN-NEXT: v_cmp_lt_u32_e32 vcc, 3, v1
653-
; GCN-NEXT: v_cndmask_b32_e32 v2, 4, v2, vcc
652+
; GCN-NEXT: v_cndmask_b32_e32 v2, v1, v2, vcc
654653
; GCN-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
655654
; GCN-NEXT: v_mov_b32_e32 v0, v1
656655
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
@@ -660,12 +659,9 @@ define i32 @atomicrmw_usub_cond_private_i32(ptr addrspace(5) %ptr) {
660659
}
661660

662661
define i32 @atomicrmw_usub_sat_private_i32(ptr addrspace(5) %ptr) {
663-
; IR-LABEL: define i32 @atomicrmw_usub_sat_private_i32(
664-
; IR-SAME: ptr addrspace(5) [[PTR:%.*]]) #[[ATTR0]] {
665-
; IR-NEXT: [[TMP1:%.*]] = load i32, ptr addrspace(5) [[PTR]], align 4
666-
; IR-NEXT: [[TMP2:%.*]] = icmp uge i32 [[TMP1]], 4
667-
; IR-NEXT: [[TMP3:%.*]] = sub i32 [[TMP1]], 4
668-
; IR-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i32 [[TMP3]], i32 0
662+
; IR-LABEL: @atomicrmw_usub_sat_private_i32(
663+
; IR-NEXT: [[TMP1:%.*]] = load i32, ptr addrspace(5) [[PTR:%.*]], align 4
664+
; IR-NEXT: [[NEW:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[TMP1]], i32 4)
669665
; IR-NEXT: store i32 [[NEW]], ptr addrspace(5) [[PTR]], align 4
670666
; IR-NEXT: ret i32 [[TMP1]]
671667
;
@@ -674,9 +670,8 @@ define i32 @atomicrmw_usub_sat_private_i32(ptr addrspace(5) %ptr) {
674670
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
675671
; GCN-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
676672
; GCN-NEXT: s_waitcnt vmcnt(0)
677-
; GCN-NEXT: v_add_i32_e32 v2, vcc, -4, v1
678-
; GCN-NEXT: v_cmp_lt_u32_e32 vcc, 3, v1
679-
; GCN-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
673+
; GCN-NEXT: v_max_u32_e32 v2, 4, v1
674+
; GCN-NEXT: v_add_i32_e32 v2, vcc, -4, v2
680675
; GCN-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
681676
; GCN-NEXT: v_mov_b32_e32 v0, v1
682677
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)

Diff for: ‎llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-i16.ll

+42-54
Original file line numberDiff line numberDiff line change
@@ -1381,9 +1381,9 @@ define i16 @test_atomicrmw_usub_cond_i16_global_agent(ptr addrspace(1) %ptr, i16
13811381
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP3]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
13821382
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[LOADED]], [[SHIFTAMT]]
13831383
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
1384-
; CHECK-NEXT: [[TMP5:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1385-
; CHECK-NEXT: [[TMP7:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1386-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP5]], i16 [[TMP7]], i16 [[VALUE]]
1384+
; CHECK-NEXT: [[TMP4:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1385+
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1386+
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP4]], i16 [[TMP5]], i16 [[EXTRACTED]]
13871387
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
13881388
; CHECK-NEXT: [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
13891389
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
@@ -1408,9 +1408,9 @@ define i16 @test_atomicrmw_usub_cond_i16_global_agent_align4(ptr addrspace(1) %p
14081408
; CHECK: atomicrmw.start:
14091409
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP1]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
14101410
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[LOADED]] to i16
1411-
; CHECK-NEXT: [[TMP3:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1412-
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1413-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP3]], i16 [[TMP5]], i16 [[VALUE]]
1411+
; CHECK-NEXT: [[TMP2:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1412+
; CHECK-NEXT: [[TMP3:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1413+
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i16 [[TMP3]], i16 [[EXTRACTED]]
14141414
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
14151415
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], -65536
14161416
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[EXTENDED]]
@@ -1440,9 +1440,9 @@ define i16 @test_atomicrmw_usub_cond_i16_local(ptr addrspace(3) %ptr, i16 %value
14401440
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP3]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
14411441
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[LOADED]], [[TMP2]]
14421442
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
1443-
; CHECK-NEXT: [[TMP5:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1444-
; CHECK-NEXT: [[TMP7:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1445-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP5]], i16 [[TMP7]], i16 [[VALUE]]
1443+
; CHECK-NEXT: [[TMP4:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1444+
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1445+
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP4]], i16 [[TMP5]], i16 [[EXTRACTED]]
14461446
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
14471447
; CHECK-NEXT: [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[TMP2]]
14481448
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
@@ -1467,9 +1467,9 @@ define i16 @test_atomicrmw_usub_cond_i16_local_align4(ptr addrspace(3) %ptr, i16
14671467
; CHECK: atomicrmw.start:
14681468
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP1]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
14691469
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[LOADED]] to i16
1470-
; CHECK-NEXT: [[TMP3:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1471-
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1472-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP3]], i16 [[TMP5]], i16 [[VALUE]]
1470+
; CHECK-NEXT: [[TMP2:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1471+
; CHECK-NEXT: [[TMP3:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1472+
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i16 [[TMP3]], i16 [[EXTRACTED]]
14731473
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
14741474
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], -65536
14751475
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[EXTENDED]]
@@ -1500,9 +1500,9 @@ define i16 @test_atomicrmw_usub_cond_i16_flat_agent(ptr %ptr, i16 %value) {
15001500
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP3]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
15011501
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[LOADED]], [[SHIFTAMT]]
15021502
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
1503-
; CHECK-NEXT: [[TMP5:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1504-
; CHECK-NEXT: [[TMP7:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1505-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP5]], i16 [[TMP7]], i16 [[VALUE]]
1503+
; CHECK-NEXT: [[TMP4:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1504+
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1505+
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP4]], i16 [[TMP5]], i16 [[EXTRACTED]]
15061506
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
15071507
; CHECK-NEXT: [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
15081508
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
@@ -1527,9 +1527,9 @@ define i16 @test_atomicrmw_usub_cond_i16_flat_agent_align4(ptr %ptr, i16 %value)
15271527
; CHECK: atomicrmw.start:
15281528
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP1]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
15291529
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[LOADED]] to i16
1530-
; CHECK-NEXT: [[TMP3:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1531-
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1532-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP3]], i16 [[TMP5]], i16 [[VALUE]]
1530+
; CHECK-NEXT: [[TMP2:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1531+
; CHECK-NEXT: [[TMP3:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1532+
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i16 [[TMP3]], i16 [[EXTRACTED]]
15331533
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
15341534
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], -65536
15351535
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[EXTENDED]]
@@ -1560,16 +1560,14 @@ define i16 @test_atomicrmw_usub_sat_i16_global_agent(ptr addrspace(1) %ptr, i16
15601560
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP3]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
15611561
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[LOADED]], [[SHIFTAMT]]
15621562
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
1563-
; CHECK-NEXT: [[TMP4:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1564-
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1565-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP4]], i16 [[TMP5]], i16 0
1563+
; CHECK-NEXT: [[NEW:%.*]] = call i16 @llvm.usub.sat.i16(i16 [[EXTRACTED]], i16 [[VALUE:%.*]])
15661564
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
15671565
; CHECK-NEXT: [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
15681566
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
15691567
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
1570-
; CHECK-NEXT: [[TMP8:%.*]] = cmpxchg ptr addrspace(1) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] syncscope("agent") seq_cst seq_cst, align 4
1571-
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP8]], 1
1572-
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP8]], 0
1568+
; CHECK-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] syncscope("agent") seq_cst seq_cst, align 4
1569+
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1570+
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP4]], 0
15731571
; CHECK-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
15741572
; CHECK: atomicrmw.end:
15751573
; CHECK-NEXT: [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
@@ -1587,15 +1585,13 @@ define i16 @test_atomicrmw_usub_sat_i16_global_agent_align4(ptr addrspace(1) %pt
15871585
; CHECK: atomicrmw.start:
15881586
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP1]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
15891587
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[LOADED]] to i16
1590-
; CHECK-NEXT: [[TMP2:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1591-
; CHECK-NEXT: [[TMP3:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1592-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i16 [[TMP3]], i16 0
1588+
; CHECK-NEXT: [[NEW:%.*]] = call i16 @llvm.usub.sat.i16(i16 [[EXTRACTED]], i16 [[VALUE:%.*]])
15931589
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
15941590
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], -65536
15951591
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[EXTENDED]]
1596-
; CHECK-NEXT: [[TMP6:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[LOADED]], i32 [[INSERTED]] syncscope("agent") seq_cst seq_cst, align 4
1597-
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
1598-
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP6]], 0
1592+
; CHECK-NEXT: [[TMP2:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[LOADED]], i32 [[INSERTED]] syncscope("agent") seq_cst seq_cst, align 4
1593+
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP2]], 1
1594+
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP2]], 0
15991595
; CHECK-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
16001596
; CHECK: atomicrmw.end:
16011597
; CHECK-NEXT: [[EXTRACTED1:%.*]] = trunc i32 [[NEWLOADED]] to i16
@@ -1619,16 +1615,14 @@ define i16 @test_atomicrmw_usub_sat_i16_local(ptr addrspace(3) %ptr, i16 %value)
16191615
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP3]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
16201616
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[LOADED]], [[TMP2]]
16211617
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
1622-
; CHECK-NEXT: [[TMP4:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1623-
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1624-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP4]], i16 [[TMP5]], i16 0
1618+
; CHECK-NEXT: [[NEW:%.*]] = call i16 @llvm.usub.sat.i16(i16 [[EXTRACTED]], i16 [[VALUE:%.*]])
16251619
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
16261620
; CHECK-NEXT: [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[TMP2]]
16271621
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
16281622
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
1629-
; CHECK-NEXT: [[TMP8:%.*]] = cmpxchg ptr addrspace(3) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
1630-
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP8]], 1
1631-
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP8]], 0
1623+
; CHECK-NEXT: [[TMP4:%.*]] = cmpxchg ptr addrspace(3) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
1624+
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1625+
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP4]], 0
16321626
; CHECK-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
16331627
; CHECK: atomicrmw.end:
16341628
; CHECK-NEXT: [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[TMP2]]
@@ -1646,15 +1640,13 @@ define i16 @test_atomicrmw_usub_sat_i16_local_align4(ptr addrspace(3) %ptr, i16
16461640
; CHECK: atomicrmw.start:
16471641
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP1]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
16481642
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[LOADED]] to i16
1649-
; CHECK-NEXT: [[TMP2:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1650-
; CHECK-NEXT: [[TMP3:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1651-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i16 [[TMP3]], i16 0
1643+
; CHECK-NEXT: [[NEW:%.*]] = call i16 @llvm.usub.sat.i16(i16 [[EXTRACTED]], i16 [[VALUE:%.*]])
16521644
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
16531645
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], -65536
16541646
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[EXTENDED]]
1655-
; CHECK-NEXT: [[TMP6:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
1656-
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
1657-
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP6]], 0
1647+
; CHECK-NEXT: [[TMP2:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
1648+
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP2]], 1
1649+
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP2]], 0
16581650
; CHECK-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
16591651
; CHECK: atomicrmw.end:
16601652
; CHECK-NEXT: [[EXTRACTED1:%.*]] = trunc i32 [[NEWLOADED]] to i16
@@ -1679,16 +1671,14 @@ define i16 @test_atomicrmw_usub_sat_i16_flat_agent(ptr %ptr, i16 %value) {
16791671
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP3]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
16801672
; CHECK-NEXT: [[SHIFTED:%.*]] = lshr i32 [[LOADED]], [[SHIFTAMT]]
16811673
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[SHIFTED]] to i16
1682-
; CHECK-NEXT: [[TMP4:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1683-
; CHECK-NEXT: [[TMP5:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1684-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP4]], i16 [[TMP5]], i16 0
1674+
; CHECK-NEXT: [[NEW:%.*]] = call i16 @llvm.usub.sat.i16(i16 [[EXTRACTED]], i16 [[VALUE:%.*]])
16851675
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
16861676
; CHECK-NEXT: [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
16871677
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
16881678
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
1689-
; CHECK-NEXT: [[TMP8:%.*]] = cmpxchg ptr [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] syncscope("agent") seq_cst seq_cst, align 4
1690-
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP8]], 1
1691-
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP8]], 0
1679+
; CHECK-NEXT: [[TMP4:%.*]] = cmpxchg ptr [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] syncscope("agent") seq_cst seq_cst, align 4
1680+
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
1681+
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP4]], 0
16921682
; CHECK-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
16931683
; CHECK: atomicrmw.end:
16941684
; CHECK-NEXT: [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
@@ -1706,15 +1696,13 @@ define i16 @test_atomicrmw_usub_sat_i16_flat_agent_align4(ptr %ptr, i16 %value)
17061696
; CHECK: atomicrmw.start:
17071697
; CHECK-NEXT: [[LOADED:%.*]] = phi i32 [ [[TMP1]], [[TMP0:%.*]] ], [ [[NEWLOADED:%.*]], [[ATOMICRMW_START]] ]
17081698
; CHECK-NEXT: [[EXTRACTED:%.*]] = trunc i32 [[LOADED]] to i16
1709-
; CHECK-NEXT: [[TMP2:%.*]] = icmp uge i16 [[EXTRACTED]], [[VALUE:%.*]]
1710-
; CHECK-NEXT: [[TMP3:%.*]] = sub i16 [[EXTRACTED]], [[VALUE]]
1711-
; CHECK-NEXT: [[NEW:%.*]] = select i1 [[TMP2]], i16 [[TMP3]], i16 0
1699+
; CHECK-NEXT: [[NEW:%.*]] = call i16 @llvm.usub.sat.i16(i16 [[EXTRACTED]], i16 [[VALUE:%.*]])
17121700
; CHECK-NEXT: [[EXTENDED:%.*]] = zext i16 [[NEW]] to i32
17131701
; CHECK-NEXT: [[UNMASKED:%.*]] = and i32 [[LOADED]], -65536
17141702
; CHECK-NEXT: [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[EXTENDED]]
1715-
; CHECK-NEXT: [[TMP6:%.*]] = cmpxchg ptr [[PTR]], i32 [[LOADED]], i32 [[INSERTED]] syncscope("agent") seq_cst seq_cst, align 4
1716-
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
1717-
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP6]], 0
1703+
; CHECK-NEXT: [[TMP2:%.*]] = cmpxchg ptr [[PTR]], i32 [[LOADED]], i32 [[INSERTED]] syncscope("agent") seq_cst seq_cst, align 4
1704+
; CHECK-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP2]], 1
1705+
; CHECK-NEXT: [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP2]], 0
17181706
; CHECK-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
17191707
; CHECK: atomicrmw.end:
17201708
; CHECK-NEXT: [[EXTRACTED1:%.*]] = trunc i32 [[NEWLOADED]] to i16

0 commit comments

Comments
 (0)