[llvm-branch-commits] [llvm] AMDGPU/GlobalISel: Handle atomic sextload and zextload (PR #111721)
Matt Arsenault via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Oct 9 11:07:49 PDT 2024
https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/111721
>From 81dad077b21128cfc827ffb18b12631407a2bde4 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Wed, 9 Oct 2024 19:35:29 +0400
Subject: [PATCH] AMDGPU/GlobalISel: Handle atomic sextload and zextload
Atomic loads are handled differently from the DAG, and have separate opcodes
and explicit control over the extensions, like ordinary loads. Add
new patterns for these.
There's room for cleanup and improvement. d16 cases aren't handled.
Fixes #111645
---
llvm/lib/Target/AMDGPU/AMDGPUGISel.td | 2 +
llvm/lib/Target/AMDGPU/AMDGPUInstructions.td | 21 +
llvm/lib/Target/AMDGPU/BUFInstructions.td | 7 +
llvm/lib/Target/AMDGPU/DSInstructions.td | 7 +
llvm/lib/Target/AMDGPU/FLATInstructions.td | 16 +
llvm/lib/Target/AMDGPU/SIInstrInfo.td | 45 ++
.../AMDGPU/GlobalISel/atomic_load_flat.ll | 331 ++++++++++++
.../AMDGPU/GlobalISel/atomic_load_global.ll | 478 ++++++++++++++++
.../AMDGPU/GlobalISel/atomic_load_local_2.ll | 509 ++++++++++++++++++
9 files changed, 1416 insertions(+)
create mode 100644 llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_flat.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_global.ll
create mode 100644 llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_local_2.ll
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
index 278d3536add916..d348f489d95dd3 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
@@ -207,6 +207,8 @@ def : GINodeEquiv<G_STORE, AMDGPUst_glue> {
def : GINodeEquiv<G_LOAD, AMDGPUatomic_ld_glue> {
bit CheckMMOIsAtomic = 1;
+ let IfSignExtend = G_SEXTLOAD;
+ let IfZeroExtend = G_ZEXTLOAD;
}
def : GINodeEquiv<G_STORE, AMDGPUatomic_st_glue> {
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
index 09987a6504b9d0..7816ae911312c2 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -521,6 +521,27 @@ def atomic_load_64_#as : PatFrag<(ops node:$ptr), (atomic_load_64 node:$ptr)> {
let IsAtomic = 1;
let MemoryVT = i64;
}
+
+def atomic_load_zext_8_#as : PatFrag<(ops node:$ptr), (atomic_load_zext node:$ptr)> {
+ let IsAtomic = 1;
+ let MemoryVT = i8;
+}
+
+def atomic_load_sext_8_#as : PatFrag<(ops node:$ptr), (atomic_load_sext node:$ptr)> {
+ let IsAtomic = 1;
+ let MemoryVT = i8;
+}
+
+def atomic_load_zext_16_#as : PatFrag<(ops node:$ptr), (atomic_load_zext node:$ptr)> {
+ let IsAtomic = 1;
+ let MemoryVT = i16;
+}
+
+def atomic_load_sext_16_#as : PatFrag<(ops node:$ptr), (atomic_load_sext node:$ptr)> {
+ let IsAtomic = 1;
+ let MemoryVT = i16;
+}
+
} // End let AddressSpaces
} // End foreach as
diff --git a/llvm/lib/Target/AMDGPU/BUFInstructions.td b/llvm/lib/Target/AMDGPU/BUFInstructions.td
index 6bdff9862e55ac..4ed0f9ade871f1 100644
--- a/llvm/lib/Target/AMDGPU/BUFInstructions.td
+++ b/llvm/lib/Target/AMDGPU/BUFInstructions.td
@@ -983,15 +983,22 @@ defm BUFFER_LOAD_LDS_U16 : MUBUF_Pseudo_Loads_LDSOpc <
>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_UBYTE", i32, atomic_load_8_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_UBYTE", i32, atomic_load_zext_8_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_USHORT", i32, atomic_load_16_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_USHORT", i32, atomic_load_zext_16_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_UBYTE", i16, atomic_load_8_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_USHORT", i16, atomic_load_16_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_UBYTE", i16, atomic_load_zext_8_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_USHORT", i16, atomic_load_zext_16_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_UBYTE", i32, extloadi8_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_UBYTE", i32, zextloadi8_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_SBYTE", i32, sextloadi8_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_SBYTE", i32, atomic_load_sext_8_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_SBYTE", i32, atomic_load_sext_16_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_USHORT", i32, extloadi16_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_USHORT", i32, zextloadi16_global>;
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_SSHORT", i32, sextloadi16_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_SSHORT", i32, atomic_load_sext_16_global>;
foreach vt = Reg32Types.types in {
defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_DWORD", vt, load_global>;
diff --git a/llvm/lib/Target/AMDGPU/DSInstructions.td b/llvm/lib/Target/AMDGPU/DSInstructions.td
index e9283fde85a48d..7724821bbd7c36 100644
--- a/llvm/lib/Target/AMDGPU/DSInstructions.td
+++ b/llvm/lib/Target/AMDGPU/DSInstructions.td
@@ -795,12 +795,19 @@ defm : DSReadPat_mc <DS_READ_B32, vt, "load_local">;
defm : DSReadPat_mc <DS_READ_U8, i16, "atomic_load_8_local">;
defm : DSReadPat_mc <DS_READ_U8, i32, "atomic_load_8_local">;
+defm : DSReadPat_mc <DS_READ_U8, i16, "atomic_load_zext_8_local">;
+defm : DSReadPat_mc <DS_READ_U8, i32, "atomic_load_zext_8_local">;
+defm : DSReadPat_mc <DS_READ_I8, i16, "atomic_load_sext_8_local">;
+defm : DSReadPat_mc <DS_READ_I8, i32, "atomic_load_sext_8_local">;
defm : DSReadPat_mc <DS_READ_U16, i16, "atomic_load_16_local">;
defm : DSReadPat_mc <DS_READ_U16, i32, "atomic_load_16_local">;
+defm : DSReadPat_mc <DS_READ_U16, i32, "atomic_load_zext_16_local">;
+defm : DSReadPat_mc <DS_READ_I16, i32, "atomic_load_sext_16_local">;
defm : DSReadPat_mc <DS_READ_B32, i32, "atomic_load_32_local">;
defm : DSReadPat_mc <DS_READ_B64, i64, "atomic_load_64_local">;
let OtherPredicates = [D16PreservesUnusedBits] in {
+// TODO: Atomic loads
def : DSReadPat_D16<DS_READ_U16_D16_HI, load_d16_hi_local, v2i16>;
def : DSReadPat_D16<DS_READ_U16_D16_HI, load_d16_hi_local, v2f16>;
def : DSReadPat_D16<DS_READ_U8_D16_HI, az_extloadi8_d16_hi_local, v2i16>;
diff --git a/llvm/lib/Target/AMDGPU/FLATInstructions.td b/llvm/lib/Target/AMDGPU/FLATInstructions.td
index a9ab0c5a453e8e..0f2566937022e8 100644
--- a/llvm/lib/Target/AMDGPU/FLATInstructions.td
+++ b/llvm/lib/Target/AMDGPU/FLATInstructions.td
@@ -1355,11 +1355,17 @@ let OtherPredicates = [HasFlatAddressSpace] in {
def : FlatLoadPat <FLAT_LOAD_UBYTE, atomic_load_8_flat, i32>;
def : FlatLoadPat <FLAT_LOAD_UBYTE, atomic_load_8_flat, i16>;
+def : FlatLoadPat <FLAT_LOAD_UBYTE, atomic_load_zext_8_flat, i32>;
+def : FlatLoadPat <FLAT_LOAD_UBYTE, atomic_load_zext_8_flat, i16>;
def : FlatLoadPat <FLAT_LOAD_USHORT, atomic_load_16_flat, i32>;
def : FlatLoadPat <FLAT_LOAD_USHORT, atomic_load_16_flat, i16>;
+def : FlatLoadPat <FLAT_LOAD_USHORT, atomic_load_zext_16_flat, i32>;
+def : FlatLoadPat <FLAT_LOAD_USHORT, atomic_load_zext_16_flat, i16>;
def : FlatLoadPat <FLAT_LOAD_UBYTE, extloadi8_flat, i32>;
def : FlatLoadPat <FLAT_LOAD_UBYTE, zextloadi8_flat, i32>;
def : FlatLoadPat <FLAT_LOAD_SBYTE, sextloadi8_flat, i32>;
+def : FlatLoadPat <FLAT_LOAD_SBYTE, atomic_load_sext_8_flat, i32>;
+def : FlatLoadPat <FLAT_LOAD_SBYTE, atomic_load_sext_8_flat, i16>;
def : FlatLoadPat <FLAT_LOAD_UBYTE, extloadi8_flat, i16>;
def : FlatLoadPat <FLAT_LOAD_UBYTE, zextloadi8_flat, i16>;
def : FlatLoadPat <FLAT_LOAD_SBYTE, sextloadi8_flat, i16>;
@@ -1456,6 +1462,7 @@ def : FlatStorePat <FLAT_STORE_BYTE_D16_HI, truncstorei8_hi16_flat, i32>;
}
let OtherPredicates = [D16PreservesUnusedBits] in {
+// TODO: Handle atomic loads
def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_flat, v2i16>;
def : FlatLoadPat_D16 <FLAT_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_flat, v2f16>;
def : FlatLoadPat_D16 <FLAT_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_flat, v2i16>;
@@ -1477,8 +1484,14 @@ let OtherPredicates = [HasFlatGlobalInsts] in {
defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, atomic_load_8_global, i32>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, atomic_load_8_global, i16>;
+defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, atomic_load_zext_8_global, i32>;
+defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, atomic_load_zext_8_global, i16>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, atomic_load_16_global, i32>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, atomic_load_16_global, i16>;
+defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, atomic_load_zext_16_global, i32>;
+defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, atomic_load_zext_16_global, i16>;
+defm : GlobalFLATLoadPats <GLOBAL_LOAD_SBYTE, atomic_load_sext_8_global, i32>;
+defm : GlobalFLATLoadPats <GLOBAL_LOAD_SBYTE, atomic_load_sext_8_global, i16>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, extloadi8_global, i32>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_UBYTE, zextloadi8_global, i32>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_SBYTE, sextloadi8_global, i32>;
@@ -1488,6 +1501,8 @@ defm : GlobalFLATLoadPats <GLOBAL_LOAD_SBYTE, sextloadi8_global, i16>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, extloadi16_global, i32>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, zextloadi16_global, i32>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_SSHORT, sextloadi16_global, i32>;
+defm : GlobalFLATLoadPats <GLOBAL_LOAD_SSHORT, atomic_load_sext_16_global, i32>;
+defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, atomic_load_zext_16_global, i32>;
defm : GlobalFLATLoadPats <GLOBAL_LOAD_USHORT, load_global, i16>;
foreach vt = Reg32Types.types in {
@@ -1525,6 +1540,7 @@ defm : GlobalFLATStorePats <GLOBAL_STORE_BYTE_D16_HI, truncstorei8_hi16_global,
}
let OtherPredicates = [D16PreservesUnusedBits] in {
+// TODO: Handle atomic loads
defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_global, v2i16>;
defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_UBYTE_D16_HI, az_extloadi8_d16_hi_global, v2f16>;
defm : GlobalFLATLoadPats_D16 <GLOBAL_LOAD_SBYTE_D16_HI, sextloadi8_d16_hi_global, v2i16>;
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index 087ca1f954464d..7095563de4a1b1 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -348,6 +348,18 @@ def load_glue : PatFrag <(ops node:$ptr), (unindexedload_glue node:$ptr)> {
let IsNonExtLoad = 1;
}
+def atomic_load_zext_glue :
+ PatFrag<(ops node:$ptr), (AMDGPUatomic_ld_glue node:$ptr)> {
+ let IsAtomic = true; // FIXME: Should be IsLoad and/or IsAtomic?
+ let IsZeroExtLoad = true;
+}
+
+def atomic_load_sext_glue :
+ PatFrag<(ops node:$ptr), (AMDGPUatomic_ld_glue node:$ptr)> {
+ let IsAtomic = true; // FIXME: Should be IsLoad and/or IsAtomic?
+ let IsSignExtLoad = true;
+}
+
def atomic_load_8_glue : PatFrag<(ops node:$ptr),
(AMDGPUatomic_ld_glue node:$ptr)> {
let IsAtomic = 1;
@@ -372,6 +384,30 @@ def atomic_load_64_glue : PatFrag<(ops node:$ptr),
let MemoryVT = i64;
}
+def atomic_load_zext_8_glue : PatFrag<(ops node:$ptr),
+ (atomic_load_zext_glue node:$ptr)> {
+ let IsAtomic = 1;
+ let MemoryVT = i8;
+}
+
+def atomic_load_sext_8_glue : PatFrag<(ops node:$ptr),
+ (atomic_load_sext_glue node:$ptr)> {
+ let IsAtomic = 1;
+ let MemoryVT = i8;
+}
+
+def atomic_load_zext_16_glue : PatFrag<(ops node:$ptr),
+ (atomic_load_zext_glue node:$ptr)> {
+ let IsAtomic = 1;
+ let MemoryVT = i16;
+}
+
+def atomic_load_sext_16_glue : PatFrag<(ops node:$ptr),
+ (atomic_load_sext_glue node:$ptr)> {
+ let IsAtomic = 1;
+ let MemoryVT = i16;
+}
+
def extload_glue : PatFrag<(ops node:$ptr), (unindexedload_glue node:$ptr)> {
let IsLoad = 1;
let IsAnyExtLoad = 1;
@@ -453,6 +489,15 @@ def atomic_load_32_local_m0 : PatFrag<(ops node:$ptr),
(atomic_load_32_glue node:$ptr)>;
def atomic_load_64_local_m0 : PatFrag<(ops node:$ptr),
(atomic_load_64_glue node:$ptr)>;
+
+def atomic_load_zext_8_local_m0 : PatFrag<(ops node:$ptr),
+ (atomic_load_zext_8_glue node:$ptr)>;
+def atomic_load_sext_8_local_m0 : PatFrag<(ops node:$ptr),
+ (atomic_load_sext_8_glue node:$ptr)>;
+def atomic_load_zext_16_local_m0 : PatFrag<(ops node:$ptr),
+ (atomic_load_zext_16_glue node:$ptr)>;
+def atomic_load_sext_16_local_m0 : PatFrag<(ops node:$ptr),
+ (atomic_load_sext_16_glue node:$ptr)>;
} // End let AddressSpaces = LoadAddress_local.AddrSpaces
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_flat.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_flat.ll
new file mode 100644
index 00000000000000..788fb04e842b4e
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_flat.ll
@@ -0,0 +1,331 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri < %s | FileCheck -check-prefixes=GCN,GFX7 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
+
+define i8 @atomic_load_flat_monotonic_i8(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_i8:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr %ptr monotonic, align 1
+ ret i8 %load
+}
+
+define i32 @atomic_load_flat_monotonic_i8_zext_to_i32(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_i8_zext_to_i32:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr %ptr monotonic, align 1
+ %ext = zext i8 %load to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_flat_monotonic_i8_sext_to_i32(ptr %ptr) {
+; GFX7-LABEL: atomic_load_flat_monotonic_i8_sext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_flat_monotonic_i8_sext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_flat_monotonic_i8_sext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX9-NEXT: flat_load_ubyte v3, v[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v0, v2
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr %ptr monotonic, align 1
+ %ext = sext i8 %load to i32
+ ret i32 %ext
+}
+
+define i16 @atomic_load_flat_monotonic_i8_zext_to_i16(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_i8_zext_to_i16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr %ptr monotonic, align 1
+ %ext = zext i8 %load to i16
+ ret i16 %ext
+}
+
+define i16 @atomic_load_flat_monotonic_i8_sext_to_i16(ptr %ptr) {
+; GFX7-LABEL: atomic_load_flat_monotonic_i8_sext_to_i16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_flat_monotonic_i8_sext_to_i16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_flat_monotonic_i8_sext_to_i16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX9-NEXT: flat_load_ubyte v3, v[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v0, v2
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr %ptr monotonic, align 1
+ %ext = sext i8 %load to i16
+ ret i16 %ext
+}
+
+define i16 @atomic_load_flat_monotonic_i16(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_i16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ushort v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr %ptr monotonic, align 2
+ ret i16 %load
+}
+
+define i32 @atomic_load_flat_monotonic_i16_zext_to_i32(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_i16_zext_to_i32:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_flat_monotonic_i16_sext_to_i32(ptr %ptr) {
+; GFX7-LABEL: atomic_load_flat_monotonic_i16_sext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX7-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_flat_monotonic_i16_sext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX8-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_flat_monotonic_i16_sext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX9-NEXT: flat_load_ushort v3, v[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v0, v2
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr %ptr monotonic, align 2
+ %ext = sext i16 %load to i32
+ ret i32 %ext
+}
+
+define half @atomic_load_flat_monotonic_f16(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_f16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ushort v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic half, ptr %ptr monotonic, align 2
+ ret half %load
+}
+
+define bfloat @atomic_load_flat_monotonic_bf16(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_bf16:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ushort v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic bfloat, ptr %ptr monotonic, align 2
+ ret bfloat %load
+}
+
+define i32 @atomic_load_flat_monotonic_f16_zext_to_i32(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_f16_zext_to_i32:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic half, ptr %ptr monotonic, align 2
+ %cast = bitcast half %load to i16
+ %ext = zext i16 %cast to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_flat_monotonic_bf16_zext_to_i32(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_bf16_zext_to_i32:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic bfloat, ptr %ptr monotonic, align 2
+ %cast = bitcast bfloat %load to i16
+ %ext = zext i16 %cast to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_flat_monotonic_i16_d16_hi_shift(ptr %ptr) {
+; GCN-LABEL: atomic_load_flat_monotonic_i16_d16_hi_shift:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ %shl = shl i32 %ext, 16
+ ret i32 %shl
+}
+
+define <2 x i16> @atomic_load_flat_monotonic_i16_d16_hi_vector_insert(ptr %ptr, <2 x i16> %vec) {
+; GFX7-LABEL: atomic_load_flat_monotonic_i16_d16_hi_vector_insert:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_flat_monotonic_i16_d16_hi_vector_insert:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_flat_monotonic_i16_d16_hi_vector_insert:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff
+; GFX9-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX9-NEXT: v_and_or_b32 v0, v2, v1, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr %ptr monotonic, align 2
+ %insert = insertelement <2 x i16> %vec, i16 %load, i32 1
+ ret <2 x i16> %insert
+}
+
+define i32 @atomic_load_flat_monotonic_i16_d16_lo_or(ptr %ptr, i16 %high) {
+; GFX7-LABEL: atomic_load_flat_monotonic_i16_d16_lo_or:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_flat_monotonic_i16_d16_lo_or:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_flat_monotonic_i16_d16_lo_or:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX9-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_lshl_or_b32 v0, v1, 16, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ %high.ext = zext i16 %high to i32
+ %shl = shl i32 %high.ext, 16
+ %or = or i32 %shl, %ext
+ ret i32 %or
+}
+
+define <2 x i16> @atomic_load_flat_monotonic_i16_d16_lo_vector_insert(ptr %ptr, <2 x i16> %vec) {
+; GFX7-LABEL: atomic_load_flat_monotonic_i16_d16_lo_vector_insert:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_flat_monotonic_i16_d16_lo_vector_insert:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_flat_monotonic_i16_d16_lo_vector_insert:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff0000
+; GFX9-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_and_or_b32 v0, v2, v1, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr %ptr monotonic, align 2
+ %insert = insertelement <2 x i16> %vec, i16 %load, i32 0
+ ret <2 x i16> %insert
+}
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_global.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_global.ll
new file mode 100644
index 00000000000000..139d841590f85a
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_global.ll
@@ -0,0 +1,478 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri < %s | FileCheck -check-prefixes=GCN,GFX7 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
+
+define i8 @atomic_load_global_monotonic_i8(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i8:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i8:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i8:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(1) %ptr monotonic, align 1
+ ret i8 %load
+}
+
+define i32 @atomic_load_global_monotonic_i8_zext_to_i32(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i8_zext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i8_zext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i8_zext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(1) %ptr monotonic, align 1
+ %ext = zext i8 %load to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_global_monotonic_i8_sext_to_i32(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i8_sext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i8_sext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i8_sext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_sbyte v2, v[0:1], off glc
+; GFX9-NEXT: global_load_ubyte v3, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_mov_b32_e32 v0, v2
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(1) %ptr monotonic, align 1
+ %ext = sext i8 %load to i32
+ ret i32 %ext
+}
+
+define i16 @atomic_load_global_monotonic_i8_zext_to_i16(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i8_zext_to_i16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i8_zext_to_i16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i8_zext_to_i16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(1) %ptr monotonic, align 1
+ %ext = zext i8 %load to i16
+ ret i16 %ext
+}
+
+define i16 @atomic_load_global_monotonic_i8_sext_to_i16(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i8_sext_to_i16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i8_sext_to_i16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i8_sext_to_i16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_sbyte v2, v[0:1], off glc
+; GFX9-NEXT: global_load_ubyte v3, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_mov_b32_e32 v0, v2
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(1) %ptr monotonic, align 1
+ %ext = sext i8 %load to i16
+ ret i16 %ext
+}
+
+define i16 @atomic_load_global_monotonic_i16(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ushort v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(1) %ptr monotonic, align 2
+ ret i16 %load
+}
+
+define i32 @atomic_load_global_monotonic_i16_zext_to_i32(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i16_zext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i16_zext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i16_zext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(1) %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_global_monotonic_i16_sext_to_i32(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i16_sext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX7-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v2
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i16_sext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_sbyte v2, v[0:1] glc
+; GFX8-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v2
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i16_sext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_sbyte v2, v[0:1], off glc
+; GFX9-NEXT: global_load_ushort v3, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(1)
+; GFX9-NEXT: v_mov_b32_e32 v0, v2
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(1) %ptr monotonic, align 2
+ %ext = sext i16 %load to i32
+ ret i32 %ext
+}
+
+define half @atomic_load_global_monotonic_f16(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_f16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ushort v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic half, ptr addrspace(1) %ptr monotonic, align 2
+ ret half %load
+}
+
+define bfloat @atomic_load_global_monotonic_bf16(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ushort v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ushort v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic bfloat, ptr addrspace(1) %ptr monotonic, align 2
+ ret bfloat %load
+}
+
+define i32 @atomic_load_global_monotonic_f16_zext_to_i32(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_f16_zext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_f16_zext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_f16_zext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic half, ptr addrspace(1) %ptr monotonic, align 2
+ %cast = bitcast half %load to i16
+ %ext = zext i16 %cast to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_global_monotonic_bf16_zext_to_i32(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_bf16_zext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_bf16_zext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_bf16_zext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic bfloat, ptr addrspace(1) %ptr monotonic, align 2
+ %cast = bitcast bfloat %load to i16
+ %ext = zext i16 %cast to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_global_monotonic_i16_d16_hi_shift(ptr addrspace(1) %ptr) {
+; GFX7-LABEL: atomic_load_global_monotonic_i16_d16_hi_shift:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i16_d16_hi_shift:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i16_d16_hi_shift:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(1) %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ %shl = shl i32 %ext, 16
+ ret i32 %shl
+}
+
+define <2 x i16> @atomic_load_global_monotonic_i16_d16_hi_vector_insert(ptr addrspace(1) %ptr, <2 x i16> %vec) {
+; GFX7-LABEL: atomic_load_global_monotonic_i16_d16_hi_vector_insert:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i16_d16_hi_vector_insert:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i16_d16_hi_vector_insert:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX9-NEXT: v_and_or_b32 v0, v2, v1, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(1) %ptr monotonic, align 2
+ %insert = insertelement <2 x i16> %vec, i16 %load, i32 1
+ ret <2 x i16> %insert
+}
+
+define i32 @atomic_load_global_monotonic_i16_d16_lo_or(ptr addrspace(1) %ptr, i16 %high) {
+; GFX7-LABEL: atomic_load_global_monotonic_i16_d16_lo_or:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i16_d16_lo_or:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i16_d16_lo_or:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_lshl_or_b32 v0, v1, 16, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(1) %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ %high.ext = zext i16 %high to i32
+ %shl = shl i32 %high.ext, 16
+ %or = or i32 %shl, %ext
+ ret i32 %or
+}
+
+define <2 x i16> @atomic_load_global_monotonic_i16_d16_lo_vector_insert(ptr addrspace(1) %ptr, <2 x i16> %vec) {
+; GFX7-LABEL: atomic_load_global_monotonic_i16_d16_lo_vector_insert:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v3
+; GFX7-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT: v_or_b32_e32 v1, v1, v2
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt vmcnt(0)
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_global_monotonic_i16_d16_lo_vector_insert:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: flat_load_ubyte v0, v[0:1] glc
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_global_monotonic_i16_d16_lo_vector_insert:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
+; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff0000
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_and_or_b32 v0, v2, v1, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(1) %ptr monotonic, align 2
+ %insert = insertelement <2 x i16> %vec, i16 %load, i32 0
+ ret <2 x i16> %insert
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; GCN: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_local_2.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_local_2.ll
new file mode 100644
index 00000000000000..5823bc3dfd3f5b
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/atomic_load_local_2.ll
@@ -0,0 +1,509 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri < %s | FileCheck -check-prefixes=GCN,GFX7 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8 %s
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
+
+; TODO: Merge with atomic_load_local.ll
+
+define i8 @atomic_load_local_monotonic_i8(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i8:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u8 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i8:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u8 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i8:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(3) %ptr monotonic, align 1
+ ret i8 %load
+}
+
+define i32 @atomic_load_local_monotonic_i8_zext_to_i32(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i8_zext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u8 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i8_zext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u8 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i8_zext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(3) %ptr monotonic, align 1
+ %ext = zext i8 %load to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_local_monotonic_i8_sext_to_i32(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i8_sext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_i8 v1, v0
+; GFX7-NEXT: ds_read_u8 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i8_sext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_i8 v1, v0
+; GFX8-NEXT: ds_read_u8 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i8_sext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_i8 v1, v0
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v0, v1
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(3) %ptr monotonic, align 1
+ %ext = sext i8 %load to i32
+ ret i32 %ext
+}
+
+define i16 @atomic_load_local_monotonic_i8_zext_to_i16(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i8_zext_to_i16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u8 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i8_zext_to_i16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u8 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i8_zext_to_i16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(3) %ptr monotonic, align 1
+ %ext = zext i8 %load to i16
+ ret i16 %ext
+}
+
+define i16 @atomic_load_local_monotonic_i8_sext_to_i16(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i8_sext_to_i16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_i8 v1, v0
+; GFX7-NEXT: ds_read_u8 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i8_sext_to_i16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_i8 v1, v0
+; GFX8-NEXT: ds_read_u8 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i8_sext_to_i16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_i8 v1, v0
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v0, v1
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i8, ptr addrspace(3) %ptr monotonic, align 1
+ %ext = sext i8 %load to i16
+ ret i16 %ext
+}
+
+define i16 @atomic_load_local_monotonic_i16(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u16 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(3) %ptr monotonic, align 2
+ ret i16 %load
+}
+
+define i32 @atomic_load_local_monotonic_i16_zext_to_i32(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i16_zext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i16_zext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i16_zext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(3) %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_local_monotonic_i16_sext_to_i32(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i16_sext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_i16 v1, v0
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_mov_b32_e32 v0, v1
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i16_sext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_i16 v1, v0
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_mov_b32_e32 v0, v1
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i16_sext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_i8 v1, v0
+; GFX9-NEXT: ds_read_u16 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_mov_b32_e32 v0, v1
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(3) %ptr monotonic, align 2
+ %ext = sext i16 %load to i32
+ ret i32 %ext
+}
+
+define half @atomic_load_local_monotonic_f16(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_f16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_f16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_f16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u16 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic half, ptr addrspace(3) %ptr monotonic, align 2
+ ret half %load
+}
+
+define bfloat @atomic_load_local_monotonic_bf16(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_bf16:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_bf16:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_bf16:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u16 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic bfloat, ptr addrspace(3) %ptr monotonic, align 2
+ ret bfloat %load
+}
+
+define i32 @atomic_load_local_monotonic_f16_zext_to_i32(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_f16_zext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_f16_zext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_f16_zext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic half, ptr addrspace(3) %ptr monotonic, align 2
+ %cast = bitcast half %load to i16
+ %ext = zext i16 %cast to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_local_monotonic_bf16_zext_to_i32(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_bf16_zext_to_i32:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_bf16_zext_to_i32:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_bf16_zext_to_i32:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic bfloat, ptr addrspace(3) %ptr monotonic, align 2
+ %cast = bitcast bfloat %load to i16
+ %ext = zext i16 %cast to i32
+ ret i32 %ext
+}
+
+define i32 @atomic_load_local_monotonic_i16_d16_hi_shift(ptr addrspace(3) %ptr) {
+; GFX7-LABEL: atomic_load_local_monotonic_i16_d16_hi_shift:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i16_d16_hi_shift:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i16_d16_hi_shift:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(3) %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ %shl = shl i32 %ext, 16
+ ret i32 %shl
+}
+
+define <2 x i16> @atomic_load_local_monotonic_i16_d16_hi_vector_insert(ptr addrspace(3) %ptr, <2 x i16> %vec) {
+; GFX7-LABEL: atomic_load_local_monotonic_i16_d16_hi_vector_insert:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v2, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i16_d16_hi_vector_insert:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i16_d16_hi_vector_insert:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: v_mov_b32_e32 v2, 0xffff
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
+; GFX9-NEXT: v_and_or_b32 v0, v1, v2, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(3) %ptr monotonic, align 2
+ %insert = insertelement <2 x i16> %vec, i16 %load, i32 1
+ ret <2 x i16> %insert
+}
+
+define i32 @atomic_load_local_monotonic_i16_d16_lo_or(ptr addrspace(3) %ptr, i16 %high) {
+; GFX7-LABEL: atomic_load_local_monotonic_i16_d16_lo_or:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i16_d16_lo_or:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i16_d16_lo_or:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_lshl_or_b32 v0, v1, 16, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(3) %ptr monotonic, align 2
+ %ext = zext i16 %load to i32
+ %high.ext = zext i16 %high to i32
+ %shl = shl i32 %high.ext, 16
+ %or = or i32 %shl, %ext
+ ret i32 %or
+}
+
+define <2 x i16> @atomic_load_local_monotonic_i16_d16_lo_vector_insert(ptr addrspace(3) %ptr, <2 x i16> %vec) {
+; GFX7-LABEL: atomic_load_local_monotonic_i16_d16_lo_vector_insert:
+; GFX7: ; %bb.0:
+; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT: s_mov_b32 m0, -1
+; GFX7-NEXT: ds_read_u16 v0, v0
+; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX7-NEXT: v_or_b32_e32 v1, v2, v1
+; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX7-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX7-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: atomic_load_local_monotonic_i16_d16_lo_vector_insert:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_mov_b32 m0, -1
+; GFX8-NEXT: ds_read_u16 v0, v0
+; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
+; GFX8-NEXT: s_waitcnt lgkmcnt(0)
+; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: atomic_load_local_monotonic_i16_d16_lo_vector_insert:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: ds_read_u8 v0, v0
+; GFX9-NEXT: v_mov_b32_e32 v2, 0xffff0000
+; GFX9-NEXT: s_waitcnt lgkmcnt(0)
+; GFX9-NEXT: v_and_or_b32 v0, v1, v2, v0
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %load = load atomic i16, ptr addrspace(3) %ptr monotonic, align 2
+ %insert = insertelement <2 x i16> %vec, i16 %load, i32 0
+ ret <2 x i16> %insert
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; GCN: {{.*}}
More information about the llvm-branch-commits
mailing list