[llvm] [NVPTX] Add support for atomic add for f16 type (PR #84295)

Adrian Kuegel via llvm-commits llvm-commits at lists.llvm.org
Tue Mar 12 00:13:47 PDT 2024


https://github.com/akuegel updated https://github.com/llvm/llvm-project/pull/84295

>From 7bac5a88e37b016d4797072aa40cb791792c8749 Mon Sep 17 00:00:00 2001
From: Adrian Kuegel <akuegel at google.com>
Date: Thu, 7 Mar 2024 09:56:57 +0000
Subject: [PATCH 1/6] [NVPTX] Add support for atomic add for f16 type.

atom.add.noftz.f16 is supported since SM 7.0
---
 llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp |  2 ++
 llvm/lib/Target/NVPTX/NVPTXIntrinsics.td    | 15 +++++++++++++++
 2 files changed, 17 insertions(+)

diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
index c979c03dc1b835..7e39fd0c81d86c 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -6100,6 +6100,8 @@ NVPTXTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
 
   if (AI->isFloatingPointOperation()) {
     if (AI->getOperation() == AtomicRMWInst::BinOp::FAdd) {
+      if (Ty->isHalfTy() && STI.getSmVersion() >= 70)
+        return AtomicExpansionKind::None;
       if (Ty->isFloatTy())
         return AtomicExpansionKind::None;
       if (Ty->isDoubleTy() && STI.hasAtomAddF64())
diff --git a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
index 477789a164ead2..31b5338b0d0f5d 100644
--- a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
+++ b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
@@ -1630,6 +1630,13 @@ defm INT_PTX_ATOM_ADD_GEN_64 : F_ATOMIC_2<i64, Int64Regs, "", ".u64", ".add",
 defm INT_PTX_ATOM_ADD_GEN_64_USE_G : F_ATOMIC_2<i64, Int64Regs, ".global", ".u64",
   ".add", atomic_load_add_64_gen, i64imm, imm>;
 
+defm INT_PTX_ATOM_ADD_G_F16 : F_ATOMIC_2<f16, Int16Regs, ".global", ".f16", ".add.noftz",
+  atomic_load_add_g, f16imm, fpimm, [hasSM<70>]>;
+defm INT_PTX_ATOM_ADD_S_F16 : F_ATOMIC_2<f16, Int16Regs, ".shared", ".f16", ".add.noftz",
+  atomic_load_add_s, f16imm, fpimm, [hasSM<70>]>;
+defm INT_PTX_ATOM_ADD_GEN_F16 : F_ATOMIC_2<f16, Int16Regs, "", ".f16", ".add.noftz",
+  atomic_load_add_gen, f16imm, fpimm, [hasSM<70>]>;
+
 defm INT_PTX_ATOM_ADD_G_F32 : F_ATOMIC_2<f32, Float32Regs, ".global", ".f32", ".add",
   atomic_load_add_g, f32imm, fpimm>;
 defm INT_PTX_ATOM_ADD_S_F32 : F_ATOMIC_2<f32, Float32Regs, ".shared", ".f32", ".add",
@@ -2007,6 +2014,9 @@ multiclass ATOM2P_impl<string AsmStr,  Intrinsic Intr,
                        SDNode Imm, ValueType ImmTy,
                        list<Predicate> Preds> {
   let AddedComplexity = 1 in {
+    def : ATOM23_impl<AsmStr, regT, regclass, Preds,
+                      (ins Int16Regs:$src, regclass:$b),
+                      (Intr (i16 Int16Regs:$src), (regT regclass:$b))>;
     def : ATOM23_impl<AsmStr, regT, regclass, Preds,
                       (ins Int32Regs:$src, regclass:$b),
                       (Intr (i32 Int32Regs:$src), (regT regclass:$b))>;
@@ -2017,6 +2027,9 @@ multiclass ATOM2P_impl<string AsmStr,  Intrinsic Intr,
   // tablegen can't infer argument types from Intrinsic (though it can
   // from Instruction) so we have to enforce specific type on
   // immediates via explicit cast to ImmTy.
+  def : ATOM23_impl<AsmStr, regT, regclass, Preds,
+                    (ins Int16Regs:$src, ImmType:$b),
+                    (Intr (i16 Int16Regs:$src), (ImmTy Imm:$b))>;
   def : ATOM23_impl<AsmStr, regT, regclass, Preds,
                     (ins Int32Regs:$src, ImmType:$b),
                     (Intr (i32 Int32Regs:$src), (ImmTy Imm:$b))>;
@@ -2136,6 +2149,8 @@ multiclass ATOM2_add_impl<string OpStr> {
    defm _s32  : ATOM2S_impl<OpStr, "i", "s32", i32, Int32Regs, i32imm, imm, i32, []>;
    defm _u32  : ATOM2S_impl<OpStr, "i", "u32", i32, Int32Regs, i32imm, imm, i32, []>;
    defm _u64  : ATOM2S_impl<OpStr, "i", "u64", i64, Int64Regs, i64imm, imm, i64, []>;
+   defm _f16  : ATOM2S_impl<OpStr, "f", "f16", f16, Int16Regs, f16imm, fpimm, f16,
+                            [hasSM<70>]>;
    defm _f32  : ATOM2S_impl<OpStr, "f", "f32", f32, Float32Regs, f32imm, fpimm, f32,
                             []>;
    defm _f64  : ATOM2S_impl<OpStr, "f", "f64", f64, Float64Regs, f64imm, fpimm, f64,

>From 9aefbbee78df14eedac9ffce4238ed520c7a2fe2 Mon Sep 17 00:00:00 2001
From: Adrian Kuegel <akuegel at google.com>
Date: Thu, 7 Mar 2024 10:00:19 +0000
Subject: [PATCH 2/6] [NVPTX] Add atomic add test.

---
 llvm/test/CodeGen/NVPTX/atomics-sm70.ll | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
 create mode 100644 llvm/test/CodeGen/NVPTX/atomics-sm70.ll

diff --git a/llvm/test/CodeGen/NVPTX/atomics-sm70.ll b/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
new file mode 100644
index 00000000000000..71d9d7b62f9107
--- /dev/null
+++ b/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
@@ -0,0 +1,17 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_70 | FileCheck %s
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 | FileCheck %s
+; RUN: %if ptxas && !ptxas-12.0 %{ llc < %s -march=nvptx -mcpu=sm_70 | %ptxas-verify -arch=sm_70 %}
+; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_70 | %ptxas-verify -arch=sm_70 %}
+
+; CHECK-LABEL: .func test(
+define void @test(ptr %dp0, ptr addrspace(1) %dp1, ptr addrspace(3) %dp3, half %d) {
+; CHECK: atom.add.noftz.f16
+  %r1 = atomicrmw fadd ptr %dp0, half %d seq_cst
+; CHECK: atom.global.add.noftz.f16
+  %r2 = atomicrmw fadd ptr addrspace(1) %dp1, half %d seq_cst
+; CHECK: atom.shared.add.noftz.f16
+  %ret = atomicrmw fadd ptr addrspace(3) %dp3, half %d seq_cst
+  ret void
+}
+
+attributes #1 = { argmemonly nounwind }

>From dbd3f9a2b5707246932a17c3271298ba2e53fa1a Mon Sep 17 00:00:00 2001
From: Adrian Kuegel <akuegel at google.com>
Date: Fri, 8 Mar 2024 07:37:35 +0000
Subject: [PATCH 3/6] Also check PTX version.

---
 llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp | 3 ++-
 llvm/lib/Target/NVPTX/NVPTXIntrinsics.td    | 8 ++++----
 2 files changed, 6 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
index 7e39fd0c81d86c..c411c8ef9528d7 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -6100,7 +6100,8 @@ NVPTXTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
 
   if (AI->isFloatingPointOperation()) {
     if (AI->getOperation() == AtomicRMWInst::BinOp::FAdd) {
-      if (Ty->isHalfTy() && STI.getSmVersion() >= 70)
+      if (Ty->isHalfTy() && STI.getSmVersion() >= 70 &&
+          STI.getPTXVersion() >= 63)
         return AtomicExpansionKind::None;
       if (Ty->isFloatTy())
         return AtomicExpansionKind::None;
diff --git a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
index 31b5338b0d0f5d..869b13369e87e1 100644
--- a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
+++ b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
@@ -1631,11 +1631,11 @@ defm INT_PTX_ATOM_ADD_GEN_64_USE_G : F_ATOMIC_2<i64, Int64Regs, ".global", ".u64
   ".add", atomic_load_add_64_gen, i64imm, imm>;
 
 defm INT_PTX_ATOM_ADD_G_F16 : F_ATOMIC_2<f16, Int16Regs, ".global", ".f16", ".add.noftz",
-  atomic_load_add_g, f16imm, fpimm, [hasSM<70>]>;
+  atomic_load_add_g, f16imm, fpimm, [hasSM<70>, hasPTX<63>]>;
 defm INT_PTX_ATOM_ADD_S_F16 : F_ATOMIC_2<f16, Int16Regs, ".shared", ".f16", ".add.noftz",
-  atomic_load_add_s, f16imm, fpimm, [hasSM<70>]>;
+  atomic_load_add_s, f16imm, fpimm, [hasSM<70>, hasPTX<63>]>;
 defm INT_PTX_ATOM_ADD_GEN_F16 : F_ATOMIC_2<f16, Int16Regs, "", ".f16", ".add.noftz",
-  atomic_load_add_gen, f16imm, fpimm, [hasSM<70>]>;
+  atomic_load_add_gen, f16imm, fpimm, [hasSM<70>, hasPTX<63>]>;
 
 defm INT_PTX_ATOM_ADD_G_F32 : F_ATOMIC_2<f32, Float32Regs, ".global", ".f32", ".add",
   atomic_load_add_g, f32imm, fpimm>;
@@ -2150,7 +2150,7 @@ multiclass ATOM2_add_impl<string OpStr> {
    defm _u32  : ATOM2S_impl<OpStr, "i", "u32", i32, Int32Regs, i32imm, imm, i32, []>;
    defm _u64  : ATOM2S_impl<OpStr, "i", "u64", i64, Int64Regs, i64imm, imm, i64, []>;
    defm _f16  : ATOM2S_impl<OpStr, "f", "f16", f16, Int16Regs, f16imm, fpimm, f16,
-                            [hasSM<70>]>;
+                            [hasSM<70>, hasPTX<63>]>;
    defm _f32  : ATOM2S_impl<OpStr, "f", "f32", f32, Float32Regs, f32imm, fpimm, f32,
                             []>;
    defm _f64  : ATOM2S_impl<OpStr, "f", "f64", f64, Float64Regs, f64imm, fpimm, f64,

>From 9765a9a1362abe3536c6888b7d44ab1afbaecfb8 Mon Sep 17 00:00:00 2001
From: Adrian Kuegel <akuegel at google.com>
Date: Mon, 11 Mar 2024 07:05:11 +0000
Subject: [PATCH 4/6] Update tests and add more tests

---
 llvm/test/CodeGen/NVPTX/atomics-sm70.ll | 27 +++++++++++++++++--------
 llvm/test/CodeGen/NVPTX/atomics.ll      |  7 +++++++
 2 files changed, 26 insertions(+), 8 deletions(-)

diff --git a/llvm/test/CodeGen/NVPTX/atomics-sm70.ll b/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
index 71d9d7b62f9107..e43cb6f9dec524 100644
--- a/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
+++ b/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
@@ -1,16 +1,27 @@
-; RUN: llc < %s -march=nvptx -mcpu=sm_70 | FileCheck %s
-; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 | FileCheck %s
-; RUN: %if ptxas && !ptxas-12.0 %{ llc < %s -march=nvptx -mcpu=sm_70 | %ptxas-verify -arch=sm_70 %}
-; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_70 | %ptxas-verify -arch=sm_70 %}
+; RUN: llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s --dump-input=always
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s
+; RUN: llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx62 | FileCheck %s --check-prefixes=CHECKPTX62
+; RUN: %if ptxas && !ptxas-12.0 %{ llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx63 | %ptxas-verify -arch=sm_70 %}
+; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | %ptxas-verify -arch=sm_70 %}
+; RUN: %if ptxas && !ptxas-12.0 %{ llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx62 | %ptxas-verify -arch=sm_70 %}
 
 ; CHECK-LABEL: .func test(
-define void @test(ptr %dp0, ptr addrspace(1) %dp1, ptr addrspace(3) %dp3, half %d) {
+define void @test(ptr %dp0, ptr addrspace(1) %dp1, ptr addrspace(3) %dp3, half %val) {
 ; CHECK: atom.add.noftz.f16
-  %r1 = atomicrmw fadd ptr %dp0, half %d seq_cst
+; CHECK-NOT: atom.cas
+; CHECKPTX62: atom.cas
+; CHECKPTX62-NOT: atom.add.noftz.f16
+  %r1 = atomicrmw fadd ptr %dp0, half %val seq_cst
 ; CHECK: atom.global.add.noftz.f16
-  %r2 = atomicrmw fadd ptr addrspace(1) %dp1, half %d seq_cst
+; CHECK-NOT: atom.global.cas
+; CHECKPTX62: atom.global.cas
+; CHECKPTX62-NOT: atom.global.add.noftz.f16
+  %r2 = atomicrmw fadd ptr addrspace(1) %dp1, half %val seq_cst
 ; CHECK: atom.shared.add.noftz.f16
-  %ret = atomicrmw fadd ptr addrspace(3) %dp3, half %d seq_cst
+; CHECK-NOT: atom.shared.cas
+; CHECKPTX62: atom.shared.cas
+; CHECKPTX62-NOT: atom.shared.add.noftz.f16
+  %ret = atomicrmw fadd ptr addrspace(3) %dp3, half %val seq_cst
   ret void
 }
 
diff --git a/llvm/test/CodeGen/NVPTX/atomics.ll b/llvm/test/CodeGen/NVPTX/atomics.ll
index e99d0fd05e346b..6f2b5dcf47f13b 100644
--- a/llvm/test/CodeGen/NVPTX/atomics.ll
+++ b/llvm/test/CodeGen/NVPTX/atomics.ll
@@ -175,6 +175,13 @@ define float @atomicrmw_add_f32_generic(ptr %addr, float %val) {
   ret float %ret
 }
 
+; CHECK-LABEL: atomicrmw_add_f16_generic
+define half @atomicrmw_add_f16_generic(ptr %addr, half %val) {
+; CHECK: atom.cas
+  %ret = atomicrmw fadd ptr %addr, half %val seq_cst
+  ret half %ret
+}
+
 ; CHECK-LABEL: atomicrmw_add_f32_addrspace1
 define float @atomicrmw_add_f32_addrspace1(ptr addrspace(1) %addr, float %val) {
 ; CHECK: atom.global.add.f32

>From 3bd7abaa7f1582f6f2706aac14bd8026925d7730 Mon Sep 17 00:00:00 2001
From: Adrian Kuegel <akuegel at google.com>
Date: Mon, 11 Mar 2024 07:11:18 +0000
Subject: [PATCH 5/6] Remove dump-input=always left over from debugging.

---
 llvm/test/CodeGen/NVPTX/atomics-sm70.ll | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/test/CodeGen/NVPTX/atomics-sm70.ll b/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
index e43cb6f9dec524..e1ae9661cb1617 100644
--- a/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
+++ b/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s --dump-input=always
+; RUN: llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s
 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s
 ; RUN: llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx62 | FileCheck %s --check-prefixes=CHECKPTX62
 ; RUN: %if ptxas && !ptxas-12.0 %{ llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx63 | %ptxas-verify -arch=sm_70 %}

>From 34ce80d71b8bf6bd0f8755ccbb2b40554a61e706 Mon Sep 17 00:00:00 2001
From: Adrian Kuegel <akuegel at google.com>
Date: Tue, 12 Mar 2024 07:13:12 +0000
Subject: [PATCH 6/6] Autogenerate tests.

---
 llvm/test/CodeGen/NVPTX/atomics-sm70.ll | 123 +++++++++++++++++++++---
 1 file changed, 108 insertions(+), 15 deletions(-)

diff --git a/llvm/test/CodeGen/NVPTX/atomics-sm70.ll b/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
index e1ae9661cb1617..f68bb2049d0061 100644
--- a/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
+++ b/llvm/test/CodeGen/NVPTX/atomics-sm70.ll
@@ -1,26 +1,119 @@
-; RUN: llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s
-; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s --check-prefixes=CHECK
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s --check-prefixes=CHECK64
 ; RUN: llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx62 | FileCheck %s --check-prefixes=CHECKPTX62
 ; RUN: %if ptxas && !ptxas-12.0 %{ llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx63 | %ptxas-verify -arch=sm_70 %}
 ; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | %ptxas-verify -arch=sm_70 %}
 ; RUN: %if ptxas && !ptxas-12.0 %{ llc < %s -march=nvptx -mcpu=sm_70 -mattr=+ptx62 | %ptxas-verify -arch=sm_70 %}
 
-; CHECK-LABEL: .func test(
+target triple = "nvptx64-nvidia-cuda"
+
 define void @test(ptr %dp0, ptr addrspace(1) %dp1, ptr addrspace(3) %dp3, half %val) {
-; CHECK: atom.add.noftz.f16
-; CHECK-NOT: atom.cas
-; CHECKPTX62: atom.cas
-; CHECKPTX62-NOT: atom.add.noftz.f16
+; CHECK-LABEL: test(
+; CHECK:       {
+; CHECK-NEXT:    .reg .b16 %rs<5>;
+; CHECK-NEXT:    .reg .b32 %r<4>;
+; CHECK-EMPTY:
+; CHECK-NEXT:  // %bb.0:
+; CHECK-NEXT:    ld.param.u32 %r1, [test_param_0];
+; CHECK-NEXT:    ld.param.b16 %rs1, [test_param_3];
+; CHECK-NEXT:    atom.add.noftz.f16 %rs2, [%r1], %rs1;
+; CHECK-NEXT:    ld.param.u32 %r2, [test_param_1];
+; CHECK-NEXT:    atom.global.add.noftz.f16 %rs3, [%r2], %rs1;
+; CHECK-NEXT:    ld.param.u32 %r3, [test_param_2];
+; CHECK-NEXT:    atom.shared.add.noftz.f16 %rs4, [%r3], %rs1;
+; CHECK-NEXT:    ret;
+;
+; CHECK64-LABEL: test(
+; CHECK64:       {
+; CHECK64-NEXT:    .reg .b16 %rs<5>;
+; CHECK64-NEXT:    .reg .b64 %rd<4>;
+; CHECK64-EMPTY:
+; CHECK64-NEXT:  // %bb.0:
+; CHECK64-NEXT:    ld.param.u64 %rd1, [test_param_0];
+; CHECK64-NEXT:    ld.param.b16 %rs1, [test_param_3];
+; CHECK64-NEXT:    atom.add.noftz.f16 %rs2, [%rd1], %rs1;
+; CHECK64-NEXT:    ld.param.u64 %rd2, [test_param_1];
+; CHECK64-NEXT:    atom.global.add.noftz.f16 %rs3, [%rd2], %rs1;
+; CHECK64-NEXT:    ld.param.u64 %rd3, [test_param_2];
+; CHECK64-NEXT:    atom.shared.add.noftz.f16 %rs4, [%rd3], %rs1;
+; CHECK64-NEXT:    ret;
+;
+; CHECKPTX62-LABEL: test(
+; CHECKPTX62:       {
+; CHECKPTX62-NEXT:    .reg .pred %p<4>;
+; CHECKPTX62-NEXT:    .reg .b16 %rs<14>;
+; CHECKPTX62-NEXT:    .reg .b32 %r<49>;
+; CHECKPTX62-EMPTY:
+; CHECKPTX62-NEXT:  // %bb.0:
+; CHECKPTX62-NEXT:    ld.param.b16 %rs1, [test_param_3];
+; CHECKPTX62-NEXT:    ld.param.u32 %r20, [test_param_2];
+; CHECKPTX62-NEXT:    ld.param.u32 %r19, [test_param_1];
+; CHECKPTX62-NEXT:    ld.param.u32 %r21, [test_param_0];
+; CHECKPTX62-NEXT:    and.b32 %r1, %r21, -4;
+; CHECKPTX62-NEXT:    and.b32 %r22, %r21, 3;
+; CHECKPTX62-NEXT:    shl.b32 %r2, %r22, 3;
+; CHECKPTX62-NEXT:    mov.b32 %r23, 65535;
+; CHECKPTX62-NEXT:    shl.b32 %r24, %r23, %r2;
+; CHECKPTX62-NEXT:    not.b32 %r3, %r24;
+; CHECKPTX62-NEXT:    ld.u32 %r46, [%r1];
+; CHECKPTX62-NEXT:  $L__BB0_1: // %atomicrmw.start
+; CHECKPTX62-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECKPTX62-NEXT:    shr.u32 %r25, %r46, %r2;
+; CHECKPTX62-NEXT:    cvt.u16.u32 %rs2, %r25;
+; CHECKPTX62-NEXT:    add.rn.f16 %rs4, %rs2, %rs1;
+; CHECKPTX62-NEXT:    cvt.u32.u16 %r26, %rs4;
+; CHECKPTX62-NEXT:    shl.b32 %r27, %r26, %r2;
+; CHECKPTX62-NEXT:    and.b32 %r28, %r46, %r3;
+; CHECKPTX62-NEXT:    or.b32 %r29, %r28, %r27;
+; CHECKPTX62-NEXT:    atom.cas.b32 %r6, [%r1], %r46, %r29;
+; CHECKPTX62-NEXT:    setp.ne.s32 %p1, %r6, %r46;
+; CHECKPTX62-NEXT:    mov.u32 %r46, %r6;
+; CHECKPTX62-NEXT:    @%p1 bra $L__BB0_1;
+; CHECKPTX62-NEXT:  // %bb.2: // %atomicrmw.end
+; CHECKPTX62-NEXT:    and.b32 %r7, %r19, -4;
+; CHECKPTX62-NEXT:    shl.b32 %r30, %r19, 3;
+; CHECKPTX62-NEXT:    and.b32 %r8, %r30, 24;
+; CHECKPTX62-NEXT:    shl.b32 %r32, %r23, %r8;
+; CHECKPTX62-NEXT:    not.b32 %r9, %r32;
+; CHECKPTX62-NEXT:    ld.global.u32 %r47, [%r7];
+; CHECKPTX62-NEXT:  $L__BB0_3: // %atomicrmw.start9
+; CHECKPTX62-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECKPTX62-NEXT:    shr.u32 %r33, %r47, %r8;
+; CHECKPTX62-NEXT:    cvt.u16.u32 %rs6, %r33;
+; CHECKPTX62-NEXT:    add.rn.f16 %rs8, %rs6, %rs1;
+; CHECKPTX62-NEXT:    cvt.u32.u16 %r34, %rs8;
+; CHECKPTX62-NEXT:    shl.b32 %r35, %r34, %r8;
+; CHECKPTX62-NEXT:    and.b32 %r36, %r47, %r9;
+; CHECKPTX62-NEXT:    or.b32 %r37, %r36, %r35;
+; CHECKPTX62-NEXT:    atom.global.cas.b32 %r12, [%r7], %r47, %r37;
+; CHECKPTX62-NEXT:    setp.ne.s32 %p2, %r12, %r47;
+; CHECKPTX62-NEXT:    mov.u32 %r47, %r12;
+; CHECKPTX62-NEXT:    @%p2 bra $L__BB0_3;
+; CHECKPTX62-NEXT:  // %bb.4: // %atomicrmw.end8
+; CHECKPTX62-NEXT:    and.b32 %r13, %r20, -4;
+; CHECKPTX62-NEXT:    shl.b32 %r38, %r20, 3;
+; CHECKPTX62-NEXT:    and.b32 %r14, %r38, 24;
+; CHECKPTX62-NEXT:    shl.b32 %r40, %r23, %r14;
+; CHECKPTX62-NEXT:    not.b32 %r15, %r40;
+; CHECKPTX62-NEXT:    ld.shared.u32 %r48, [%r13];
+; CHECKPTX62-NEXT:  $L__BB0_5: // %atomicrmw.start27
+; CHECKPTX62-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECKPTX62-NEXT:    shr.u32 %r41, %r48, %r14;
+; CHECKPTX62-NEXT:    cvt.u16.u32 %rs10, %r41;
+; CHECKPTX62-NEXT:    add.rn.f16 %rs12, %rs10, %rs1;
+; CHECKPTX62-NEXT:    cvt.u32.u16 %r42, %rs12;
+; CHECKPTX62-NEXT:    shl.b32 %r43, %r42, %r14;
+; CHECKPTX62-NEXT:    and.b32 %r44, %r48, %r15;
+; CHECKPTX62-NEXT:    or.b32 %r45, %r44, %r43;
+; CHECKPTX62-NEXT:    atom.shared.cas.b32 %r18, [%r13], %r48, %r45;
+; CHECKPTX62-NEXT:    setp.ne.s32 %p3, %r18, %r48;
+; CHECKPTX62-NEXT:    mov.u32 %r48, %r18;
+; CHECKPTX62-NEXT:    @%p3 bra $L__BB0_5;
+; CHECKPTX62-NEXT:  // %bb.6: // %atomicrmw.end26
+; CHECKPTX62-NEXT:    ret;
   %r1 = atomicrmw fadd ptr %dp0, half %val seq_cst
-; CHECK: atom.global.add.noftz.f16
-; CHECK-NOT: atom.global.cas
-; CHECKPTX62: atom.global.cas
-; CHECKPTX62-NOT: atom.global.add.noftz.f16
   %r2 = atomicrmw fadd ptr addrspace(1) %dp1, half %val seq_cst
-; CHECK: atom.shared.add.noftz.f16
-; CHECK-NOT: atom.shared.cas
-; CHECKPTX62: atom.shared.cas
-; CHECKPTX62-NOT: atom.shared.add.noftz.f16
   %ret = atomicrmw fadd ptr addrspace(3) %dp3, half %val seq_cst
   ret void
 }



More information about the llvm-commits mailing list