[clang] [llvm] [NVPTX] Support __usAtomicCAS builtin (PR #99646)
via cfe-commits
cfe-commits at lists.llvm.org
Fri Jul 19 08:26:53 PDT 2024
https://github.com/DenisGZM updated https://github.com/llvm/llvm-project/pull/99646
>From 2cc142d85e706a2a4921e831fd78c8e609b70ae2 Mon Sep 17 00:00:00 2001
From: Denis Gerasimov <Denis.Gerasimov at baikalelectronics.ru>
Date: Fri, 19 Jul 2024 15:47:57 +0300
Subject: [PATCH 1/2] [NVPTX] Support __usAtomicCAS builtin
---
clang/include/clang/Basic/BuiltinsNVPTX.def | 3 +++
clang/lib/CodeGen/CGBuiltin.cpp | 3 +++
clang/lib/Headers/__clang_cuda_device_functions.h | 12 ++++++++++++
clang/test/CodeGen/builtins-nvptx.c | 3 +++
llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp | 2 +-
llvm/lib/Target/NVPTX/NVPTXIntrinsics.td | 15 +++++++++++++++
6 files changed, 37 insertions(+), 1 deletion(-)
diff --git a/clang/include/clang/Basic/BuiltinsNVPTX.def b/clang/include/clang/Basic/BuiltinsNVPTX.def
index 504314d8d96e9..3cf683a2f2129 100644
--- a/clang/include/clang/Basic/BuiltinsNVPTX.def
+++ b/clang/include/clang/Basic/BuiltinsNVPTX.def
@@ -829,6 +829,9 @@ BUILTIN(__nvvm_atom_xor_gen_ll, "LLiLLiD*LLi", "n")
TARGET_BUILTIN(__nvvm_atom_cta_xor_gen_ll, "LLiLLiD*LLi", "n", SM_60)
TARGET_BUILTIN(__nvvm_atom_sys_xor_gen_ll, "LLiLLiD*LLi", "n", SM_60)
+TARGET_BUILTIN(__nvvm_atom_cas_gen_us, "UsUsD*UsUs", "n", SM_70)
+TARGET_BUILTIN(__nvvm_atom_cta_cas_gen_us, "UsUsD*UsUs", "n", SM_70)
+TARGET_BUILTIN(__nvvm_atom_sys_cas_gen_us, "UsUsD*UsUs", "n", SM_70)
BUILTIN(__nvvm_atom_cas_gen_i, "iiD*ii", "n")
TARGET_BUILTIN(__nvvm_atom_cta_cas_gen_i, "iiD*ii", "n", SM_60)
TARGET_BUILTIN(__nvvm_atom_sys_cas_gen_i, "iiD*ii", "n", SM_60)
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 2d32a8582cfdf..9812a48ad4258 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -20100,6 +20100,7 @@ Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID,
case NVPTX::BI__nvvm_atom_min_gen_ull:
return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E);
+ case NVPTX::BI__nvvm_atom_cas_gen_us:
case NVPTX::BI__nvvm_atom_cas_gen_i:
case NVPTX::BI__nvvm_atom_cas_gen_l:
case NVPTX::BI__nvvm_atom_cas_gen_ll:
@@ -20291,6 +20292,7 @@ Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID,
case NVPTX::BI__nvvm_atom_sys_xor_gen_l:
case NVPTX::BI__nvvm_atom_sys_xor_gen_ll:
return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_sys, *this, E);
+ case NVPTX::BI__nvvm_atom_cta_cas_gen_us:
case NVPTX::BI__nvvm_atom_cta_cas_gen_i:
case NVPTX::BI__nvvm_atom_cta_cas_gen_l:
case NVPTX::BI__nvvm_atom_cta_cas_gen_ll: {
@@ -20302,6 +20304,7 @@ Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID,
Intrinsic::nvvm_atomic_cas_gen_i_cta, {ElemTy, Ptr->getType()}),
{Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))});
}
+ case NVPTX::BI__nvvm_atom_sys_cas_gen_us:
case NVPTX::BI__nvvm_atom_sys_cas_gen_i:
case NVPTX::BI__nvvm_atom_sys_cas_gen_l:
case NVPTX::BI__nvvm_atom_sys_cas_gen_ll: {
diff --git a/clang/lib/Headers/__clang_cuda_device_functions.h b/clang/lib/Headers/__clang_cuda_device_functions.h
index f8a12cefdb81b..f66fe625a3967 100644
--- a/clang/lib/Headers/__clang_cuda_device_functions.h
+++ b/clang/lib/Headers/__clang_cuda_device_functions.h
@@ -529,6 +529,18 @@ __DEVICE__ void __threadfence(void) { __nvvm_membar_gl(); }
__DEVICE__ void __threadfence_block(void) { __nvvm_membar_cta(); };
__DEVICE__ void __threadfence_system(void) { __nvvm_membar_sys(); };
__DEVICE__ void __trap(void) { __asm__ __volatile__("trap;"); }
+__DEVICE__ unsigned short __usAtomicCAS(unsigned short *__p, unsigned short __cmp,
+ unsigned short __v) {
+ return __nvvm_atom_cas_gen_us(__p, __cmp, __v);
+}
+__DEVICE__ unsigned short __usAtomicCAS_block(unsigned short *__p, unsigned short __cmp,
+ unsigned short __v) {
+ return __nvvm_atom_cta_cas_gen_us(__p, __cmp, __v);
+}
+__DEVICE__ unsigned short __usAtomicCAS_system(unsigned short *__p, unsigned short __cmp,
+ unsigned short __v) {
+ return __nvvm_atom_sys_cas_gen_us(__p, __cmp, __v);
+}
__DEVICE__ unsigned int __uAtomicAdd(unsigned int *__p, unsigned int __v) {
return __nvvm_atom_add_gen_i((int *)__p, __v);
}
diff --git a/clang/test/CodeGen/builtins-nvptx.c b/clang/test/CodeGen/builtins-nvptx.c
index 75b9d6d1fe190..3ba1fabd05335 100644
--- a/clang/test/CodeGen/builtins-nvptx.c
+++ b/clang/test/CodeGen/builtins-nvptx.c
@@ -306,6 +306,9 @@ __device__ void nvvm_atom(float *fp, float f, double *dfp, double df, int *ip,
// CHECK: atomicrmw umin ptr {{.*}} seq_cst, align 8
__nvvm_atom_min_gen_ull((unsigned long long *)&sll, ll);
+ // CHECK: cmpxchg ptr {{.*}} seq_cst seq_cst, align 2
+ // CHECK-NEXT: extractvalue { i16, i1 } {{%[0-9]+}}, 0
+ __nvvm_atom_cas_gen_us(ip, 0, i);
// CHECK: cmpxchg ptr {{.*}} seq_cst seq_cst, align 4
// CHECK-NEXT: extractvalue { i32, i1 } {{%[0-9]+}}, 0
__nvvm_atom_cas_gen_i(ip, 0, i);
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
index bc23998455a68..7c63f9215b55e 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -871,7 +871,7 @@ NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
// actions
computeRegisterProperties(STI.getRegisterInfo());
- setMinCmpXchgSizeInBits(32);
+ setMinCmpXchgSizeInBits(16);
setMaxAtomicSizeInBitsSupported(64);
setMaxDivRemBitWidthSupported(64);
}
diff --git a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
index c81dfa68e4bd4..e35eb8f55376b 100644
--- a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
+++ b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
@@ -1997,6 +1997,12 @@ defm INT_PTX_ATOM_XOR_GEN_64_USE_G : F_ATOMIC_2<i64, Int64Regs, ".global", ".b64
// atom_cas
+def atomic_cmp_swap_i16_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b, node:$c),
+ (atomic_cmp_swap_i16 node:$a, node:$b, node:$c)>;
+def atomic_cmp_swap_i16_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b, node:$c),
+ (atomic_cmp_swap_i16 node:$a, node:$b, node:$c)>;
+def atomic_cmp_swap_i16_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b, node:$c),
+ (atomic_cmp_swap_i16 node:$a, node:$b, node:$c)>;
def atomic_cmp_swap_i32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_i32 node:$a, node:$b, node:$c)>;
def atomic_cmp_swap_i32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b, node:$c),
@@ -2010,6 +2016,14 @@ def atomic_cmp_swap_i64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b, node:$c),
def atomic_cmp_swap_i64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_i64 node:$a, node:$b, node:$c)>;
+defm INT_PTX_ATOM_CAS_G_16 : F_ATOMIC_3<i16, Int16Regs, ".global", ".b16", ".cas",
+ atomic_cmp_swap_i16_g, i16imm>;
+defm INT_PTX_ATOM_CAS_S_16 : F_ATOMIC_3<i16, Int16Regs, ".shared", ".b16", ".cas",
+ atomic_cmp_swap_i16_s, i16imm>;
+defm INT_PTX_ATOM_CAS_GEN_16 : F_ATOMIC_3<i16, Int16Regs, "", ".b16", ".cas",
+ atomic_cmp_swap_i16_gen, i16imm>;
+defm INT_PTX_ATOM_CAS_GEN_16_USE_G : F_ATOMIC_3<i16, Int16Regs, ".global", ".b16", ".cas",
+ atomic_cmp_swap_i16_gen, i16imm>;
defm INT_PTX_ATOM_CAS_G_32 : F_ATOMIC_3<i32, Int32Regs, ".global", ".b32", ".cas",
atomic_cmp_swap_i32_g, i32imm>;
defm INT_PTX_ATOM_CAS_S_32 : F_ATOMIC_3<i32, Int32Regs, ".shared", ".b32", ".cas",
@@ -2221,6 +2235,7 @@ multiclass ATOM2_incdec_impl<string OpStr> {
// atom.cas
multiclass ATOM3_cas_impl<string OpStr> {
+ defm _b16 : ATOM3S_impl<OpStr, "i", "b16", i16, Int16Regs, i16imm, imm, i16, []>;
defm _b32 : ATOM3S_impl<OpStr, "i", "b32", i32, Int32Regs, i32imm, imm, i32, []>;
defm _b64 : ATOM3S_impl<OpStr, "i", "b64", i64, Int64Regs, i64imm, imm, i64, []>;
}
>From a7ccb4e573e119defaf6ce94836d910e8c9330cb Mon Sep 17 00:00:00 2001
From: Denis Gerasimov <Denis.Gerasimov at baikalelectronics.ru>
Date: Fri, 19 Jul 2024 18:26:41 +0300
Subject: [PATCH 2/2] Consider SM and PTX versions.
---
llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp | 3 +-
llvm/lib/Target/NVPTX/NVPTXIntrinsics.td | 8 +-
llvm/test/CodeGen/NVPTX/atomics-sm90.ll | 140 +++++++-------------
3 files changed, 56 insertions(+), 95 deletions(-)
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
index 7c63f9215b55e..c800c12bda7e1 100644
--- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -871,7 +871,8 @@ NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
// actions
computeRegisterProperties(STI.getRegisterInfo());
- setMinCmpXchgSizeInBits(16);
+ bool Allow16BitCAS = STI.getSmVersion() >= 70 && STI.getPTXVersion() >= 63;
+ setMinCmpXchgSizeInBits(Allow16BitCAS ? 16 : 32);
setMaxAtomicSizeInBitsSupported(64);
setMaxDivRemBitWidthSupported(64);
}
diff --git a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
index e35eb8f55376b..86f06bd8639f5 100644
--- a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
+++ b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
@@ -2017,13 +2017,13 @@ def atomic_cmp_swap_i64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_i64 node:$a, node:$b, node:$c)>;
defm INT_PTX_ATOM_CAS_G_16 : F_ATOMIC_3<i16, Int16Regs, ".global", ".b16", ".cas",
- atomic_cmp_swap_i16_g, i16imm>;
+ atomic_cmp_swap_i16_g, i16imm, [hasSM<70>, hasPTX<63>]>;
defm INT_PTX_ATOM_CAS_S_16 : F_ATOMIC_3<i16, Int16Regs, ".shared", ".b16", ".cas",
- atomic_cmp_swap_i16_s, i16imm>;
+ atomic_cmp_swap_i16_s, i16imm, [hasSM<70>, hasPTX<63>]>;
defm INT_PTX_ATOM_CAS_GEN_16 : F_ATOMIC_3<i16, Int16Regs, "", ".b16", ".cas",
- atomic_cmp_swap_i16_gen, i16imm>;
+ atomic_cmp_swap_i16_gen, i16imm, [hasSM<70>, hasPTX<63>]>;
defm INT_PTX_ATOM_CAS_GEN_16_USE_G : F_ATOMIC_3<i16, Int16Regs, ".global", ".b16", ".cas",
- atomic_cmp_swap_i16_gen, i16imm>;
+ atomic_cmp_swap_i16_gen, i16imm, [hasSM<70>, hasPTX<63>]>;
defm INT_PTX_ATOM_CAS_G_32 : F_ATOMIC_3<i32, Int32Regs, ".global", ".b32", ".cas",
atomic_cmp_swap_i32_g, i32imm>;
defm INT_PTX_ATOM_CAS_S_32 : F_ATOMIC_3<i32, Int32Regs, ".shared", ".b32", ".cas",
diff --git a/llvm/test/CodeGen/NVPTX/atomics-sm90.ll b/llvm/test/CodeGen/NVPTX/atomics-sm90.ll
index 9301ea44c6936..d69dd8ad1c940 100644
--- a/llvm/test/CodeGen/NVPTX/atomics-sm90.ll
+++ b/llvm/test/CodeGen/NVPTX/atomics-sm90.ll
@@ -45,102 +45,62 @@ define void @test(ptr %dp0, ptr addrspace(1) %dp1, ptr addrspace(3) %dp3, bfloat
;
; CHECKPTX71-LABEL: test(
; CHECKPTX71: {
-; CHECKPTX71-NEXT: .reg .pred %p<5>;
-; CHECKPTX71-NEXT: .reg .b16 %rs<18>;
-; CHECKPTX71-NEXT: .reg .b32 %r<58>;
-; CHECKPTX71-NEXT: .reg .f32 %f<12>;
+; CHECKPTX71-NEXT: .reg .pred %p<5>;
+; CHECKPTX71-NEXT: .reg .b16 %rs<34>;
+; CHECKPTX71-NEXT: .reg .b32 %r<4>;
+; CHECKPTX71-NEXT: .reg .f32 %f<12>;
; CHECKPTX71-EMPTY:
; CHECKPTX71-NEXT: // %bb.0:
-; CHECKPTX71-NEXT: ld.param.b16 %rs1, [test_param_3];
-; CHECKPTX71-NEXT: ld.param.u32 %r23, [test_param_2];
-; CHECKPTX71-NEXT: ld.param.u32 %r22, [test_param_1];
-; CHECKPTX71-NEXT: ld.param.u32 %r24, [test_param_0];
-; CHECKPTX71-NEXT: and.b32 %r1, %r24, -4;
-; CHECKPTX71-NEXT: and.b32 %r25, %r24, 3;
-; CHECKPTX71-NEXT: shl.b32 %r2, %r25, 3;
-; CHECKPTX71-NEXT: mov.b32 %r26, 65535;
-; CHECKPTX71-NEXT: shl.b32 %r27, %r26, %r2;
-; CHECKPTX71-NEXT: not.b32 %r3, %r27;
-; CHECKPTX71-NEXT: ld.u32 %r54, [%r1];
-; CHECKPTX71-NEXT: cvt.f32.bf16 %f2, %rs1;
-; CHECKPTX71-NEXT: $L__BB0_1: // %atomicrmw.start
+; CHECKPTX71-NEXT: ld.param.b16 %rs13, [test_param_3];
+; CHECKPTX71-NEXT: ld.param.u32 %r3, [test_param_2];
+; CHECKPTX71-NEXT: ld.param.u32 %r2, [test_param_1];
+; CHECKPTX71-NEXT: ld.param.u32 %r1, [test_param_0];
+; CHECKPTX71-NEXT: ld.b16 %rs30, [%r1];
+; CHECKPTX71-NEXT: cvt.f32.bf16 %f1, %rs13;
+; CHECKPTX71-NEXT: $L__BB0_1: // %atomicrmw.start
; CHECKPTX71-NEXT: // =>This Inner Loop Header: Depth=1
-; CHECKPTX71-NEXT: shr.u32 %r28, %r54, %r2;
-; CHECKPTX71-NEXT: cvt.u16.u32 %rs2, %r28;
-; CHECKPTX71-NEXT: cvt.f32.bf16 %f1, %rs2;
-; CHECKPTX71-NEXT: add.rn.f32 %f3, %f1, %f2;
-; CHECKPTX71-NEXT: cvt.rn.bf16.f32 %rs4, %f3;
-; CHECKPTX71-NEXT: cvt.u32.u16 %r29, %rs4;
-; CHECKPTX71-NEXT: shl.b32 %r30, %r29, %r2;
-; CHECKPTX71-NEXT: and.b32 %r31, %r54, %r3;
-; CHECKPTX71-NEXT: or.b32 %r32, %r31, %r30;
-; CHECKPTX71-NEXT: atom.cas.b32 %r6, [%r1], %r54, %r32;
-; CHECKPTX71-NEXT: setp.ne.s32 %p1, %r6, %r54;
-; CHECKPTX71-NEXT: mov.u32 %r54, %r6;
-; CHECKPTX71-NEXT: @%p1 bra $L__BB0_1;
-; CHECKPTX71-NEXT: // %bb.2: // %atomicrmw.end
-; CHECKPTX71-NEXT: ld.u32 %r55, [%r1];
-; CHECKPTX71-NEXT: $L__BB0_3: // %atomicrmw.start9
+; CHECKPTX71-NEXT: cvt.f32.bf16 %f2, %rs30;
+; CHECKPTX71-NEXT: add.rn.f32 %f3, %f2, %f1;
+; CHECKPTX71-NEXT: cvt.rn.bf16.f32 %rs14, %f3;
+; CHECKPTX71-NEXT: atom.cas.b16 %rs17, [%r1], %rs30, %rs14;
+; CHECKPTX71-NEXT: setp.ne.s16 %p1, %rs17, %rs30;
+; CHECKPTX71-NEXT: mov.u16 %rs30, %rs17;
+; CHECKPTX71-NEXT: @%p1 bra $L__BB0_1;
+; CHECKPTX71-NEXT: // %bb.2: // %atomicrmw.end
+; CHECKPTX71-NEXT: ld.b16 %rs31, [%r1];
+; CHECKPTX71-NEXT: $L__BB0_3: // %atomicrmw.start2
; CHECKPTX71-NEXT: // =>This Inner Loop Header: Depth=1
-; CHECKPTX71-NEXT: shr.u32 %r33, %r55, %r2;
-; CHECKPTX71-NEXT: cvt.u16.u32 %rs6, %r33;
-; CHECKPTX71-NEXT: cvt.f32.bf16 %f4, %rs6;
-; CHECKPTX71-NEXT: add.rn.f32 %f5, %f4, 0f3F800000;
-; CHECKPTX71-NEXT: cvt.rn.bf16.f32 %rs8, %f5;
-; CHECKPTX71-NEXT: cvt.u32.u16 %r34, %rs8;
-; CHECKPTX71-NEXT: shl.b32 %r35, %r34, %r2;
-; CHECKPTX71-NEXT: and.b32 %r36, %r55, %r3;
-; CHECKPTX71-NEXT: or.b32 %r37, %r36, %r35;
-; CHECKPTX71-NEXT: atom.cas.b32 %r9, [%r1], %r55, %r37;
-; CHECKPTX71-NEXT: setp.ne.s32 %p2, %r9, %r55;
-; CHECKPTX71-NEXT: mov.u32 %r55, %r9;
-; CHECKPTX71-NEXT: @%p2 bra $L__BB0_3;
-; CHECKPTX71-NEXT: // %bb.4: // %atomicrmw.end8
-; CHECKPTX71-NEXT: and.b32 %r10, %r22, -4;
-; CHECKPTX71-NEXT: shl.b32 %r38, %r22, 3;
-; CHECKPTX71-NEXT: and.b32 %r11, %r38, 24;
-; CHECKPTX71-NEXT: shl.b32 %r40, %r26, %r11;
-; CHECKPTX71-NEXT: not.b32 %r12, %r40;
-; CHECKPTX71-NEXT: ld.global.u32 %r56, [%r10];
-; CHECKPTX71-NEXT: $L__BB0_5: // %atomicrmw.start27
+; CHECKPTX71-NEXT: cvt.f32.bf16 %f4, %rs31;
+; CHECKPTX71-NEXT: add.rn.f32 %f5, %f4, 0f3F800000;
+; CHECKPTX71-NEXT: cvt.rn.bf16.f32 %rs18, %f5;
+; CHECKPTX71-NEXT: atom.cas.b16 %rs21, [%r1], %rs31, %rs18;
+; CHECKPTX71-NEXT: setp.ne.s16 %p2, %rs21, %rs31;
+; CHECKPTX71-NEXT: mov.u16 %rs31, %rs21;
+; CHECKPTX71-NEXT: @%p2 bra $L__BB0_3;
+; CHECKPTX71-NEXT: // %bb.4: // %atomicrmw.end1
+; CHECKPTX71-NEXT: ld.global.b16 %rs32, [%r2];
+; CHECKPTX71-NEXT: $L__BB0_5: // %atomicrmw.start8
; CHECKPTX71-NEXT: // =>This Inner Loop Header: Depth=1
-; CHECKPTX71-NEXT: shr.u32 %r41, %r56, %r11;
-; CHECKPTX71-NEXT: cvt.u16.u32 %rs10, %r41;
-; CHECKPTX71-NEXT: cvt.f32.bf16 %f6, %rs10;
-; CHECKPTX71-NEXT: add.rn.f32 %f8, %f6, %f2;
-; CHECKPTX71-NEXT: cvt.rn.bf16.f32 %rs12, %f8;
-; CHECKPTX71-NEXT: cvt.u32.u16 %r42, %rs12;
-; CHECKPTX71-NEXT: shl.b32 %r43, %r42, %r11;
-; CHECKPTX71-NEXT: and.b32 %r44, %r56, %r12;
-; CHECKPTX71-NEXT: or.b32 %r45, %r44, %r43;
-; CHECKPTX71-NEXT: atom.global.cas.b32 %r15, [%r10], %r56, %r45;
-; CHECKPTX71-NEXT: setp.ne.s32 %p3, %r15, %r56;
-; CHECKPTX71-NEXT: mov.u32 %r56, %r15;
-; CHECKPTX71-NEXT: @%p3 bra $L__BB0_5;
-; CHECKPTX71-NEXT: // %bb.6: // %atomicrmw.end26
-; CHECKPTX71-NEXT: and.b32 %r16, %r23, -4;
-; CHECKPTX71-NEXT: shl.b32 %r46, %r23, 3;
-; CHECKPTX71-NEXT: and.b32 %r17, %r46, 24;
-; CHECKPTX71-NEXT: shl.b32 %r48, %r26, %r17;
-; CHECKPTX71-NEXT: not.b32 %r18, %r48;
-; CHECKPTX71-NEXT: ld.shared.u32 %r57, [%r16];
-; CHECKPTX71-NEXT: $L__BB0_7: // %atomicrmw.start45
+; CHECKPTX71-NEXT: cvt.f32.bf16 %f7, %rs32;
+; CHECKPTX71-NEXT: add.rn.f32 %f8, %f7, %f1;
+; CHECKPTX71-NEXT: cvt.rn.bf16.f32 %rs22, %f8;
+; CHECKPTX71-NEXT: atom.global.cas.b16 %rs25, [%r2], %rs32, %rs22;
+; CHECKPTX71-NEXT: setp.ne.s16 %p3, %rs25, %rs32;
+; CHECKPTX71-NEXT: mov.u16 %rs32, %rs25;
+; CHECKPTX71-NEXT: @%p3 bra $L__BB0_5;
+; CHECKPTX71-NEXT: // %bb.6: // %atomicrmw.end7
+; CHECKPTX71-NEXT: ld.shared.b16 %rs33, [%r3];
+; CHECKPTX71-NEXT: $L__BB0_7: // %atomicrmw.start14
; CHECKPTX71-NEXT: // =>This Inner Loop Header: Depth=1
-; CHECKPTX71-NEXT: shr.u32 %r49, %r57, %r17;
-; CHECKPTX71-NEXT: cvt.u16.u32 %rs14, %r49;
-; CHECKPTX71-NEXT: cvt.f32.bf16 %f9, %rs14;
-; CHECKPTX71-NEXT: add.rn.f32 %f11, %f9, %f2;
-; CHECKPTX71-NEXT: cvt.rn.bf16.f32 %rs16, %f11;
-; CHECKPTX71-NEXT: cvt.u32.u16 %r50, %rs16;
-; CHECKPTX71-NEXT: shl.b32 %r51, %r50, %r17;
-; CHECKPTX71-NEXT: and.b32 %r52, %r57, %r18;
-; CHECKPTX71-NEXT: or.b32 %r53, %r52, %r51;
-; CHECKPTX71-NEXT: atom.shared.cas.b32 %r21, [%r16], %r57, %r53;
-; CHECKPTX71-NEXT: setp.ne.s32 %p4, %r21, %r57;
-; CHECKPTX71-NEXT: mov.u32 %r57, %r21;
-; CHECKPTX71-NEXT: @%p4 bra $L__BB0_7;
-; CHECKPTX71-NEXT: // %bb.8: // %atomicrmw.end44
-; CHECKPTX71-NEXT: ret;
+; CHECKPTX71-NEXT: cvt.f32.bf16 %f10, %rs33;
+; CHECKPTX71-NEXT: add.rn.f32 %f11, %f10, %f1;
+; CHECKPTX71-NEXT: cvt.rn.bf16.f32 %rs26, %f11;
+; CHECKPTX71-NEXT: atom.shared.cas.b16 %rs29, [%r3], %rs33, %rs26;
+; CHECKPTX71-NEXT: setp.ne.s16 %p4, %rs29, %rs33;
+; CHECKPTX71-NEXT: mov.u16 %rs33, %rs29;
+; CHECKPTX71-NEXT: @%p4 bra $L__BB0_7;
+; CHECKPTX71-NEXT: // %bb.8: // %atomicrmw.end13
+; CHECKPTX71-NEXT: ret;
%r1 = atomicrmw fadd ptr %dp0, bfloat %val seq_cst
%r2 = atomicrmw fadd ptr %dp0, bfloat 1.0 seq_cst
%r3 = atomicrmw fadd ptr addrspace(1) %dp1, bfloat %val seq_cst
More information about the cfe-commits
mailing list