[PATCH 5/9] R600/SI: Add global atomicrmw max/umax
Aaron Watry
awatry at gmail.com
Thu Oct 16 18:01:29 PDT 2014
v2: Add separate offset/no-offset tests
Signed-off-by: Aaron Watry <awatry at gmail.com>
Reviewed-by (v1): Matt Arsenault <matthew.arsenault at amd.com>
---
lib/Target/R600/AMDGPUInstructions.td | 2 +
lib/Target/R600/SIInstructions.td | 8 +-
test/CodeGen/R600/global_atomics.ll | 160 ++++++++++++++++++++++++++++++++++
3 files changed, 168 insertions(+), 2 deletions(-)
diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td
index 972ef1d..f295f33 100644
--- a/lib/Target/R600/AMDGPUInstructions.td
+++ b/lib/Target/R600/AMDGPUInstructions.td
@@ -388,7 +388,9 @@ class global_binary_atomic_op<SDNode atomic_op> : PatFrag<
def atomic_add_global : global_binary_atomic_op<atomic_load_add>;
def atomic_and_global : global_binary_atomic_op<atomic_load_and>;
+def atomic_max_global : global_binary_atomic_op<atomic_load_max>;
def atomic_sub_global : global_binary_atomic_op<atomic_load_sub>;
+def atomic_umax_global : global_binary_atomic_op<atomic_load_umax>;
//===----------------------------------------------------------------------===//
// Misc Pattern Fragments
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 2f0a74c..5a7cdf3 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -906,8 +906,12 @@ defm BUFFER_ATOMIC_SUB : MUBUF_Atomic <
//def BUFFER_ATOMIC_RSUB : MUBUF_ <0x00000034, "BUFFER_ATOMIC_RSUB", []>;
//def BUFFER_ATOMIC_SMIN : MUBUF_ <0x00000035, "BUFFER_ATOMIC_SMIN", []>;
//def BUFFER_ATOMIC_UMIN : MUBUF_ <0x00000036, "BUFFER_ATOMIC_UMIN", []>;
-//def BUFFER_ATOMIC_SMAX : MUBUF_ <0x00000037, "BUFFER_ATOMIC_SMAX", []>;
-//def BUFFER_ATOMIC_UMAX : MUBUF_ <0x00000038, "BUFFER_ATOMIC_UMAX", []>;
+defm BUFFER_ATOMIC_SMAX : MUBUF_Atomic <
+ 0x00000037, "BUFFER_ATOMIC_SMAX", VReg_32, i32, atomic_max_global
+>;
+defm BUFFER_ATOMIC_UMAX : MUBUF_Atomic <
+ 0x00000038, "BUFFER_ATOMIC_UMAX", VReg_32, i32, atomic_umax_global
+>;
defm BUFFER_ATOMIC_AND : MUBUF_Atomic <
0x00000039, "BUFFER_ATOMIC_AND", VReg_32, i32, atomic_and_global
>;
diff --git a/test/CodeGen/R600/global_atomics.ll b/test/CodeGen/R600/global_atomics.ll
index 5feecfc..93dba4a 100644
--- a/test/CodeGen/R600/global_atomics.ll
+++ b/test/CodeGen/R600/global_atomics.ll
@@ -239,3 +239,163 @@ entry:
store i32 %0, i32 addrspace(1)* %out2
ret void
}
+
+; FUNC-LABEL: {{^}}atomic_max_i32_offset:
+; SI: BUFFER_ATOMIC_SMAX v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offset:0x10{{$}}
+define void @atomic_max_i32_offset(i32 addrspace(1)* %out, i32 %in) {
+entry:
+ %gep = getelementptr i32 addrspace(1)* %out, i32 4
+ %0 = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_max_i32_ret_offset:
+; SI: BUFFER_ATOMIC_SMAX [[RET:v[0-9]+]], s[{{[0-9]+}}:{{[0-9]+}}], 0 offset:0x10 glc {{$}}
+; SI: BUFFER_STORE_DWORD [[RET]]
+define void @atomic_max_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+entry:
+ %gep = getelementptr i32 addrspace(1)* %out, i32 4
+ %0 = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+ store i32 %0, i32 addrspace(1)* %out2
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_max_i32_addr64_offset:
+; SI: BUFFER_ATOMIC_SMAX v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], s[{{[0-9]+}}:{{[0-9]+}}], 0 addr64 offset:0x10{{$}}
+define void @atomic_max_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+entry:
+ %ptr = getelementptr i32 addrspace(1)* %out, i64 %index
+ %gep = getelementptr i32 addrspace(1)* %ptr, i32 4
+ %0 = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_max_i32_ret_addr64_offset:
+; SI: BUFFER_ATOMIC_SMAX [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}], s[{{[0-9]+}}:{{[0-9]+}}], 0 addr64 offset:0x10 glc{{$}}
+; SI: BUFFER_STORE_DWORD [[RET]]
+define void @atomic_max_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+entry:
+ %ptr = getelementptr i32 addrspace(1)* %out, i64 %index
+ %gep = getelementptr i32 addrspace(1)* %ptr, i32 4
+ %0 = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
+ store i32 %0, i32 addrspace(1)* %out2
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_max_i32:
+; SI: BUFFER_ATOMIC_SMAX v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0{{$}}
+define void @atomic_max_i32(i32 addrspace(1)* %out, i32 %in) {
+entry:
+ %0 = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in seq_cst
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_max_i32_ret:
+; SI: BUFFER_ATOMIC_SMAX [[RET:v[0-9]+]], s[{{[0-9]+}}:{{[0-9]+}}], 0 glc
+; SI: BUFFER_STORE_DWORD [[RET]]
+define void @atomic_max_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+entry:
+ %0 = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in seq_cst
+ store i32 %0, i32 addrspace(1)* %out2
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_max_i32_addr64:
+; SI: BUFFER_ATOMIC_SMAX v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], s[{{[0-9]+}}:{{[0-9]+}}], 0 addr64{{$}}
+define void @atomic_max_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+entry:
+ %ptr = getelementptr i32 addrspace(1)* %out, i64 %index
+ %0 = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in seq_cst
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_max_i32_ret_addr64:
+; SI: BUFFER_ATOMIC_SMAX [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}], s[{{[0-9]+}}:{{[0-9]+}}], 0 addr64 glc{{$}}
+; SI: BUFFER_STORE_DWORD [[RET]]
+define void @atomic_max_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+entry:
+ %ptr = getelementptr i32 addrspace(1)* %out, i64 %index
+ %0 = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in seq_cst
+ store i32 %0, i32 addrspace(1)* %out2
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_i32_offset:
+; SI: BUFFER_ATOMIC_UMAX v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offset:0x10{{$}}
+define void @atomic_umax_i32_offset(i32 addrspace(1)* %out, i32 %in) {
+entry:
+ %gep = getelementptr i32 addrspace(1)* %out, i32 4
+ %0 = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_i32_ret_offset:
+; SI: BUFFER_ATOMIC_UMAX [[RET:v[0-9]+]], s[{{[0-9]+}}:{{[0-9]+}}], 0 offset:0x10 glc {{$}}
+; SI: BUFFER_STORE_DWORD [[RET]]
+define void @atomic_umax_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+entry:
+ %gep = getelementptr i32 addrspace(1)* %out, i32 4
+ %0 = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+ store i32 %0, i32 addrspace(1)* %out2
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_i32_addr64_offset:
+; SI: BUFFER_ATOMIC_UMAX v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], s[{{[0-9]+}}:{{[0-9]+}}], 0 addr64 offset:0x10{{$}}
+define void @atomic_umax_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+entry:
+ %ptr = getelementptr i32 addrspace(1)* %out, i64 %index
+ %gep = getelementptr i32 addrspace(1)* %ptr, i32 4
+ %0 = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_i32_ret_addr64_offset:
+; SI: BUFFER_ATOMIC_UMAX [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}], s[{{[0-9]+}}:{{[0-9]+}}], 0 addr64 offset:0x10 glc{{$}}
+; SI: BUFFER_STORE_DWORD [[RET]]
+define void @atomic_umax_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+entry:
+ %ptr = getelementptr i32 addrspace(1)* %out, i64 %index
+ %gep = getelementptr i32 addrspace(1)* %ptr, i32 4
+ %0 = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
+ store i32 %0, i32 addrspace(1)* %out2
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_i32:
+; SI: BUFFER_ATOMIC_UMAX v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0{{$}}
+define void @atomic_umax_i32(i32 addrspace(1)* %out, i32 %in) {
+entry:
+ %0 = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in seq_cst
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_i32_ret:
+; SI: BUFFER_ATOMIC_UMAX [[RET:v[0-9]+]], s[{{[0-9]+}}:{{[0-9]+}}], 0 glc
+; SI: BUFFER_STORE_DWORD [[RET]]
+define void @atomic_umax_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
+entry:
+ %0 = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in seq_cst
+ store i32 %0, i32 addrspace(1)* %out2
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_i32_addr64:
+; SI: BUFFER_ATOMIC_UMAX v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], s[{{[0-9]+}}:{{[0-9]+}}], 0 addr64{{$}}
+define void @atomic_umax_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
+entry:
+ %ptr = getelementptr i32 addrspace(1)* %out, i64 %index
+ %0 = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in seq_cst
+ ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_i32_ret_addr64:
+; SI: BUFFER_ATOMIC_UMAX [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}], s[{{[0-9]+}}:{{[0-9]+}}], 0 addr64 glc{{$}}
+; SI: BUFFER_STORE_DWORD [[RET]]
+define void @atomic_umax_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
+entry:
+ %ptr = getelementptr i32 addrspace(1)* %out, i64 %index
+ %0 = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in seq_cst
+ store i32 %0, i32 addrspace(1)* %out2
+ ret void
+}
--
2.1.0
More information about the llvm-commits
mailing list