[AMDGPU] Add an AMDGPU specific atomic optimizer.

This commit adds a new IR level pass to the AMDGPU backend to perform
atomic optimizations. It works by:

- Running through a function and finding atomicrmw add/sub or uses of
  the atomic buffer intrinsics for add/sub.
- If all arguments except the value to be added/subtracted are uniform,
  record the value to be optimized.
- Run through the atomic operations we can optimize and, depending on
  whether the value is uniform/divergent use wavefront wide operations
  (DPP in the divergent case) to calculate the total amount to be
  atomically added/subtracted.
- Then let only a single lane of each wavefront perform the atomic
  operation, reducing the total number of atomic operations in flight.
- Lastly we recombine the result from the single lane to each lane of
  the wavefront, and calculate our individual lanes offset into the
  final result.

Differential Revision: https://reviews.llvm.org/D51969

llvm-svn: 343973

1 files changed, 129 insertions, 0 deletions

diff --git a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_raw_buffer.ll b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_raw_buffer.ll
new file mode 100644
index 00000000000..3e71e02ebfa
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_raw_buffer.ll
@@ -0,0 +1,129 @@
+; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX7LESS %s
+; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=tonga -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX8MORE %s
+; RUN: llc -march=amdgcn -mtriple=amdgcn---amdgiz -mcpu=gfx900 -mattr=-flat-for-global -amdgpu-atomic-optimizations=true -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX8MORE %s
+
+declare i32 @llvm.amdgcn.workitem.id.x()
+declare i32 @llvm.amdgcn.raw.buffer.atomic.add(i32, <4 x i32>, i32, i32, i32)
+declare i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32, <4 x i32>, i32, i32, i32)
+
+; Show that what the atomic optimization pass will do for raw buffers.
+
+; GCN-LABEL: add_i32_constant:
+; GCN: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
+; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_lo:[0-9]+]], s[[exec_lo]], 0
+; GCN: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_hi:[0-9]+]], s[[exec_hi]], v[[mbcnt_lo]]
+; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc, 0, v[[mbcnt_hi]]
+; GCN: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
+; GCN: v_mul_u32_u24{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[popcount]], 5
+; GCN: buffer_atomic_add v[[value]]
+define amdgpu_kernel void @add_i32_constant(i32 addrspace(1)* %out, <4 x i32> %inout) {
+entry:
+  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.add(i32 5, <4 x i32> %inout, i32 0, i32 0, i32 0)
+  store i32 %old, i32 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: add_i32_uniform:
+; GCN: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
+; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_lo:[0-9]+]], s[[exec_lo]], 0
+; GCN: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_hi:[0-9]+]], s[[exec_hi]], v[[mbcnt_lo]]
+; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc, 0, v[[mbcnt_hi]]
+; GCN: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
+; GCN: s_mul_i32 s[[scalar_value:[0-9]+]], s{{[0-9]+}}, s[[popcount]]
+; GCN: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
+; GCN: buffer_atomic_add v[[value]]
+define amdgpu_kernel void @add_i32_uniform(i32 addrspace(1)* %out, <4 x i32> %inout, i32 %additive) {
+entry:
+  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.add(i32 %additive, <4 x i32> %inout, i32 0, i32 0, i32 0)
+  store i32 %old, i32 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: add_i32_varying_vdata:
+; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
+; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
+; GFX7LESS-NOT: s_bcnt1_i32_b64
+; GFX7LESS: buffer_atomic_add v{{[0-9]+}}
+; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
+; GFX8MORE: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
+; GFX8MORE: buffer_atomic_add v[[value]]
+define amdgpu_kernel void @add_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
+entry:
+  %lane = call i32 @llvm.amdgcn.workitem.id.x()
+  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.add(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i32 0)
+  store i32 %old, i32 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: add_i32_varying_offset:
+; GCN-NOT: v_mbcnt_lo_u32_b32
+; GCN-NOT: v_mbcnt_hi_u32_b32
+; GCN-NOT: s_bcnt1_i32_b64
+; GCN: buffer_atomic_add v{{[0-9]+}}
+define amdgpu_kernel void @add_i32_varying_offset(i32 addrspace(1)* %out, <4 x i32> %inout) {
+entry:
+  %lane = call i32 @llvm.amdgcn.workitem.id.x()
+  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.add(i32 1, <4 x i32> %inout, i32 %lane, i32 0, i32 0)
+  store i32 %old, i32 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: sub_i32_constant:
+; GCN: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
+; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_lo:[0-9]+]], s[[exec_lo]], 0
+; GCN: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_hi:[0-9]+]], s[[exec_hi]], v[[mbcnt_lo]]
+; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc, 0, v[[mbcnt_hi]]
+; GCN: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
+; GCN: v_mul_u32_u24{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[popcount]], 5
+; GCN: buffer_atomic_sub v[[value]]
+define amdgpu_kernel void @sub_i32_constant(i32 addrspace(1)* %out, <4 x i32> %inout) {
+entry:
+  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 5, <4 x i32> %inout, i32 0, i32 0, i32 0)
+  store i32 %old, i32 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: sub_i32_uniform:
+; GCN: s_mov_b64 s{{\[}}[[exec_lo:[0-9]+]]:[[exec_hi:[0-9]+]]{{\]}}, exec
+; GCN: v_mbcnt_lo_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_lo:[0-9]+]], s[[exec_lo]], 0
+; GCN: v_mbcnt_hi_u32_b32{{(_e[0-9]+)?}} v[[mbcnt_hi:[0-9]+]], s[[exec_hi]], v[[mbcnt_lo]]
+; GCN: v_cmp_eq_u32{{(_e[0-9]+)?}} vcc, 0, v[[mbcnt_hi]]
+; GCN: s_bcnt1_i32_b64 s[[popcount:[0-9]+]], s{{\[}}[[exec_lo]]:[[exec_hi]]{{\]}}
+; GCN: s_mul_i32 s[[scalar_value:[0-9]+]], s{{[0-9]+}}, s[[popcount]]
+; GCN: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
+; GCN: buffer_atomic_sub v[[value]]
+define amdgpu_kernel void @sub_i32_uniform(i32 addrspace(1)* %out, <4 x i32> %inout, i32 %subitive) {
+entry:
+  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 %subitive, <4 x i32> %inout, i32 0, i32 0, i32 0)
+  store i32 %old, i32 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: sub_i32_varying_vdata:
+; GFX7LESS-NOT: v_mbcnt_lo_u32_b32
+; GFX7LESS-NOT: v_mbcnt_hi_u32_b32
+; GFX7LESS-NOT: s_bcnt1_i32_b64
+; GFX7LESS: buffer_atomic_sub v{{[0-9]+}}
+; GFX8MORE: v_readlane_b32 s[[scalar_value:[0-9]+]], v{{[0-9]+}}, 63
+; GFX8MORE: v_mov_b32{{(_e[0-9]+)?}} v[[value:[0-9]+]], s[[scalar_value]]
+; GFX8MORE: buffer_atomic_sub v[[value]]
+define amdgpu_kernel void @sub_i32_varying_vdata(i32 addrspace(1)* %out, <4 x i32> %inout) {
+entry:
+  %lane = call i32 @llvm.amdgcn.workitem.id.x()
+  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 %lane, <4 x i32> %inout, i32 0, i32 0, i32 0)
+  store i32 %old, i32 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: sub_i32_varying_offset:
+; GCN-NOT: v_mbcnt_lo_u32_b32
+; GCN-NOT: v_mbcnt_hi_u32_b32
+; GCN-NOT: s_bcnt1_i32_b64
+; GCN: buffer_atomic_sub v{{[0-9]+}}
+define amdgpu_kernel void @sub_i32_varying_offset(i32 addrspace(1)* %out, <4 x i32> %inout) {
+entry:
+  %lane = call i32 @llvm.amdgcn.workitem.id.x()
+  %old = call i32 @llvm.amdgcn.raw.buffer.atomic.sub(i32 1, <4 x i32> %inout, i32 %lane, i32 0, i32 0)
+  store i32 %old, i32 addrspace(1)* %out
+  ret void
+}