[llvm] r325567 - [AMDGPU] stop buffer_store being moved illegally

[Tim Renouf via llvm-commits]

Author: tpr
Date: Tue Feb 20 02:03:38 2018
New Revision: 325567

URL: http://llvm.org/viewvc/llvm-project?rev=325567&view=rev
Log:
[AMDGPU] stop buffer_store being moved illegally

Summary:
The machine instruction scheduler was illegally moving a buffer store
past a buffer load with the same descriptor and offset. Fixed by marking
buffer ops as mayAlias and isAliased. This may be overly conservative,
and we may need to revisit.

Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, t-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D43332

Change-Id: Iff3173d9e0653e830474546276ab9d30318b8ef7

Added:
    llvm/trunk/test/CodeGen/AMDGPU/buffer-schedule.ll
Modified:
    llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h

Modified: llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h?rev=325567&r1=325566&r2=325567&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h (original)
+++ llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h Tue Feb 20 02:03:38 2018
@@ -70,15 +70,11 @@ public:
   }
 
   bool isAliased(const MachineFrameInfo *) const override {
-    // FIXME: If we ever change image intrinsics to accept fat pointers, then
-    // this could be true for some cases.
-    return false;
+    return true;
   }
 
   bool mayAlias(const MachineFrameInfo *) const override {
-    // FIXME: If we ever change image intrinsics to accept fat pointers, then
-    // this could be true for some cases.
-    return false;
+    return true;
   }
 };
 

Added: llvm/trunk/test/CodeGen/AMDGPU/buffer-schedule.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/buffer-schedule.ll?rev=325567&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/buffer-schedule.ll (added)
+++ llvm/trunk/test/CodeGen/AMDGPU/buffer-schedule.ll Tue Feb 20 02:03:38 2018
@@ -0,0 +1,53 @@
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s
+
+; The buffer_loads and buffer_stores all access the same location. Check they do
+; not get reordered by the scheduler.
+
+; GCN-LABEL: {{^}}_amdgpu_cs_main:
+; GCN: buffer_load_dword
+; GCN: buffer_store_dword
+; GCN: buffer_load_dword
+; GCN: buffer_store_dword
+; GCN: buffer_load_dword
+; GCN: buffer_store_dword
+; GCN: buffer_load_dword
+; GCN: buffer_store_dword
+
+; Function Attrs: nounwind
+define amdgpu_cs void @_amdgpu_cs_main(<3 x i32> inreg %arg3, <3 x i32> %arg5) {
+.entry:
+  %tmp9 = add <3 x i32> %arg3, %arg5
+  %tmp10 = extractelement <3 x i32> %tmp9, i32 0
+  %tmp11 = shl i32 %tmp10, 2
+  %tmp12 = inttoptr i64 undef to <4 x i32> addrspace(4)*
+  %tmp13 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp12, align 16
+  %tmp14 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> %tmp13, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp17 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp12, align 16
+  call void @llvm.amdgcn.buffer.store.f32(float %tmp14, <4 x i32> %tmp17, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp12, align 16
+  %tmp21 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> %tmp20, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp22 = fadd reassoc nnan arcp contract float %tmp21, 1.000000e+00
+  call void @llvm.amdgcn.buffer.store.f32(float %tmp22, <4 x i32> %tmp20, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp25 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp12, align 16
+  %tmp26 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> %tmp25, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp27 = fadd reassoc nnan arcp contract float %tmp26, 1.000000e+00
+  call void @llvm.amdgcn.buffer.store.f32(float %tmp27, <4 x i32> %tmp25, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp30 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp12, align 16
+  %tmp31 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> %tmp30, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp32 = fadd reassoc nnan arcp contract float %tmp31, 1.000000e+00
+  call void @llvm.amdgcn.buffer.store.f32(float %tmp32, <4 x i32> %tmp30, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp35 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp12, align 16
+  %tmp36 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> %tmp35, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  %tmp37 = fadd reassoc nnan arcp contract float %tmp36, 1.000000e+00
+  call void @llvm.amdgcn.buffer.store.f32(float %tmp37, <4 x i32> %tmp35, i32 0, i32 %tmp11, i1 false, i1 false) #0
+  ret void
+}
+
+declare float @llvm.amdgcn.buffer.load.f32(<4 x i32>, i32, i32, i1, i1) #2
+
+declare void @llvm.amdgcn.buffer.store.f32(float, <4 x i32>, i32, i32, i1, i1) #3
+
+attributes #2 = { nounwind readonly }
+attributes #3 = { nounwind writeonly }
+