[AMDGPU] stop buffer_store being moved illegally

[tpr]

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.

Change-Id: Iff3173d9e0653e830474546276ab9d30318b8ef7

[arsenm]

LGTM except test cleanups

test/CodeGen/AMDGPU/buffer-schedule.ll

4–5

Remove these and specify the full triple to the llc run line

20–21

Extra copies

23

Delete attribute comments

lib/Target/AMDGPU/SIMachineFunctionInfo.h

public:
 
   bool isConstant(const MachineFrameInfo *) const override {
     // This should probably be true for most images, but we will start by being
     // conservative.
     return false;
   }
 
   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;
   }
 };
 
 /// This class keeps track of the SPI_SP_INPUT_ADDR config register, which
 /// tells the hardware which interpolation parameters to load.
 class SIMachineFunctionInfo final : public AMDGPUMachineFunction {
   unsigned TIDReg = AMDGPU::NoRegister;
 

test/CodeGen/AMDGPU/buffer-schedule.ll

+; RUN: llc -march=amdgcn -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s
+
+target datalayout = "e-p:64:64-p1:64:64-p2:64:64-p3:32:32-p4:32:32-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-A5"
+target triple = "amdgcn--amdpal"
+
+; 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
+
+target datalayout = "e-p:64:64-p1:64:64-p2:64:64-p3:32:32-p4:32:32-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-A5"
+target triple = "amdgcn--amdpal"
+
+; 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(2)*
+  %tmp13 = load <4 x i32>, <4 x i32> addrspace(2)* %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(2)* %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(2)* %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(2)* %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(2)* %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(2)* %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
+}
+
+; Function Attrs: nounwind readonly
+declare float @llvm.amdgcn.buffer.load.f32(<4 x i32>, i32, i32, i1, i1) #2
+
+; Function Attrs: nounwind writeonly
+declare void @llvm.amdgcn.buffer.store.f32(float, <4 x i32>, i32, i32, i1, i1) #3
+
+attributes #2 = { nounwind readonly }
+attributes #3 = { nounwind writeonly }
+