AMDGPU: Split MUBUF offset into aligned components

Summary:
Atomic buffer operations do not work (and trap on gfx9) when the
components are unaligned, even if their sum is aligned.

Previously, we generated an offset of 4156 without an SGPR by
splitting it as 4095 + 61 (immediate + inline constant). The
highest offset for which we can do this correctly is 4156 = 4092 + 64.

Fixes dEQP-GLES31.functional.ssbo.atomic.*

Reviewers: arsenm

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

Differential Revision: https://reviews.llvm.org/D37850

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@315302 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index 7faf3e1..5bd1092 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -1235,24 +1235,30 @@ bool AMDGPUDAGToDAGISel::SelectMUBUFConstant(SDValue Constant,
                                              SDValue &SOffset,
                                              SDValue &ImmOffset) const {
   SDLoc DL(Constant);
+  const uint32_t Align = 4;
+  const uint32_t MaxImm = alignDown(4095, Align);
   uint32_t Imm = cast<ConstantSDNode>(Constant)->getZExtValue();
   uint32_t Overflow = 0;
 
-  if (Imm >= 4096) {
-    if (Imm <= 4095 + 64) {
-      // Use an SOffset inline constant for 1..64
-      Overflow = Imm - 4095;
-      Imm = 4095;
+  if (Imm > MaxImm) {
+    if (Imm <= MaxImm + 64) {
+      // Use an SOffset inline constant for 4..64
+      Overflow = Imm - MaxImm;
+      Imm = MaxImm;
     } else {
       // Try to keep the same value in SOffset for adjacent loads, so that
       // the corresponding register contents can be re-used.
       //
-      // Load values with all low-bits set into SOffset, so that a larger
-      // range of values can be covered using s_movk_i32
-      uint32_t High = (Imm + 1) & ~4095;
-      uint32_t Low = (Imm + 1) & 4095;
+      // Load values with all low-bits (except for alignment bits) set into
+      // SOffset, so that a larger range of values can be covered using
+      // s_movk_i32.
+      //
+      // Atomic operations fail to work correctly when individual address
+      // components are unaligned, even if their sum is aligned.
+      uint32_t High = (Imm + Align) & ~4095;
+      uint32_t Low = (Imm + Align) & 4095;
       Imm = Low;
-      Overflow = High - 1;
+      Overflow = High - Align;
     }
   }
 

diff --git a/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.atomic.ll b/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.atomic.ll
index 98f7058..b6f72a1 100644 (file)
--- a/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.atomic.ll
+++ b/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.atomic.ll
@@ -3,7 +3,7 @@
 
 ;CHECK-LABEL: {{^}}test1:
 ;CHECK: buffer_atomic_swap v0, off, s[0:3], 0 glc
-;VI: s_movk_i32 [[SOFS:s[0-9]+]], 0x1fff
+;VI: s_movk_i32 [[SOFS:s[0-9]+]], 0x1ffc
 ;CHECK: s_waitcnt vmcnt(0)
 ;CHECK: buffer_atomic_swap v0, v1, s[0:3], 0 idxen glc
 ;CHECK: s_waitcnt vmcnt(0)
@@ -14,7 +14,7 @@
 ;CHECK: buffer_atomic_swap v0, v2, s[0:3], 0 offen offset:42 glc
 ;CHECK-DAG: s_waitcnt vmcnt(0)
 ;SICI: buffer_atomic_swap v0, v1, s[0:3], 0 offen glc
-;VI: buffer_atomic_swap v0, off, s[0:3], [[SOFS]] offset:1 glc
+;VI: buffer_atomic_swap v0, off, s[0:3], [[SOFS]] offset:4 glc
 ;CHECK: s_waitcnt vmcnt(0)
 ;CHECK: buffer_atomic_swap v0, off, s[0:3], 0{{$}}
 define amdgpu_ps float @test1(<4 x i32> inreg %rsrc, i32 %data, i32 %vindex, i32 %voffset) {
@@ -71,24 +71,24 @@ main_body:
 ;CHECK-LABEL: {{^}}test3:
 ;CHECK: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], 0 glc
 ;CHECK: s_waitcnt vmcnt(0)
-;VI: s_movk_i32 [[SOFS:s[0-9]+]], 0x1fff
+;VI: s_movk_i32 [[SOFS:s[0-9]+]], 0x1ffc
 ;CHECK: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, v2, s[0:3], 0 idxen glc
 ;CHECK: s_waitcnt vmcnt(0)
 ;CHECK: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, v3, s[0:3], 0 offen glc
 ;CHECK: s_waitcnt vmcnt(0)
 ;CHECK: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, v[2:3], s[0:3], 0 idxen offen glc
 ;CHECK: s_waitcnt vmcnt(0)
-;CHECK: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, v3, s[0:3], 0 offen offset:42 glc
+;CHECK: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, v3, s[0:3], 0 offen offset:44 glc
 ;CHECK-DAG: s_waitcnt vmcnt(0)
 ;SICI: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}}, s[0:3], 0 offen glc
-;VI: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[SOFS]] offset:1 glc
+;VI: buffer_atomic_cmpswap {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[SOFS]] offset:4 glc
 define amdgpu_ps float @test3(<4 x i32> inreg %rsrc, i32 %data, i32 %cmp, i32 %vindex, i32 %voffset) {
 main_body:
   %o1 = call i32 @llvm.amdgcn.buffer.atomic.cmpswap(i32 %data, i32 %cmp, <4 x i32> %rsrc, i32 0, i32 0, i1 0)
   %o2 = call i32 @llvm.amdgcn.buffer.atomic.cmpswap(i32 %o1, i32 %cmp, <4 x i32> %rsrc, i32 %vindex, i32 0, i1 0)
   %o3 = call i32 @llvm.amdgcn.buffer.atomic.cmpswap(i32 %o2, i32 %cmp, <4 x i32> %rsrc, i32 0, i32 %voffset, i1 0)
   %o4 = call i32 @llvm.amdgcn.buffer.atomic.cmpswap(i32 %o3, i32 %cmp, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i1 0)
-  %ofs.5 = add i32 %voffset, 42
+  %ofs.5 = add i32 %voffset, 44
   %o5 = call i32 @llvm.amdgcn.buffer.atomic.cmpswap(i32 %o4, i32 %cmp, <4 x i32> %rsrc, i32 0, i32 %ofs.5, i1 0)
   %o6 = call i32 @llvm.amdgcn.buffer.atomic.cmpswap(i32 %o5, i32 %cmp, <4 x i32> %rsrc, i32 0, i32 8192, i1 0)
 

diff --git a/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.format.ll b/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.format.ll
index 9cb9f25..d515993 100644 (file)
--- a/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.format.ll
+++ b/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.format.ll
@@ -27,20 +27,20 @@ main_body:
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_immoffs_large:
-;SICI: v_mov_b32_e32 [[VOFS:v[0-9]+]], 0x103c
+;SICI: v_mov_b32_e32 [[VOFS:v[0-9]+]], 0x1038
 ;SICI: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, [[VOFS]], s[0:3], 0 offen
 ;SICI: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}}, s[0:3], 0 offen
-;VI-DAG: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], 61 offset:4095
-;VI-DAG: s_movk_i32 [[OFS1:s[0-9]+]], 0x7fff
-;VI-DAG: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[OFS1]] offset:4093
+;VI-DAG: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], 60 offset:4092
+;VI-DAG: s_movk_i32 [[OFS1:s[0-9]+]], 0x7ffc
+;VI-DAG: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[OFS1]] offset:4092
 ;SICI: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}}, s[0:3], 0 offen
-;VI-DAG: s_mov_b32 [[OFS2:s[0-9]+]], 0x8fff
-;VI-DAG: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[OFS2]] offset:1
+;VI-DAG: s_mov_b32 [[OFS2:s[0-9]+]], 0x8ffc
+;VI-DAG: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[OFS2]] offset:4
 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_immoffs_large(<4 x i32> inreg) {
 main_body:
-  %d.0 = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 4156, i1 0, i1 0)
-  %d.1 = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 36860, i1 0, i1 0)
+  %d.0 = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 4152, i1 0, i1 0)
+  %d.1 = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 36856, i1 0, i1 0)
   %d.2 = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 36864, i1 0, i1 0)
   %d.3 = fadd <4 x float> %d.0, %d.1
   %data = fadd <4 x float> %d.2, %d.3
@@ -48,10 +48,10 @@ main_body:
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_immoffs_reuse:
-;VI: s_movk_i32 [[OFS:s[0-9]+]], 0xfff
-;VI: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[OFS]] offset:65
+;VI: s_movk_i32 [[OFS:s[0-9]+]], 0xffc
+;VI: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[OFS]] offset:68
 ;VI-NOT: s_mov
-;VI: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[OFS]] offset:81
+;VI: buffer_load_format_xyzw {{v\[[0-9]+:[0-9]+\]}}, off, s[0:3], [[OFS]] offset:84
 ;VI: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_immoffs_reuse(<4 x i32> inreg) {
 main_body:
@@ -80,11 +80,11 @@ main_body:
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_ofs_imm:
-;CHECK: buffer_load_format_xyzw v[0:3], v0, s[0:3], 0 offen offset:58
+;CHECK: buffer_load_format_xyzw v[0:3], v0, s[0:3], 0 offen offset:60
 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_ofs_imm(<4 x i32> inreg, i32) {
 main_body:
-  %ofs = add i32 %1, 58
+  %ofs = add i32 %1, 60
   %data = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 %ofs, i1 0, i1 0)
   ret <4 x float> %data
 }

diff --git a/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.ll b/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.ll
index 5c93ae0..03caca8 100644 (file)
--- a/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.ll
+++ b/test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.ll
@@ -18,18 +18,18 @@ main_body:
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_immoffs:
-;CHECK: buffer_load_dwordx4 v[0:3], off, s[0:3], 0 offset:42
+;CHECK: buffer_load_dwordx4 v[0:3], off, s[0:3], 0 offset:40
 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_immoffs(<4 x i32> inreg) {
 main_body:
-  %data = call <4 x float> @llvm.amdgcn.buffer.load.v4f32(<4 x i32> %0, i32 0, i32 42, i1 0, i1 0)
+  %data = call <4 x float> @llvm.amdgcn.buffer.load.v4f32(<4 x i32> %0, i32 0, i32 40, i1 0, i1 0)
   ret <4 x float> %data
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_immoffs_large:
 ;SICI: buffer_load_dwordx4 v[0:3], {{v[0-9]+}}, s[0:3], 0 offen
-;VI: s_movk_i32 [[OFFSET:s[0-9]+]], 0x1fff
-;VI: buffer_load_dwordx4 v[0:3], off, s[0:3], [[OFFSET]] offset:1
+;VI: s_movk_i32 [[OFFSET:s[0-9]+]], 0x1ffc
+;VI: buffer_load_dwordx4 v[0:3], off, s[0:3], [[OFFSET]] offset:4
 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_immoffs_large(<4 x i32> inreg) {
 main_body:
@@ -56,11 +56,11 @@ main_body:
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_ofs_imm:
-;CHECK: buffer_load_dwordx4 v[0:3], v0, s[0:3], 0 offen offset:58
+;CHECK: buffer_load_dwordx4 v[0:3], v0, s[0:3], 0 offen offset:60
 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_ofs_imm(<4 x i32> inreg, i32) {
 main_body:
-  %ofs = add i32 %1, 58
+  %ofs = add i32 %1, 60
   %data = call <4 x float> @llvm.amdgcn.buffer.load.v4f32(<4 x i32> %0, i32 0, i32 %ofs, i1 0, i1 0)
   ret <4 x float> %data
 }