AMDGPU: Split MUBUF offset into aligned components

[nhaehnle]

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 4152 = 4092 + 60.

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

[nhaehnle]

Updated commit message: the highest value that can be represented with
immediate + inline is 4156 = 4092 + 64.

[arsenm]

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1236–1239

Can you keep using the bit ranges, and then use the various alignTo functions instead of changing the hardcoded values

[nhaehnle]

Use alignDown instead of magic 4092.

[arsenm]

LGTM

lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

 bool AMDGPUDAGToDAGISel::SelectMUBUFConstant(SDValue Constant,
                                              SDValue &SOffset,
                                              SDValue &ImmOffset) const {
   SDLoc DL(Constant);
   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 <= 4092 + 64) {
+      // Use an SOffset inline constant for 4..64
+      Overflow = Imm - 4092;
+      Imm = 4092;
     } 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 + 4) & ~4095;
+      uint32_t Low = (Imm + 4) & 4095;
       Imm = Low;
-      Overflow = High - 1;
+      Overflow = High - 4;
     }
   }
 
   // There is a hardware bug in SI and CI which prevents address clamping in
   // MUBUF instructions from working correctly with SOffsets. The immediate
   // offset is unaffected.
   if (Overflow > 0 &&
       Subtarget->getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS)

test/CodeGen/AMDGPU/llvm.amdgcn.buffer.atomic.ll

 ;RUN: llc < %s -march=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck %s -check-prefix=CHECK -check-prefix=SICI
 ;RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck %s -check-prefix=CHECK -check-prefix=VI
 
 ;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)
 ;CHECK: buffer_atomic_swap v0, v2, s[0:3], 0 offen glc
 ;CHECK: s_waitcnt vmcnt(0)
 ;CHECK: buffer_atomic_swap v0, v[1:2], s[0:3], 0 idxen offen glc
 ;CHECK: s_waitcnt vmcnt(0)
 ;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) {
 main_body:
   %o1 = call i32 @llvm.amdgcn.buffer.atomic.swap(i32 %data, <4 x i32> %rsrc, i32 0, i32 0, i1 0)
   %o2 = call i32 @llvm.amdgcn.buffer.atomic.swap(i32 %o1, <4 x i32> %rsrc, i32 %vindex, i32 0, i1 0)
   %o3 = call i32 @llvm.amdgcn.buffer.atomic.swap(i32 %o2, <4 x i32> %rsrc, i32 0, i32 %voffset, i1 0)
   %o4 = call i32 @llvm.amdgcn.buffer.atomic.swap(i32 %o3, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i1 0)

 
 ; Ideally, we would teach tablegen & friends that cmpswap only modifies the
 ; first vgpr. Since we don't do that yet, the register allocator will have to
 ; create copies which we don't bother to track here.
 ;
 ;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)
 
 ; Detecting the no-return variant doesn't work right now because of how the
 ; intrinsic is replaced by an instruction that feeds into an EXTRACT_SUBREG.
 ; Since there probably isn't a reasonable use-case of cmpswap that discards
 ; the return value, that seems okay.
 ;

test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.format.ll

 ;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.format.v4f32(<4 x i32> %0, i32 0, i32 42, i1 0, i1 0)
   ret <4 x float> %data
 }
 
 ;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
   ret <4 x float> %data
 }
 
 ;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:
   %d.0 = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 4160, i1 0, i1 0)
   %d.1 = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 4176, i1 0, i1 0)
   %data = fadd <4 x float> %d.0, %d.1
   ret <4 x float> %data
 }

 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_ofs(<4 x i32> inreg, i32) {
 main_body:
   %data = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 0, i32 %1, i1 0, i1 0)
   ret <4 x float> %data
 }
 
 ;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
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_both:
 ;CHECK: buffer_load_format_xyzw v[0:3], v[0:1], s[0:3], 0 idxen offen
 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_both(<4 x i32> inreg, i32, i32) {

test/CodeGen/AMDGPU/llvm.amdgcn.buffer.load.ll

main_body:
   %data_slc = call <4 x float> @llvm.amdgcn.buffer.load.v4f32(<4 x i32> %0, i32 0, i32 0, i1 0, i1 1)
   %r0 = insertvalue {<4 x float>, <4 x float>, <4 x float>} undef, <4 x float> %data, 0
   %r1 = insertvalue {<4 x float>, <4 x float>, <4 x float>} %r0, <4 x float> %data_glc, 1
   %r2 = insertvalue {<4 x float>, <4 x float>, <4 x float>} %r1, <4 x float> %data_slc, 2
   ret {<4 x float>, <4 x float>, <4 x float>} %r2
 }
 
 ;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:
   %data = call <4 x float> @llvm.amdgcn.buffer.load.v4f32(<4 x i32> %0, i32 0, i32 8192, i1 0, i1 0)
   ret <4 x float> %data
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_idx:

 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_ofs(<4 x i32> inreg, i32) {
 main_body:
   %data = call <4 x float> @llvm.amdgcn.buffer.load.v4f32(<4 x i32> %0, i32 0, i32 %1, i1 0, i1 0)
   ret <4 x float> %data
 }
 
 ;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
 }
 
 ;CHECK-LABEL: {{^}}buffer_load_both:
 ;CHECK: buffer_load_dwordx4 v[0:3], v[0:1], s[0:3], 0 idxen offen
 ;CHECK: s_waitcnt
 define amdgpu_ps <4 x float> @buffer_load_both(<4 x i32> inreg, i32, i32) {