[spirv] add -fspv-reduce-load-size option (#3931)

Since loading a big object takes the memory pressure, reduction of the
load size can have some performance benefit. In particular, it is
useful for mobile GPUs. `-fspv-reduce-load-size` removes
OpLoad/OpCompositeExtract of struct/array types by running spirv-opt
--reduce-load-size pass.

Fixes #3889

external/SPIRV-Headers

@@ -1 +1 @@
-Subproject commit 449bc986ba6f4c5e10e32828783f9daef2a77644
+Subproject commit 6cae8216a6ea19ff3f237af01e54378c1ff81fcd

external/SPIRV-Tools

@@ -1 +1 @@
-Subproject commit bd3a271ce3fcc8b9ed8e6eb2e94392d1b220adc9
+Subproject commit 0c09258e073bdbae19d2de8386125b35b783928b

include/dxc/Support/HLSLOptions.td

@@ -339,6 +339,8 @@ def fspv_target_env_EQ : Joined<["-"], "fspv-target-env=">, Group<spirv_Group>,
   HelpText<"Specify the target environment: vulkan1.0 (default) or vulkan1.1">;
 def fspv_flatten_resource_arrays: Flag<["-"], "fspv-flatten-resource-arrays">, Group<spirv_Group>, Flags<[CoreOption, DriverOption]>,
   HelpText<"Flatten arrays of resources so each array element takes one binding number">;
+def fspv_reduce_load_size: Flag<["-"], "fspv-reduce-load-size">, Group<spirv_Group>, Flags<[CoreOption, DriverOption]>,
+  HelpText<"Replaces loads of composite objects to reduce memory pressure for the loads">;
 def fvk_auto_shift_bindings: Flag<["-"], "fvk-auto-shift-bindings">, Group<spirv_Group>, Flags<[CoreOption, DriverOption]>,
   HelpText<"Apply fvk-*-shift to resources without an explicit register assignment.">;
 def Wno_vk_ignored_features : Joined<["-"], "Wno-vk-ignored-features">, Group<spirv_Group>, Flags<[CoreOption, DriverOption, HelpHidden]>,

include/dxc/Support/SPIRVOptions.h

@@ -55,6 +55,7 @@ struct SpirvCodeGenOptions {
   bool useGlLayout;
   bool useScalarLayout;
   bool flattenResourceArrays;
+  bool reduceLoadSize;
   bool autoShiftBindings;
   bool supportNonzeroBaseInstance;
   SpirvLayoutRule cBufferLayoutRule;

lib/DxcSupport/HLSLOptions.cpp

@@ -859,6 +859,8 @@ int ReadDxcOpts(const OptTable *optionTable, unsigned flagsToInclude,
   opts.SpirvOptions.noWarnEmulatedFeatures = Args.hasFlag(OPT_Wno_vk_emulated_features, OPT_INVALID, false);
   opts.SpirvOptions.flattenResourceArrays =
       Args.hasFlag(OPT_fspv_flatten_resource_arrays, OPT_INVALID, false);
+  opts.SpirvOptions.reduceLoadSize =
+      Args.hasFlag(OPT_fspv_reduce_load_size, OPT_INVALID, false);
   opts.SpirvOptions.autoShiftBindings = Args.hasFlag(OPT_fvk_auto_shift_bindings, OPT_INVALID, false);
 
   if (!handleVkShiftArgs(Args, OPT_fvk_b_shift, "b", &opts.SpirvOptions.bShift, errors) ||
@@ -947,6 +949,7 @@ int ReadDxcOpts(const OptTable *optionTable, unsigned flagsToInclude,
       Args.hasFlag(OPT_fvk_use_dx_layout, OPT_INVALID, false) ||
       Args.hasFlag(OPT_fvk_use_scalar_layout, OPT_INVALID, false) ||
       Args.hasFlag(OPT_fspv_flatten_resource_arrays, OPT_INVALID, false) ||
+      Args.hasFlag(OPT_fspv_reduce_load_size, OPT_INVALID, false) ||
       Args.hasFlag(OPT_fspv_reflect, OPT_INVALID, false) ||
       Args.hasFlag(OPT_Wno_vk_ignored_features, OPT_INVALID, false) ||
       Args.hasFlag(OPT_Wno_vk_emulated_features, OPT_INVALID, false) ||

tools/clang/lib/SPIRV/SpirvEmitter.cpp

@@ -740,9 +740,9 @@ void SpirvEmitter::HandleTranslationUnit(ASTContext &context) {
 
   // In order to flatten composite resources, we must also unroll loops.
   // Therefore we should run legalization before optimization.
-  needsLegalization = needsLegalization ||
-                      declIdMapper.requiresLegalization() ||
-                      spirvOptions.flattenResourceArrays ||
+  needsLegalization =
+      needsLegalization || declIdMapper.requiresLegalization() ||
+      spirvOptions.flattenResourceArrays || spirvOptions.reduceLoadSize ||
       declIdMapper.requiresFlatteningCompositeResources() ||
       !dsetbindingsToCombineImageSampler.empty();
 
@@ -12466,6 +12466,13 @@ bool SpirvEmitter::spirvToolsLegalize(std::vector<uint32_t> *mod,
     // types.
     optimizer.RegisterPass(spvtools::CreateAggressiveDCEPass());
   }
+  if (spirvOptions.reduceLoadSize) {
+    // The threshold must be bigger than 1.0 to reduce all possible loads.
+    optimizer.RegisterPass(spvtools::CreateReduceLoadSizePass(1.1));
+    // ADCE should be run after reduce-load-size pass in order to remove
+    // dead instructions.
+    optimizer.RegisterPass(spvtools::CreateAggressiveDCEPass());
+  }
   optimizer.RegisterPass(spvtools::CreateReplaceInvalidOpcodePass());
   optimizer.RegisterPass(spvtools::CreateCompactIdsPass());
 

tools/clang/test/CodeGenSPIRV/reduce.load.size.hlsl

@@ -0,0 +1,32 @@
+// Run: %dxc -T cs_6_0 -E main -fspv-reduce-load-size -O0
+
+struct S {
+  uint f;
+};
+
+cbuffer gBuffer { uint a[6]; };
+
+RWStructuredBuffer<S> gRWSBuffer;
+
+// CHECK-NOT: OpCompositeExtract
+
+// CHECK: [[p0:%\w+]] = OpAccessChain %_ptr_Uniform_uint {{%\w+}} %uint_0
+// CHECK: OpLoad %uint [[p0]]
+// CHECK: [[p1:%\w+]] = OpAccessChain %_ptr_Uniform_uint {{%\w+}} %uint_1
+// CHECK: OpLoad %uint [[p1]]
+// CHECK: [[p2:%\w+]] = OpAccessChain %_ptr_Uniform_uint {{%\w+}} %uint_2
+// CHECK: OpLoad %uint [[p2]]
+// CHECK: [[p3:%\w+]] = OpAccessChain %_ptr_Uniform_uint {{%\w+}} %uint_3
+// CHECK: OpLoad %uint [[p3]]
+// CHECK: [[p4:%\w+]] = OpAccessChain %_ptr_Uniform_uint {{%\w+}} %uint_4
+// CHECK: OpLoad %uint [[p4]]
+// CHECK: [[p5:%\w+]] = OpAccessChain %_ptr_Uniform_uint {{%\w+}} %uint_5
+// CHECK: OpLoad %uint [[p5]]
+uint foo(uint p[6]) {
+  return p[0] + p[1] + p[2] + p[3] + p[4] + p[5];
+}
+
+[numthreads(1,1,1)]
+void main() {
+  gRWSBuffer[0].f = foo(a);
+}

tools/clang/unittests/SPIRV/CodeGenSpirvTest.cpp

@@ -2656,6 +2656,8 @@ TEST_F(FileTest, MeshShadingNVAmplificationError4) {
   runFileTest("meshshading.nv.error3.amplification.hlsl", Expect::Failure);
 }
 
+TEST_F(FileTest, ReduceLoadSize) { runFileTest("reduce.load.size.hlsl"); }
+
 // Test OpEntryPoint in the Vulkan1.2 target environment
 TEST_F(FileTest, Vk1p2EntryPoint) {
   runFileTest("vk.1p2.entry-point.hlsl");