Register DxilCounters.cpp with hctgen and regenerate it

- Add TARGET_NAME optional parameter to add_hlsl_hctgen() in HCT.cmake
  to allow different CMake target name and hctgen mode (needed since
  DxilCounters is already used as the target name for DxilCounters.h)
- Add add_hlsl_hctgen(DxilCounters OUTPUT DxilCounters.cpp CODE_TAG ...)
  to lib/DXIL/CMakeLists.txt so DxilCounters.cpp participates in source
  regeneration when HLSL_COPY_GENERATED_SOURCES is set
- Regenerate DxilCounters.cpp: adds BarrierByMemoryType/Handle/Node,
  FDot, VectorReduceAnd/Or, SampleCmpGrad/Bias, RawBufferVectorLoad/Store

Co-authored-by: damyanp <[email protected]>
Agent-Logs-Url: https://github.com/damyanp/DirectXShaderCompiler/sessions/12071741-931b-464e-90a5-ce03c9b6e9c3

Author: Copilot

cmake/modules/HCT.cmake

@@ -38,9 +38,12 @@ endif()
 function(add_hlsl_hctgen mode)
   cmake_parse_arguments(ARG
     "BUILD_DIR;CODE_TAG"
-    "OUTPUT"
+    "OUTPUT;TARGET_NAME"
     ""
     ${ARGN})
+  if (NOT ARG_TARGET_NAME)
+    set(ARG_TARGET_NAME ${mode})
+  endif()
 
   if (NOT ARG_OUTPUT)
     message(FATAL_ERROR "add_hlsl_hctgen requires OUTPUT argument")
@@ -132,8 +135,8 @@ function(add_hlsl_hctgen mode)
                      DEPENDS ${output}
                      COMMENT "Verifying clang-format results...")
 
-  add_custom_target(${mode}
+  add_custom_target(${ARG_TARGET_NAME}
                     DEPENDS ${temp_output}.stamp)
 
-  add_dependencies(HCTGen ${mode})
+  add_dependencies(HCTGen ${ARG_TARGET_NAME})
 endfunction()

lib/DXIL/CMakeLists.txt

@@ -2,6 +2,7 @@
 # This file is distributed under the University of Illinois Open Source License.
 # See LICENSE.TXT for details.
 
+add_hlsl_hctgen(DxilCounters OUTPUT DxilCounters.cpp CODE_TAG TARGET_NAME DxilCountersCpp)
 add_hlsl_hctgen(DxilOperations OUTPUT DxilOperations.cpp CODE_TAG)
 add_hlsl_hctgen(DxilShaderModel OUTPUT DxilShaderModel.cpp CODE_TAG)
 add_hlsl_hctgen(DxilMetadata OUTPUT DxilMetadataHelper.cpp CODE_TAG)

lib/DXIL/DxilCounters.cpp

@@ -133,17 +133,18 @@ bool CountDxilOp_atomic(unsigned op) {
   return op == 70 || (78 <= op && op <= 79);
 }
 bool CountDxilOp_barrier(unsigned op) {
-  // Instructions: Barrier=80
-  return op == 80;
+  // Instructions: Barrier=80, BarrierByMemoryType=244,
+  // BarrierByMemoryHandle=245, BarrierByNodeRecordHandle=246
+  return op == 80 || (244 <= op && op <= 246);
 }
 bool CountDxilOp_floats(unsigned op) {
   // Instructions: FAbs=6, Saturate=7, IsNaN=8, IsInf=9, IsFinite=10,
   // IsNormal=11, Cos=12, Sin=13, Tan=14, Acos=15, Asin=16, Atan=17, Hcos=18,
   // Hsin=19, Htan=20, Exp=21, Frc=22, Log=23, Sqrt=24, Rsqrt=25, Round_ne=26,
   // Round_ni=27, Round_pi=28, Round_z=29, FMax=35, FMin=36, Fma=47, Dot2=54,
-  // Dot3=55, Dot4=56, Dot2AddHalf=162
+  // Dot3=55, Dot4=56, Dot2AddHalf=162, FDot=311
   return (6 <= op && op <= 29) || (35 <= op && op <= 36) || op == 47 ||
-         (54 <= op && op <= 56) || op == 162;
+         (54 <= op && op <= 56) || op == 162 || op == 311;
 }
 bool CountDxilOp_gs_cut(unsigned op) {
   // Instructions: CutStream=98, EmitThenCutStream=99
@@ -155,9 +156,9 @@ bool CountDxilOp_gs_emit(unsigned op) {
 }
 bool CountDxilOp_ints(unsigned op) {
   // Instructions: IMax=37, IMin=38, IMul=41, IMad=48, Ibfe=51,
-  // Dot4AddI8Packed=163
+  // Dot4AddI8Packed=163, VectorReduceAnd=309, VectorReduceOr=310
   return (37 <= op && op <= 38) || op == 41 || op == 48 || op == 51 ||
-         op == 163;
+         op == 163 || (309 <= op && op <= 310);
 }
 bool CountDxilOp_sig_ld(unsigned op) {
   // Instructions: LoadInput=4, LoadOutputControlPoint=103,
@@ -175,16 +176,18 @@ bool CountDxilOp_tex_bias(unsigned op) {
 }
 bool CountDxilOp_tex_cmp(unsigned op) {
   // Instructions: SampleCmp=64, SampleCmpLevelZero=65, TextureGatherCmp=74,
-  // SampleCmpLevel=224
-  return (64 <= op && op <= 65) || op == 74 || op == 224;
+  // SampleCmpLevel=224, SampleCmpGrad=254, SampleCmpBias=255
+  return (64 <= op && op <= 65) || op == 74 || op == 224 ||
+         (254 <= op && op <= 255);
 }
 bool CountDxilOp_tex_grad(unsigned op) {
   // Instructions: SampleGrad=63
   return op == 63;
 }
 bool CountDxilOp_tex_load(unsigned op) {
-  // Instructions: TextureLoad=66, BufferLoad=68, RawBufferLoad=139
-  return op == 66 || op == 68 || op == 139;
+  // Instructions: TextureLoad=66, BufferLoad=68, RawBufferLoad=139,
+  // RawBufferVectorLoad=303
+  return op == 66 || op == 68 || op == 139 || op == 303;
 }
 bool CountDxilOp_tex_norm(unsigned op) {
   // Instructions: Sample=60, SampleLevel=62, TextureGather=73,
@@ -195,17 +198,18 @@ bool CountDxilOp_tex_store(unsigned op) {
   // Instructions: TextureStore=67, BufferStore=69, RawBufferStore=140,
   // WriteSamplerFeedback=174, WriteSamplerFeedbackBias=175,
   // WriteSamplerFeedbackLevel=176, WriteSamplerFeedbackGrad=177,
-  // TextureStoreSample=225
+  // TextureStoreSample=225, RawBufferVectorStore=304
   return op == 67 || op == 69 || op == 140 || (174 <= op && op <= 177) ||
-         op == 225;
+         op == 225 || op == 304;
 }
 bool CountDxilOp_uints(unsigned op) {
   // Instructions: Bfrev=30, Countbits=31, FirstbitLo=32, FirstbitHi=33,
   // FirstbitSHi=34, UMax=39, UMin=40, UMul=42, UDiv=43, UAddc=44, USubb=45,
-  // UMad=49, Msad=50, Ubfe=52, Bfi=53, Dot4AddU8Packed=164
+  // UMad=49, Msad=50, Ubfe=52, Bfi=53, Dot4AddU8Packed=164,
+  // VectorReduceAnd=309, VectorReduceOr=310
   return (30 <= op && op <= 34) || (39 <= op && op <= 40) ||
          (42 <= op && op <= 45) || (49 <= op && op <= 50) ||
-         (52 <= op && op <= 53) || op == 164;
+         (52 <= op && op <= 53) || op == 164 || (309 <= op && op <= 310);
 }
 // Counter functions for llvm ops:
 bool CountLlvmOp_atomic(unsigned op) {