[clang] [compiler-rt] [llvm] [PGO][AMDGPU] Add uniformity-aware offload profile format and instrumentation (PR #190708)

Yaxun Liu via cfe-commits cfe-commits at lists.llvm.org
Mon Jun 1 06:54:48 PDT 2026


https://github.com/yxsamliu updated https://github.com/llvm/llvm-project/pull/190708

>From 56edfebc8f80feacfcdae621a9816a4407754d3c Mon Sep 17 00:00:00 2001
From: "Yaxun (Sam) Liu" <yaxun.liu at amd.com>
Date: Mon, 6 Apr 2026 18:58:11 -0400
Subject: [PATCH] [PGO][AMDGPU] Add uniformity-aware offload profile format and
 instrumentation

Add the producer-side format and instrumentation pieces that layer on top
of the basic HIP offload PGO support.

This extends the raw/indexed profile format with wave size and uniform
counters, teaches llvm-profdata to read and merge them, and adds the AMDGPU
sampling runtime.

The uniform counters follow the same optional-section design as the MC/DC
bitmap: they live in their own section (__llvm_prf_ucnts) with header
descriptor fields that are always present, but the section is empty
(NumUniformCounters == 0) for non-GPU targets, so host profiles are
unaffected. Because this inserts a new section into the baseline layout,
the format version is bumped (raw 10->11, indexed 13->14).
---
 clang/lib/CodeGen/CGCUDANV.cpp                |  18 +-
 compiler-rt/include/profile/InstrProfData.inc |  27 ++-
 compiler-rt/lib/profile/InstrProfiling.h      |  30 ++-
 .../lib/profile/InstrProfilingBuffer.c        |  34 +--
 compiler-rt/lib/profile/InstrProfilingFile.c  |  24 ++-
 .../lib/profile/InstrProfilingInternal.h      |  19 +-
 .../lib/profile/InstrProfilingPlatformGPU.c   |  48 ++++-
 .../profile/InstrProfilingPlatformROCm.cpp    | 102 +++++++--
 .../lib/profile/InstrProfilingWriter.c        |  38 +++-
 llvm/include/llvm/IR/FixedMetadataKinds.def   |   1 +
 llvm/include/llvm/IR/RuntimeLibcalls.td       |   5 +-
 llvm/include/llvm/ProfileData/InstrProf.h     |  43 +++-
 .../llvm/ProfileData/InstrProfData.inc        |  27 ++-
 .../llvm/ProfileData/InstrProfReader.h        |   5 +
 .../llvm/ProfileData/InstrProfWriter.h        |   1 +
 llvm/lib/ProfileData/InstrProf.cpp            |  10 +-
 llvm/lib/ProfileData/InstrProfCorrelator.cpp  |   2 +
 llvm/lib/ProfileData/InstrProfReader.cpp      | 113 +++++++++-
 llvm/lib/ProfileData/InstrProfWriter.cpp      |  45 +++-
 .../Instrumentation/InstrProfiling.cpp        | 202 ++++++++++++++++--
 .../Instrumentation/PGOInstrumentation.cpp    |  71 +++++-
 .../AArch64/global-merge-profile-sections.ll  |   4 +-
 .../InstrProfiling/amdgpu-3d-grid.ll          |  23 ++
 ...rrays.ll => amdgpu-contiguous-counters.ll} |  14 +-
 .../InstrProfiling/amdgpu-instrumentation.ll  |  45 +++-
 .../InstrProfiling/amdgpu-uniform-counters.ll |  22 ++
 .../InstrProfiling/amdgpu-wave32.ll           |  35 +++
 .../InstrProfiling/amdgpu-wave64.ll           |  34 +++
 .../InstrProfiling/coverage.ll                |   8 +-
 .../InstrProfiling/inline-data-var-create.ll  |  23 +-
 .../InstrProfiling/platform.ll                |  14 ++
 .../thinlto_indirect_call_promotion.profraw   | Bin 544 -> 600 bytes
 .../Transforms/PGOProfile/comdat_internal.ll  |   4 +-
 .../instrprof_burst_sampling_fast.ll          |   2 +-
 .../Transforms/PGOProfile/vtable_profile.ll   |   2 +-
 .../llvm-profdata/Inputs/c-general.profraw    | Bin 2032 -> 2152 bytes
 .../llvm-profdata/Inputs/compressed.profraw   | Bin 1984 -> 2104 bytes
 .../llvm-profdata/binary-ids-padding.test     |  19 +-
 .../llvm-profdata/large-binary-id-size.test   |   9 +-
 ...alformed-not-space-for-another-header.test |   7 +-
 .../malformed-num-counters-zero.test          |   8 +-
 .../malformed-ptr-to-counter-array.test       |   8 +-
 .../misaligned-binary-ids-size.test           |   2 +-
 .../mismatched-raw-profile-header.test        |   3 +
 .../tools/llvm-profdata/profile-version.test  |   2 +-
 .../tools/llvm-profdata/raw-32-bits-be.test   |  19 +-
 .../tools/llvm-profdata/raw-32-bits-le.test   |  19 +-
 .../tools/llvm-profdata/raw-64-bits-be.test   |  21 +-
 .../tools/llvm-profdata/raw-64-bits-le.test   |  19 +-
 .../tools/llvm-profdata/raw-two-profiles.test |  12 +-
 llvm/tools/llvm-profdata/llvm-profdata.cpp    |  27 +++
 .../common/include/GlobalHandler.h            |  14 +-
 .../common/src/GlobalHandler.cpp              |   4 +-
 53 files changed, 1076 insertions(+), 212 deletions(-)
 create mode 100644 llvm/test/Instrumentation/InstrProfiling/amdgpu-3d-grid.ll
 rename llvm/test/Instrumentation/InstrProfiling/{amdgpu-profc-arrays.ll => amdgpu-contiguous-counters.ll} (52%)
 create mode 100644 llvm/test/Instrumentation/InstrProfiling/amdgpu-uniform-counters.ll
 create mode 100644 llvm/test/Instrumentation/InstrProfiling/amdgpu-wave32.ll
 create mode 100644 llvm/test/Instrumentation/InstrProfiling/amdgpu-wave64.ll

diff --git a/clang/lib/CodeGen/CGCUDANV.cpp b/clang/lib/CodeGen/CGCUDANV.cpp
index a727a006969d8..518e0c3db5ed6 100644
--- a/clang/lib/CodeGen/CGCUDANV.cpp
+++ b/clang/lib/CodeGen/CGCUDANV.cpp
@@ -1328,12 +1328,14 @@ void CGNVCUDARuntime::createOffloadingEntries() {
 }
 
 // For HIP host+device compiles with PGO enabled, emit the per-TU global
-// __llvm_profile_sections_<CUID>. Device side: a 7-pointer struct holding
-// section start/stop bounds for the names/counters/data sections plus the
-// raw-version variable. Host side: an opaque void* shadow whose only
-// purpose is to give the host-runtime a registered symbol name to look up
-// via hipGetSymbolAddress; the actual device-side data lives in the
-// matching device-side global.
+// __llvm_profile_sections_<CUID>. Device side: a 9-pointer struct holding
+// section start/stop bounds for the names/counters/data/uniform-counters
+// sections plus the raw-version variable; the field order must match
+// INSTR_PROF_GPU_SECT in InstrProfData.inc, which the host runtime's
+// __llvm_profile_gpu_sections struct is generated from. Host side: an opaque
+// void* shadow whose only purpose is to give the host-runtime a registered
+// symbol name to look up via hipGetSymbolAddress; the actual device-side data
+// lives in the matching device-side global.
 void CGNVCUDARuntime::emitOffloadProfilingSections() {
   if (!CGM.getLangOpts().HIP)
     return;
@@ -1381,7 +1383,7 @@ void CGNVCUDARuntime::emitOffloadProfilingSections() {
     }
 
     auto *StructTy = llvm::StructType::get(
-        Ctx, {PtrTy, PtrTy, PtrTy, PtrTy, PtrTy, PtrTy, PtrTy});
+        Ctx, {PtrTy, PtrTy, PtrTy, PtrTy, PtrTy, PtrTy, PtrTy, PtrTy, PtrTy});
     llvm::Constant *Fields[] = {
         getOrDeclare("__start___llvm_prf_names"),
         getOrDeclare("__stop___llvm_prf_names"),
@@ -1389,6 +1391,8 @@ void CGNVCUDARuntime::emitOffloadProfilingSections() {
         getOrDeclare("__stop___llvm_prf_cnts"),
         getOrDeclare("__start___llvm_prf_data"),
         getOrDeclare("__stop___llvm_prf_data"),
+        getOrDeclare("__start___llvm_prf_ucnts"),
+        getOrDeclare("__stop___llvm_prf_ucnts"),
         VersionGV,
     };
     auto *Init = llvm::ConstantStruct::get(StructTy, Fields);
diff --git a/compiler-rt/include/profile/InstrProfData.inc b/compiler-rt/include/profile/InstrProfData.inc
index 3117cb2fddf36..1f45eb0bc8b2c 100644
--- a/compiler-rt/include/profile/InstrProfData.inc
+++ b/compiler-rt/include/profile/InstrProfData.inc
@@ -78,6 +78,8 @@ INSTR_PROF_DATA(const uint64_t, llvm::Type::getInt64Ty(Ctx), FuncHash, \
                 ConstantInt::get(llvm::Type::getInt64Ty(Ctx), \
                 Inc->getHash()->getZExtValue()))
 INSTR_PROF_DATA(const IntPtrT, IntPtrTy, CounterPtr, RelativeCounterPtr)
+INSTR_PROF_DATA(const IntPtrT, IntPtrTy, UniformCounterPtr, \
+                RelativeUniformCounterPtr)
 INSTR_PROF_DATA(const IntPtrT, IntPtrTy, BitmapPtr, RelativeBitmapPtr)
 /* This is used to map function pointers for the indirect call targets to
  * function name hashes during the conversion from raw to merged profile
@@ -91,6 +93,10 @@ INSTR_PROF_DATA(const uint32_t, llvm::Type::getInt32Ty(Ctx), NumCounters, \
                 ConstantInt::get(llvm::Type::getInt32Ty(Ctx), NumCounters))
 INSTR_PROF_DATA(const uint16_t, Int16ArrayTy, NumValueSites[IPVK_Last+1], \
                 ConstantArray::get(Int16ArrayTy, Int16ArrayVals)) \
+INSTR_PROF_DATA(const uint16_t, llvm::Type::getInt16Ty(Ctx), \
+                OffloadDeviceWaveSize, \
+                ConstantInt::get(llvm::Type::getInt16Ty(Ctx), \
+                                 OffloadDeviceWaveSizeVal))
 INSTR_PROF_DATA(const uint32_t, llvm::Type::getInt32Ty(Ctx), NumBitmapBytes, \
                 ConstantInt::get(llvm::Type::getInt32Ty(Ctx), NumBitmapBytes))
 #undef INSTR_PROF_DATA
@@ -168,6 +174,12 @@ INSTR_PROF_GPU_SECT(const __llvm_profile_data *, llvm::PointerType::getUnqual( \
 INSTR_PROF_GPU_SECT(const __llvm_profile_data *, llvm::PointerType::getUnqual( \
                     Ctx), DataStop,                                            \
                     ConstantPointerNull::get(llvm::PointerType::getUnqual(Ctx)))
+INSTR_PROF_GPU_SECT(char *, llvm::PointerType::getUnqual(Ctx),                 \
+                    UniformCountersStart,                                      \
+                    ConstantPointerNull::get(llvm::PointerType::getUnqual(Ctx)))
+INSTR_PROF_GPU_SECT(char *, llvm::PointerType::getUnqual(Ctx),                 \
+                    UniformCountersStop,                                       \
+                    ConstantPointerNull::get(llvm::PointerType::getUnqual(Ctx)))
 INSTR_PROF_GPU_SECT(uint64_t *, llvm::PointerType::getUnqual(Ctx),             \
                     VersionVar,                                                \
                     ConstantPointerNull::get(llvm::PointerType::getUnqual(Ctx)))
@@ -192,6 +204,12 @@ INSTR_PROF_RAW_HEADER(uint64_t, NumCounters, NumCounters)
 INSTR_PROF_RAW_HEADER(uint64_t, PaddingBytesAfterCounters, PaddingBytesAfterCounters)
 INSTR_PROF_RAW_HEADER(uint64_t, NumBitmapBytes, NumBitmapBytes)
 INSTR_PROF_RAW_HEADER(uint64_t, PaddingBytesAfterBitmapBytes, PaddingBytesAfterBitmapBytes)
+INSTR_PROF_RAW_HEADER(uint64_t, NumUniformCounters, NumUniformCounters)
+INSTR_PROF_RAW_HEADER(uint64_t, PaddingBytesAfterUniformCounters, PaddingBytesAfterUniformCounters)
+INSTR_PROF_RAW_HEADER(uint64_t, UniformCountersDelta,
+                      UniformCountersBegin ? (uintptr_t)UniformCountersBegin -
+                                                 (uintptr_t)DataBegin
+                                           : 0)
 INSTR_PROF_RAW_HEADER(uint64_t, NamesSize,  NamesSize)
 INSTR_PROF_RAW_HEADER(uint64_t, CountersDelta,
                       (uintptr_t)CountersBegin - (uintptr_t)DataBegin)
@@ -356,6 +374,9 @@ INSTR_PROF_SECT_ENTRY(IPSK_data, \
 INSTR_PROF_SECT_ENTRY(IPSK_cnts, \
                       INSTR_PROF_QUOTE(INSTR_PROF_CNTS_COMMON), \
                       INSTR_PROF_CNTS_COFF, "__DATA,")
+INSTR_PROF_SECT_ENTRY(IPSK_ucnts, \
+                      INSTR_PROF_QUOTE(INSTR_PROF_UCNTS_COMMON), \
+                      INSTR_PROF_UCNTS_COFF, "__DATA,")
 INSTR_PROF_SECT_ENTRY(IPSK_bitmap, \
                       INSTR_PROF_QUOTE(INSTR_PROF_BITS_COMMON), \
                       INSTR_PROF_BITS_COFF, "__DATA,")
@@ -752,9 +773,9 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
         (uint64_t)'f' << 16 | (uint64_t)'R' << 8 | (uint64_t)129
 
 /* Raw profile format version (start from 1). */
-#define INSTR_PROF_RAW_VERSION 10
+#define INSTR_PROF_RAW_VERSION 11
 /* Indexed profile format version (start from 1). */
-#define INSTR_PROF_INDEX_VERSION 13
+#define INSTR_PROF_INDEX_VERSION 14
 /* Coverage mapping format version (start from 0). */
 #define INSTR_PROF_COVMAP_VERSION 6
 
@@ -803,6 +824,7 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
 #define INSTR_PROF_NAME_COMMON __llvm_prf_names
 #define INSTR_PROF_VNAME_COMMON __llvm_prf_vns
 #define INSTR_PROF_CNTS_COMMON __llvm_prf_cnts
+#define INSTR_PROF_UCNTS_COMMON __llvm_prf_ucnts
 #define INSTR_PROF_BITS_COMMON __llvm_prf_bits
 #define INSTR_PROF_VALS_COMMON __llvm_prf_vals
 #define INSTR_PROF_VNODES_COMMON __llvm_prf_vnds
@@ -820,6 +842,7 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
 #define INSTR_PROF_NAME_COFF ".lprfn$M"
 #define INSTR_PROF_VNAME_COFF ".lprfvn$M"
 #define INSTR_PROF_CNTS_COFF ".lprfc$M"
+#define INSTR_PROF_UCNTS_COFF ".lprfuc$M"
 #define INSTR_PROF_BITS_COFF ".lprfb$M"
 #define INSTR_PROF_VALS_COFF ".lprfv$M"
 #define INSTR_PROF_VNODES_COFF ".lprfnd$M"
diff --git a/compiler-rt/lib/profile/InstrProfiling.h b/compiler-rt/lib/profile/InstrProfiling.h
index 6fc26b59d3cdc..595c258d19abd 100644
--- a/compiler-rt/lib/profile/InstrProfiling.h
+++ b/compiler-rt/lib/profile/InstrProfiling.h
@@ -308,10 +308,12 @@ uint64_t __llvm_profile_get_vtable_section_size(const VTableProfData *Begin,
  */
 int __llvm_profile_get_padding_sizes_for_counters(
     uint64_t DataSize, uint64_t CountersSize, uint64_t NumBitmapBytes,
-    uint64_t NamesSize, uint64_t VTableSize, uint64_t VNameSize,
-    uint64_t *PaddingBytesBeforeCounters, uint64_t *PaddingBytesAfterCounters,
-    uint64_t *PaddingBytesAfterBitmap, uint64_t *PaddingBytesAfterNames,
-    uint64_t *PaddingBytesAfterVTable, uint64_t *PaddingBytesAfterVNames);
+    uint64_t NumUniformCounters, uint64_t NamesSize, uint64_t VTableSize,
+    uint64_t VNameSize, uint64_t *PaddingBytesBeforeCounters,
+    uint64_t *PaddingBytesAfterCounters, uint64_t *PaddingBytesAfterBitmap,
+    uint64_t *PaddingBytesAfterUniformCounters,
+    uint64_t *PaddingBytesAfterNames, uint64_t *PaddingBytesAfterVTable,
+    uint64_t *PaddingBytesAfterVNames);
 
 /*!
  * \brief Set the flag that profile data has been dumped to the file.
@@ -323,14 +325,27 @@ void __llvm_profile_set_dumped(void);
 
 /*!
  * \brief Write custom target-specific profiling data to a separate file.
- * Used by offload PGO.
+ * Used by offload PGO (HIP and OpenMP).
+ *
+ * \param Target Target triple (e.g., "amdgcn-amd-amdhsa")
+ * \param DataBegin Start of profile data records
+ * \param DataEnd End of profile data records
+ * \param CountersBegin Start of counter data
+ * \param CountersEnd End of counter data
+ * \param UniformCountersBegin Start of uniform counters (NULL if not used)
+ * \param UniformCountersEnd End of uniform counters (NULL if not used)
+ * \param NamesBegin Start of names data
+ * \param NamesEnd End of names data
+ * \param VersionOverride Profile version override (NULL to use default)
  */
 int __llvm_write_custom_profile(const char *Target,
                                 const __llvm_profile_data *DataBegin,
                                 const __llvm_profile_data *DataEnd,
                                 const char *CountersBegin,
-                                const char *CountersEnd, const char *NamesBegin,
-                                const char *NamesEnd,
+                                const char *CountersEnd,
+                                const char *UniformCountersBegin,
+                                const char *UniformCountersEnd,
+                                const char *NamesBegin, const char *NamesEnd,
                                 const uint64_t *VersionOverride);
 
 /*!
@@ -364,4 +379,5 @@ extern char INSTR_PROF_PROFILE_NAME_VAR[1]; /* __llvm_profile_filename. */
 
 const __llvm_gcov_init_func_struct *__llvm_profile_begin_covinit();
 const __llvm_gcov_init_func_struct *__llvm_profile_end_covinit();
+
 #endif /* PROFILE_INSTRPROFILING_H_ */
diff --git a/compiler-rt/lib/profile/InstrProfilingBuffer.c b/compiler-rt/lib/profile/InstrProfilingBuffer.c
index b406e8db74f3f..5776a8643abcc 100644
--- a/compiler-rt/lib/profile/InstrProfilingBuffer.c
+++ b/compiler-rt/lib/profile/InstrProfilingBuffer.c
@@ -147,10 +147,12 @@ static int needsCounterPadding(void) {
 COMPILER_RT_VISIBILITY
 int __llvm_profile_get_padding_sizes_for_counters(
     uint64_t DataSize, uint64_t CountersSize, uint64_t NumBitmapBytes,
-    uint64_t NamesSize, uint64_t VTableSize, uint64_t VNameSize,
-    uint64_t *PaddingBytesBeforeCounters, uint64_t *PaddingBytesAfterCounters,
-    uint64_t *PaddingBytesAfterBitmapBytes, uint64_t *PaddingBytesAfterNames,
-    uint64_t *PaddingBytesAfterVTable, uint64_t *PaddingBytesAfterVName) {
+    uint64_t NumUniformCounters, uint64_t NamesSize, uint64_t VTableSize,
+    uint64_t VNameSize, uint64_t *PaddingBytesBeforeCounters,
+    uint64_t *PaddingBytesAfterCounters, uint64_t *PaddingBytesAfterBitmapBytes,
+    uint64_t *PaddingBytesAfterUniformCounters,
+    uint64_t *PaddingBytesAfterNames, uint64_t *PaddingBytesAfterVTable,
+    uint64_t *PaddingBytesAfterVName) {
   // Counter padding is needed only if continuous mode is enabled.
   if (!needsCounterPadding()) {
     *PaddingBytesBeforeCounters = 0;
@@ -158,6 +160,9 @@ int __llvm_profile_get_padding_sizes_for_counters(
         __llvm_profile_get_num_padding_bytes(CountersSize);
     *PaddingBytesAfterBitmapBytes =
         __llvm_profile_get_num_padding_bytes(NumBitmapBytes);
+    if (PaddingBytesAfterUniformCounters != NULL)
+      *PaddingBytesAfterUniformCounters = __llvm_profile_get_num_padding_bytes(
+          NumUniformCounters * sizeof(uint64_t));
     *PaddingBytesAfterNames = __llvm_profile_get_num_padding_bytes(NamesSize);
     if (PaddingBytesAfterVTable != NULL)
       *PaddingBytesAfterVTable =
@@ -179,6 +184,8 @@ int __llvm_profile_get_padding_sizes_for_counters(
   *PaddingBytesAfterCounters = calculateBytesNeededToPageAlign(CountersSize);
   *PaddingBytesAfterBitmapBytes =
       calculateBytesNeededToPageAlign(NumBitmapBytes);
+  if (PaddingBytesAfterUniformCounters != NULL)
+    *PaddingBytesAfterUniformCounters = 0;
   *PaddingBytesAfterNames = calculateBytesNeededToPageAlign(NamesSize);
   // Set these two variables to zero to avoid uninitialized variables
   // even if VTableSize and VNameSize are known to be zero.
@@ -212,20 +219,22 @@ uint64_t __llvm_profile_get_size_for_buffer_internal(
    * the names. */
   uint64_t PaddingBytesBeforeCounters, PaddingBytesAfterCounters,
       PaddingBytesAfterNames, PaddingBytesAfterBitmapBytes,
-      PaddingBytesAfterVTable, PaddingBytesAfterVNames;
+      PaddingBytesAfterUniformCounters, PaddingBytesAfterVTable,
+      PaddingBytesAfterVNames;
   __llvm_profile_get_padding_sizes_for_counters(
-      DataSize, CountersSize, NumBitmapBytes, NamesSize, 0 /* VTableSize */,
-      0 /* VNameSize */, &PaddingBytesBeforeCounters,
-      &PaddingBytesAfterCounters, &PaddingBytesAfterBitmapBytes,
+      DataSize, CountersSize, NumBitmapBytes, 0 /* NumUniformCounters */,
+      NamesSize, 0 /* VTableSize */, 0 /* VNameSize */,
+      &PaddingBytesBeforeCounters, &PaddingBytesAfterCounters,
+      &PaddingBytesAfterBitmapBytes, &PaddingBytesAfterUniformCounters,
       &PaddingBytesAfterNames, &PaddingBytesAfterVTable,
       &PaddingBytesAfterVNames);
 
   return sizeof(__llvm_profile_header) + __llvm_write_binary_ids(NULL) +
          DataSize + PaddingBytesBeforeCounters + CountersSize +
          PaddingBytesAfterCounters + NumBitmapBytes +
-         PaddingBytesAfterBitmapBytes + NamesSize + PaddingBytesAfterNames +
-         VTableSize + PaddingBytesAfterVTable + VNameSize +
-         PaddingBytesAfterVNames;
+         PaddingBytesAfterBitmapBytes + PaddingBytesAfterUniformCounters +
+         NamesSize + PaddingBytesAfterNames + VTableSize +
+         PaddingBytesAfterVTable + VNameSize + PaddingBytesAfterVNames;
 }
 
 COMPILER_RT_VISIBILITY
@@ -250,7 +259,8 @@ COMPILER_RT_VISIBILITY int __llvm_profile_write_buffer_internal(
   // Set virtual table arguments to NULL since they are not supported yet.
   return lprofWriteDataImpl(
       &BufferWriter, DataBegin, DataEnd, CountersBegin, CountersEnd,
-      BitmapBegin, BitmapEnd, /*VPDataReader=*/0, NamesBegin, NamesEnd,
+      BitmapBegin, BitmapEnd, /*UniformCountersBegin=*/NULL,
+      /*UniformCountersEnd=*/NULL, /*VPDataReader=*/0, NamesBegin, NamesEnd,
       /*VTableBegin=*/NULL, /*VTableEnd=*/NULL, /*VNamesBegin=*/NULL,
       /*VNamesEnd=*/NULL, /*SkipNameDataWrite=*/0,
       __llvm_profile_get_version());
diff --git a/compiler-rt/lib/profile/InstrProfilingFile.c b/compiler-rt/lib/profile/InstrProfilingFile.c
index 9ea5a2638fac9..99cb601b9ce65 100644
--- a/compiler-rt/lib/profile/InstrProfilingFile.c
+++ b/compiler-rt/lib/profile/InstrProfilingFile.c
@@ -168,11 +168,13 @@ static int mmapForContinuousMode(uint64_t CurrentFileOffset, FILE *File) {
    * after the names. */
   uint64_t PaddingBytesBeforeCounters, PaddingBytesAfterCounters,
       PaddingBytesAfterNames, PaddingBytesAfterBitmapBytes,
-      PaddingBytesAfterVTable, PaddingBytesAfterVNames;
+      PaddingBytesAfterUniformCounters, PaddingBytesAfterVTable,
+      PaddingBytesAfterVNames;
   __llvm_profile_get_padding_sizes_for_counters(
-      DataSize, CountersSize, NumBitmapBytes, NamesSize, /*VTableSize=*/0,
-      /*VNameSize=*/0, &PaddingBytesBeforeCounters, &PaddingBytesAfterCounters,
-      &PaddingBytesAfterBitmapBytes, &PaddingBytesAfterNames,
+      DataSize, CountersSize, NumBitmapBytes, 0 /* NumUniformCounters */,
+      NamesSize, /*VTableSize=*/0, /*VNameSize=*/0, &PaddingBytesBeforeCounters,
+      &PaddingBytesAfterCounters, &PaddingBytesAfterBitmapBytes,
+      &PaddingBytesAfterUniformCounters, &PaddingBytesAfterNames,
       &PaddingBytesAfterVTable, &PaddingBytesAfterVNames);
 
   uint64_t PageAlignedCountersLength = CountersSize + PaddingBytesAfterCounters;
@@ -1316,8 +1318,10 @@ int __llvm_write_custom_profile(const char *Target,
                                 const __llvm_profile_data *DataBegin,
                                 const __llvm_profile_data *DataEnd,
                                 const char *CountersBegin,
-                                const char *CountersEnd, const char *NamesBegin,
-                                const char *NamesEnd,
+                                const char *CountersEnd,
+                                const char *UniformCountersBegin,
+                                const char *UniformCountersEnd,
+                                const char *NamesBegin, const char *NamesEnd,
                                 const uint64_t *VersionOverride) {
   int ReturnValue = 0, FilenameLength, TargetLength;
   char *FilenameBuf, *TargetFilename;
@@ -1405,10 +1409,10 @@ int __llvm_write_custom_profile(const char *Target,
     Version = *VersionOverride;
 
   /* Write custom data to the file */
-  ReturnValue =
-      lprofWriteDataImpl(&fileWriter, DataBegin, DataEnd, CountersBegin,
-                         CountersEnd, NULL, NULL, lprofGetVPDataReader(), NULL,
-                         NULL, NULL, NULL, NamesBegin, NamesEnd, 0, Version);
+  ReturnValue = lprofWriteDataImpl(
+      &fileWriter, DataBegin, DataEnd, CountersBegin, CountersEnd, NULL, NULL,
+      UniformCountersBegin, UniformCountersEnd, lprofGetVPDataReader(), NULL,
+      NULL, NULL, NULL, NamesBegin, NamesEnd, 0, Version);
   closeFileObject(OutputFile);
 
   // Restore SIGKILL.
diff --git a/compiler-rt/lib/profile/InstrProfilingInternal.h b/compiler-rt/lib/profile/InstrProfilingInternal.h
index 5647782527eb7..61485915bddad 100644
--- a/compiler-rt/lib/profile/InstrProfilingInternal.h
+++ b/compiler-rt/lib/profile/InstrProfilingInternal.h
@@ -152,16 +152,15 @@ typedef struct VPDataReaderType {
    the name data is already in destination, we just skip over it. */
 int lprofWriteData(ProfDataWriter *Writer, VPDataReaderType *VPDataReader,
                    int SkipNameDataWrite);
-int lprofWriteDataImpl(ProfDataWriter *Writer,
-                       const __llvm_profile_data *DataBegin,
-                       const __llvm_profile_data *DataEnd,
-                       const char *CountersBegin, const char *CountersEnd,
-                       const char *BitmapBegin, const char *BitmapEnd,
-                       VPDataReaderType *VPDataReader, const char *NamesBegin,
-                       const char *NamesEnd, const VTableProfData *VTableBegin,
-                       const VTableProfData *VTableEnd, const char *VNamesBegin,
-                       const char *VNamesEnd, int SkipNameDataWrite,
-                       uint64_t Version);
+int lprofWriteDataImpl(
+    ProfDataWriter *Writer, const __llvm_profile_data *DataBegin,
+    const __llvm_profile_data *DataEnd, const char *CountersBegin,
+    const char *CountersEnd, const char *BitmapBegin, const char *BitmapEnd,
+    const char *UniformCountersBegin, const char *UniformCountersEnd,
+    VPDataReaderType *VPDataReader, const char *NamesBegin,
+    const char *NamesEnd, const VTableProfData *VTableBegin,
+    const VTableProfData *VTableEnd, const char *VNamesBegin,
+    const char *VNamesEnd, int SkipNameDataWrite, uint64_t Version);
 
 /* Merge value profile data pointed to by SrcValueProfData into
  * in-memory profile counters pointed by to DstData.  */
diff --git a/compiler-rt/lib/profile/InstrProfilingPlatformGPU.c b/compiler-rt/lib/profile/InstrProfilingPlatformGPU.c
index 3c67e2c7089d4..eb1507ecd25cb 100644
--- a/compiler-rt/lib/profile/InstrProfilingPlatformGPU.c
+++ b/compiler-rt/lib/profile/InstrProfilingPlatformGPU.c
@@ -16,6 +16,7 @@
 
 #include "InstrProfiling.h"
 #include <gpuintrin.h>
+#include <stdint.h>
 
 // Symbols exported to the GPU runtime need to be visible in the .dynsym table.
 #define COMPILER_RT_GPU_VISIBILITY __attribute__((visibility("protected")))
@@ -42,6 +43,34 @@ COMPILER_RT_VISIBILITY void INSTR_PROF_INSTRUMENT_GPU_FUNC(uint64_t *counter,
   }
 }
 
+// Block-level sampling for offload PGO. For GPU kernels with stationary
+// behavior (where all thread blocks execute the same code paths regardless of
+// block ID), partial sampling significantly reduces instrumentation overhead
+// without losing PGO performance gains.
+//
+// Returns 1 if this block should be instrumented, 0 to skip. Samples by
+// matching lower bits of the x-dimension block ID to zero.
+//   sampling_bits=0: all blocks (100%)
+//   sampling_bits=3: every 8th block in x (12.5%, default)
+//
+// Note: We use only block_id_x rather than a fully linearized 3D block ID.
+// The 3D linearization requires __gpu_num_blocks_x/y which expands to
+// __builtin_amdgcn_workgroup_size_x/y. With -mcode-object-version=none (used
+// to build compiler-rt profile runtime), the compiler emits a load of
+// __oclc_ABI_version to select the correct ABI path. Since the profile runtime
+// is linked after device libs are internalized, __oclc_ABI_version is no longer
+// available. Using block_id_x directly avoids this dependency. For typical
+// kernels with large 1D or x-dominant grids this is sufficient; blocks sharing
+// the same x-index are sampled together in 3D grids (minor uniformity loss).
+COMPILER_RT_VISIBILITY int __llvm_profile_sampling_gpu(uint32_t sampling_bits) {
+  if (sampling_bits == 0)
+    return 1;
+
+  uint32_t block_id = __gpu_block_id_x();
+  uint32_t mask = (1u << sampling_bits) - 1;
+  return (block_id & mask) == 0;
+}
+
 #if defined(__AMDGPU__)
 
 #define PROF_NAME_START INSTR_PROF_SECT_START(INSTR_PROF_NAME_COMMON)
@@ -50,6 +79,8 @@ COMPILER_RT_VISIBILITY void INSTR_PROF_INSTRUMENT_GPU_FUNC(uint64_t *counter,
 #define PROF_CNTS_STOP INSTR_PROF_SECT_STOP(INSTR_PROF_CNTS_COMMON)
 #define PROF_DATA_START INSTR_PROF_SECT_START(INSTR_PROF_DATA_COMMON)
 #define PROF_DATA_STOP INSTR_PROF_SECT_STOP(INSTR_PROF_DATA_COMMON)
+#define PROF_UCNTS_START INSTR_PROF_SECT_START(INSTR_PROF_UCNTS_COMMON)
+#define PROF_UCNTS_STOP INSTR_PROF_SECT_STOP(INSTR_PROF_UCNTS_COMMON)
 
 extern char PROF_NAME_START[] COMPILER_RT_VISIBILITY COMPILER_RT_WEAK;
 extern char PROF_NAME_STOP[] COMPILER_RT_VISIBILITY COMPILER_RT_WEAK;
@@ -59,17 +90,15 @@ extern __llvm_profile_data PROF_DATA_START[] COMPILER_RT_VISIBILITY
     COMPILER_RT_WEAK;
 extern __llvm_profile_data PROF_DATA_STOP[] COMPILER_RT_VISIBILITY
     COMPILER_RT_WEAK;
+extern char PROF_UCNTS_START[] COMPILER_RT_VISIBILITY COMPILER_RT_WEAK;
+extern char PROF_UCNTS_STOP[] COMPILER_RT_VISIBILITY COMPILER_RT_WEAK;
 
 // AMDGPU is a proper ELF target and exports the linker-defined section bounds.
 COMPILER_RT_GPU_VISIBILITY
 __llvm_profile_gpu_sections INSTR_PROF_SECT_BOUNDS_TABLE = {
-    PROF_NAME_START,
-    PROF_NAME_STOP,
-    PROF_CNTS_START,
-    PROF_CNTS_STOP,
-    PROF_DATA_START,
-    PROF_DATA_STOP,
-    &INSTR_PROF_RAW_VERSION_VAR};
+    PROF_NAME_START,  PROF_NAME_STOP,  PROF_CNTS_START,
+    PROF_CNTS_STOP,   PROF_DATA_START, PROF_DATA_STOP,
+    PROF_UCNTS_START, PROF_UCNTS_STOP, &INSTR_PROF_RAW_VERSION_VAR};
 
 #elif defined(__NVPTX__)
 
@@ -79,8 +108,9 @@ __llvm_profile_gpu_sections INSTR_PROF_SECT_BOUNDS_TABLE = {
 //        globals next to each other.
 COMPILER_RT_GPU_VISIBILITY
 __llvm_profile_gpu_sections INSTR_PROF_SECT_BOUNDS_TABLE = {
-    NULL, NULL, NULL, NULL, NULL, NULL, &INSTR_PROF_RAW_VERSION_VAR};
-
+    NULL, NULL, NULL,
+    NULL, NULL, NULL,
+    NULL, NULL, &INSTR_PROF_RAW_VERSION_VAR};
 #endif
 
 #endif
diff --git a/compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp b/compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp
index ee00c572e3a42..410fcfdd8d0c8 100644
--- a/compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp
+++ b/compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp
@@ -560,29 +560,35 @@ static int processDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex,
   const void *DevCntsBegin = HostSections.CountersStart;
   const void *DevDataBegin = HostSections.DataStart;
   const void *DevNamesBegin = HostSections.NamesStart;
+  const void *DevUniformCntsBegin = HostSections.UniformCountersStart;
   const void *DevCntsEnd = HostSections.CountersStop;
   const void *DevDataEnd = HostSections.DataStop;
   const void *DevNamesEnd = HostSections.NamesStop;
+  const void *DevUniformCntsEnd = HostSections.UniformCountersStop;
 
   size_t CountersSize = (const char *)DevCntsEnd - (const char *)DevCntsBegin;
   size_t DataSize = (const char *)DevDataEnd - (const char *)DevDataBegin;
   size_t NamesSize = (const char *)DevNamesEnd - (const char *)DevNamesBegin;
+  size_t UniformCountersSize =
+      (const char *)DevUniformCntsEnd - (const char *)DevUniformCntsBegin;
 
   if (isVerboseMode())
     PROF_NOTE("Section pointers: Cnts=[%p,%p]=%zu Data=[%p,%p]=%zu "
-              "Names=[%p,%p]=%zu\n",
+              "Names=[%p,%p]=%zu UCnts=[%p,%p]=%zu\n",
               DevCntsBegin, DevCntsEnd, CountersSize, DevDataBegin, DevDataEnd,
-              DataSize, DevNamesBegin, DevNamesEnd, NamesSize);
+              DataSize, DevNamesBegin, DevNamesEnd, NamesSize,
+              DevUniformCntsBegin, DevUniformCntsEnd, UniformCountersSize);
 
   if (CountersSize == 0 || DataSize == 0)
     return 0;
 
   int ret = -1;
-  int NamesReused = 0, CntsReused = 0, DataReused = 0;
+  int NamesReused = 0, CntsReused = 0, UCntsReused = 0, DataReused = 0;
 
   char *HostDataBegin = nullptr;
   char *HostCountersBegin = nullptr;
   char *HostNamesBegin = nullptr;
+  char *HostUniformCountersBegin = nullptr;
 
   /* Sections using linker-defined __start_/__stop_ bounds are shared across
      TU structs in RDC mode. Deduplicate by caching the last copied range. */
@@ -598,8 +604,12 @@ static int processDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex,
   static char *CachedHostData = nullptr;
   static size_t CachedDataSize = 0;
 
+  static const void *CachedDevUCntsBegin = nullptr;
+  static char *CachedHostUCnts = nullptr;
+  static size_t CachedUCntsSize = 0;
+
   // Owns freshly malloc'd buffers; release() transfers ownership to the cache.
-  UniqueFree CntsOwner, DataOwner, NamesOwner;
+  UniqueFree CntsOwner, DataOwner, NamesOwner, UCntsOwner;
 
   if (CountersSize > 0 && DevCntsBegin == CachedDevCntsBegin &&
       CountersSize == CachedCntsSize) {
@@ -634,9 +644,22 @@ static int processDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex,
     NamesOwner.reset(HostNamesBegin);
   }
 
+  if (UniformCountersSize > 0 && DevUniformCntsBegin == CachedDevUCntsBegin &&
+      UniformCountersSize == CachedUCntsSize) {
+    HostUniformCountersBegin = CachedHostUCnts;
+    UCntsReused = 1;
+    if (isVerboseMode())
+      PROF_NOTE("Reusing cached ucnts section (%zu bytes)\n",
+                UniformCountersSize);
+  } else if (UniformCountersSize > 0) {
+    HostUniformCountersBegin = (char *)malloc(UniformCountersSize);
+    UCntsOwner.reset(HostUniformCountersBegin);
+  }
+
   if ((DataSize > 0 && !HostDataBegin) ||
       (CountersSize > 0 && !HostCountersBegin) ||
-      (NamesSize > 0 && !HostNamesBegin)) {
+      (NamesSize > 0 && !HostNamesBegin) ||
+      (UniformCountersSize > 0 && !HostUniformCountersBegin)) {
     PROF_ERR("%s\n", "failed to allocate host memory for device sections");
     return -1;
   }
@@ -647,7 +670,10 @@ static int processDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex,
        memcpyDeviceToHost(HostCountersBegin, DevCntsBegin, CountersSize) !=
            0) ||
       (NamesSize > 0 && !NamesReused &&
-       memcpyDeviceToHost(HostNamesBegin, DevNamesBegin, NamesSize) != 0)) {
+       memcpyDeviceToHost(HostNamesBegin, DevNamesBegin, NamesSize) != 0) ||
+      (UniformCountersSize > 0 && !UCntsReused &&
+       memcpyDeviceToHost(HostUniformCountersBegin, DevUniformCntsBegin,
+                          UniformCountersSize) != 0)) {
     PROF_ERR("%s\n", "failed to copy profile sections from device");
     return -1;
   }
@@ -672,25 +698,34 @@ static int processDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex,
     CachedNamesSize = NamesSize;
     NamesOwner.release();
   }
+  if (!UCntsReused && UniformCountersSize > 0) {
+    CachedDevUCntsBegin = DevUniformCntsBegin;
+    CachedHostUCnts = HostUniformCountersBegin;
+    CachedUCntsSize = UniformCountersSize;
+    UCntsOwner.release();
+  }
 
   if (isVerboseMode())
-    PROF_NOTE("Copied device sections: Counters=%zu, Data=%zu, Names=%zu\n",
-              CountersSize, DataSize, NamesSize);
+    PROF_NOTE("Copied device sections: Counters=%zu, Data=%zu, Names=%zu, "
+              "UniformCounters=%zu\n",
+              CountersSize, DataSize, NamesSize, UniformCountersSize);
 
   // Arrange buffer as [Data][Padding][Counters][Names] to match the layout
   // expected by lprofWriteDataImpl (CountersDelta = CountersBegin - DataBegin).
   const uint64_t NumData = DataSize / sizeof(__llvm_profile_data);
   const uint64_t NumBitmapBytes = 0;
+  const uint64_t NumUniformCounters = UniformCountersSize / sizeof(uint64_t);
   const uint64_t VTableSectionSize = 0;
   const uint64_t VNamesSize = 0;
   uint64_t PaddingBytesBeforeCounters, PaddingBytesAfterCounters,
-      PaddingBytesAfterBitmapBytes, PaddingBytesAfterNames,
-      PaddingBytesAfterVTable, PaddingBytesAfterVNames;
+      PaddingBytesAfterBitmapBytes, PaddingBytesAfterUniformCounters,
+      PaddingBytesAfterNames, PaddingBytesAfterVTable, PaddingBytesAfterVNames;
 
   if (__llvm_profile_get_padding_sizes_for_counters(
-          DataSize, CountersSize, NumBitmapBytes, NamesSize, VTableSectionSize,
-          VNamesSize, &PaddingBytesBeforeCounters, &PaddingBytesAfterCounters,
-          &PaddingBytesAfterBitmapBytes, &PaddingBytesAfterNames,
+          DataSize, CountersSize, NumBitmapBytes, NumUniformCounters, NamesSize,
+          VTableSectionSize, VNamesSize, &PaddingBytesBeforeCounters,
+          &PaddingBytesAfterCounters, &PaddingBytesAfterBitmapBytes,
+          &PaddingBytesAfterUniformCounters, &PaddingBytesAfterNames,
           &PaddingBytesAfterVTable, &PaddingBytesAfterVNames) != 0) {
     PROF_ERR("%s\n", "failed to get padding sizes");
     return -1;
@@ -715,19 +750,25 @@ static int processDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex,
   __builtin_memcpy(BufCountersBegin, HostCountersBegin, CountersSize);
   __builtin_memcpy(BufNamesBegin, HostNamesBegin, NamesSize);
 
-  // CounterPtr is a device-relative offset; relocate it for the file layout
-  // where the Data section precedes Counters.
+  // CounterPtr and UniformCounterPtr are device-relative offsets; relocate
+  // them for the file layout where the Data section precedes the Counters and
+  // UniformCounters sections. Uniform counters are copied in linker (section)
+  // order and located via their relative pointer, exactly like the regular
+  // counters: llvm-profdata reads them through UniformCounterPtr (decrementing
+  // UniformCountersDelta per record, just like CountersDelta) and does not
+  // assume data-record order, so no reordering is needed.
+  ptrdiff_t UCFileOffset =
+      DataSize + PaddingBytesBeforeCounters + CountersSize +
+      PaddingBytesAfterCounters + NumBitmapBytes + PaddingBytesAfterBitmapBytes;
   __llvm_profile_data *RelocatedData = (__llvm_profile_data *)BufDataBegin;
   for (uint64_t i = 0; i < NumData; ++i) {
+    const char *DeviceDataStructAddr =
+        (const char *)DevDataBegin + (i * sizeof(__llvm_profile_data));
     if (RelocatedData[i].CounterPtr) {
-      ptrdiff_t DeviceCounterPtrOffset = (ptrdiff_t)RelocatedData[i].CounterPtr;
-      const char *DeviceDataStructAddr =
-          (const char *)DevDataBegin + (i * sizeof(__llvm_profile_data));
       const char *DeviceCountersAddr =
-          DeviceDataStructAddr + DeviceCounterPtrOffset;
+          DeviceDataStructAddr + (ptrdiff_t)RelocatedData[i].CounterPtr;
       ptrdiff_t OffsetIntoCountersSection =
           DeviceCountersAddr - (const char *)DevCntsBegin;
-
       ptrdiff_t NewRelativeOffset = DataSize + PaddingBytesBeforeCounters +
                                     OffsetIntoCountersSection -
                                     (i * sizeof(__llvm_profile_data));
@@ -735,6 +776,21 @@ static int processDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex,
                            offsetof(__llvm_profile_data, CounterPtr),
                        &NewRelativeOffset, sizeof(NewRelativeOffset));
     }
+    if (HostUniformCountersBegin && RelocatedData[i].UniformCounterPtr) {
+      const char *DeviceUCAddr =
+          DeviceDataStructAddr + (ptrdiff_t)RelocatedData[i].UniformCounterPtr;
+      ptrdiff_t OffsetIntoUCSection =
+          DeviceUCAddr - (const char *)DevUniformCntsBegin;
+      ptrdiff_t NewUCRelativeOffset = UCFileOffset + OffsetIntoUCSection -
+                                      (i * sizeof(__llvm_profile_data));
+      __builtin_memcpy((char *)RelocatedData + i * sizeof(__llvm_profile_data) +
+                           offsetof(__llvm_profile_data, UniformCounterPtr),
+                       &NewUCRelativeOffset, sizeof(NewUCRelativeOffset));
+    } else {
+      __builtin_memset((char *)RelocatedData + i * sizeof(__llvm_profile_data) +
+                           offsetof(__llvm_profile_data, UniformCounterPtr),
+                       0, sizeof(RelocatedData[i].UniformCounterPtr));
+    }
     __builtin_memset((char *)RelocatedData + i * sizeof(__llvm_profile_data) +
                          offsetof(__llvm_profile_data, BitmapPtr),
                      0,
@@ -749,8 +805,10 @@ static int processDeviceOffloadPrf(void *DeviceOffloadPrf, int TUIndex,
   ret = __llvm_write_custom_profile(
       Target, (__llvm_profile_data *)BufDataBegin,
       (__llvm_profile_data *)(BufDataBegin + DataSize), BufCountersBegin,
-      BufCountersBegin + CountersSize, BufNamesBegin, BufNamesBegin + NamesSize,
-      nullptr);
+      BufCountersBegin + CountersSize, HostUniformCountersBegin,
+      HostUniformCountersBegin ? HostUniformCountersBegin + UniformCountersSize
+                               : nullptr,
+      BufNamesBegin, BufNamesBegin + NamesSize, nullptr);
 
   if (ret != 0) {
     PROF_ERR("%s\n", "failed to write device profile using shared API");
diff --git a/compiler-rt/lib/profile/InstrProfilingWriter.c b/compiler-rt/lib/profile/InstrProfilingWriter.c
index 633fdb9661162..2c459acd39af0 100644
--- a/compiler-rt/lib/profile/InstrProfilingWriter.c
+++ b/compiler-rt/lib/profile/InstrProfilingWriter.c
@@ -256,7 +256,9 @@ COMPILER_RT_VISIBILITY int lprofWriteData(ProfDataWriter *Writer,
   const char *VNamesEnd = __llvm_profile_end_vtabnames();
   uint64_t Version = __llvm_profile_get_version();
   return lprofWriteDataImpl(Writer, DataBegin, DataEnd, CountersBegin,
-                            CountersEnd, BitmapBegin, BitmapEnd, VPDataReader,
+                            CountersEnd, BitmapBegin, BitmapEnd,
+                            /*UniformCountersBegin=*/NULL,
+                            /*UniformCountersEnd=*/NULL, VPDataReader,
                             NamesBegin, NamesEnd, VTableBegin, VTableEnd,
                             VNamesBegin, VNamesEnd, SkipNameDataWrite, Version);
 }
@@ -265,6 +267,7 @@ COMPILER_RT_VISIBILITY int lprofWriteDataImpl(
     ProfDataWriter *Writer, const __llvm_profile_data *DataBegin,
     const __llvm_profile_data *DataEnd, const char *CountersBegin,
     const char *CountersEnd, const char *BitmapBegin, const char *BitmapEnd,
+    const char *UniformCountersBegin, const char *UniformCountersEnd,
     VPDataReaderType *VPDataReader, const char *NamesBegin,
     const char *NamesEnd, const VTableProfData *VTableBegin,
     const VTableProfData *VTableEnd, const char *VNamesBegin,
@@ -286,6 +289,13 @@ COMPILER_RT_VISIBILITY int lprofWriteDataImpl(
       __llvm_profile_get_vtable_section_size(VTableBegin, VTableEnd);
   const uint64_t VNamesSize =
       __llvm_profile_get_name_size(VNamesBegin, VNamesEnd);
+  const uint64_t NumUniformCounters =
+      (UniformCountersBegin && UniformCountersEnd &&
+       UniformCountersEnd > UniformCountersBegin)
+          ? (UniformCountersEnd - UniformCountersBegin) / sizeof(uint64_t)
+          : 0;
+  const uint64_t UniformCountersSectionSize =
+      NumUniformCounters * sizeof(uint64_t);
 
   /* Create the header. */
   __llvm_profile_header Header;
@@ -293,12 +303,13 @@ COMPILER_RT_VISIBILITY int lprofWriteDataImpl(
   /* Determine how much padding is needed before/after the counters and after
    * the names. */
   uint64_t PaddingBytesBeforeCounters, PaddingBytesAfterCounters,
-      PaddingBytesAfterBitmapBytes, PaddingBytesAfterNames,
-      PaddingBytesAfterVTable, PaddingBytesAfterVNames;
+      PaddingBytesAfterBitmapBytes, PaddingBytesAfterUniformCounters,
+      PaddingBytesAfterNames, PaddingBytesAfterVTable, PaddingBytesAfterVNames;
   if (__llvm_profile_get_padding_sizes_for_counters(
-          DataSectionSize, CountersSectionSize, NumBitmapBytes, NamesSize,
-          VTableSectionSize, VNamesSize, &PaddingBytesBeforeCounters,
-          &PaddingBytesAfterCounters, &PaddingBytesAfterBitmapBytes,
+          DataSectionSize, CountersSectionSize, NumBitmapBytes,
+          NumUniformCounters, NamesSize, VTableSectionSize, VNamesSize,
+          &PaddingBytesBeforeCounters, &PaddingBytesAfterCounters,
+          &PaddingBytesAfterBitmapBytes, &PaddingBytesAfterUniformCounters,
           &PaddingBytesAfterNames, &PaddingBytesAfterVTable,
           &PaddingBytesAfterVNames) == -1)
     return -1;
@@ -315,12 +326,25 @@ COMPILER_RT_VISIBILITY int lprofWriteDataImpl(
 #ifdef _WIN64
   Header.CountersDelta = (uint32_t)Header.CountersDelta;
   Header.BitmapDelta = (uint32_t)Header.BitmapDelta;
+  Header.UniformCountersDelta = (uint32_t)Header.UniformCountersDelta;
 #endif
 
+  /* Recompute UniformCountersDelta from file layout. The macro initializer
+     uses in-memory pointer arithmetic which is wrong when sections are in
+     separate allocations (e.g., custom profiles from device PGO). */
+  if (NumUniformCounters > 0)
+    Header.UniformCountersDelta = DataSectionSize + PaddingBytesBeforeCounters +
+                                  CountersSectionSize +
+                                  PaddingBytesAfterCounters + NumBitmapBytes +
+                                  PaddingBytesAfterBitmapBytes;
+  else
+    Header.UniformCountersDelta = 0;
+
   /* The data and names sections are omitted in lightweight mode. */
   if (NumData == 0 && NamesSize == 0) {
     Header.CountersDelta = 0;
     Header.NamesDelta = 0;
+    Header.UniformCountersDelta = 0;
   }
 
   /* Write the profile header. */
@@ -340,6 +364,8 @@ COMPILER_RT_VISIBILITY int lprofWriteDataImpl(
       {NULL, sizeof(uint8_t), PaddingBytesAfterCounters, 1},
       {BitmapBegin, sizeof(uint8_t), NumBitmapBytes, 0},
       {NULL, sizeof(uint8_t), PaddingBytesAfterBitmapBytes, 1},
+      {UniformCountersBegin, sizeof(uint8_t), UniformCountersSectionSize, 0},
+      {NULL, sizeof(uint8_t), PaddingBytesAfterUniformCounters, 1},
       {SkipNameDataWrite ? NULL : NamesBegin, sizeof(uint8_t), NamesSize, 0},
       {NULL, sizeof(uint8_t), PaddingBytesAfterNames, 1},
       {VTableBegin, sizeof(uint8_t), VTableSectionSize, 0},
diff --git a/llvm/include/llvm/IR/FixedMetadataKinds.def b/llvm/include/llvm/IR/FixedMetadataKinds.def
index 552c1f60f2d38..6271633ca6d66 100644
--- a/llvm/include/llvm/IR/FixedMetadataKinds.def
+++ b/llvm/include/llvm/IR/FixedMetadataKinds.def
@@ -64,3 +64,4 @@ LLVM_FIXED_MD_KIND(MD_inline_history, "inline_history", 49)
 LLVM_FIXED_MD_KIND(MD_elf_section_properties, "elf_section_properties", 50)
 LLVM_FIXED_MD_KIND(MD_unique_id, "guid", 51)
 LLVM_FIXED_MD_KIND(MD_mem_cache_hint, "mem.cache_hint", 52)
+LLVM_FIXED_MD_KIND(MD_block_uniformity_profile, "block.uniformity.profile", 53)
diff --git a/llvm/include/llvm/IR/RuntimeLibcalls.td b/llvm/include/llvm/IR/RuntimeLibcalls.td
index 37bad559f49e7..a703fe773ef88 100644
--- a/llvm/include/llvm/IR/RuntimeLibcalls.td
+++ b/llvm/include/llvm/IR/RuntimeLibcalls.td
@@ -502,6 +502,7 @@ def RETURN_ADDRESS : RuntimeLibcall;
 
 // GPU profiling
 def PROFILE_INSTRUMENT_GPU : RuntimeLibcall;
+def PROFILE_SAMPLING_GPU : RuntimeLibcall;
 
 // Clear cache
 def CLEAR_CACHE : RuntimeLibcall;
@@ -2245,9 +2246,11 @@ def WindowsARM64ECSystemLibrary
 def isAMDGPU : RuntimeLibcallPredicate<"TT.isAMDGPU()">;
 
 def __llvm_profile_instrument_gpu : RuntimeLibcallImpl<PROFILE_INSTRUMENT_GPU>;
+def __llvm_profile_sampling_gpu : RuntimeLibcallImpl<PROFILE_SAMPLING_GPU>;
 
 def AMDGPUSystemLibrary
-    : SystemRuntimeLibrary<isAMDGPU, (add __llvm_profile_instrument_gpu)>;
+    : SystemRuntimeLibrary<isAMDGPU, (add __llvm_profile_instrument_gpu,
+                                      __llvm_profile_sampling_gpu)>;
 
 //===----------------------------------------------------------------------===//
 // ARM Runtime Libcalls
diff --git a/llvm/include/llvm/ProfileData/InstrProf.h b/llvm/include/llvm/ProfileData/InstrProf.h
index b7f917ec39b70..7bf3932074ebc 100644
--- a/llvm/include/llvm/ProfileData/InstrProf.h
+++ b/llvm/include/llvm/ProfileData/InstrProf.h
@@ -902,6 +902,16 @@ struct InstrProfValueSiteRecord {
 struct InstrProfRecord {
   std::vector<uint64_t> Counts;
   std::vector<uint8_t> BitmapBytes;
+  /// For AMDGPU offload profiling: raw uniform counters embedded in the
+  /// profraw file. One uint64_t per instrumented block, tracking wave-uniform
+  /// execution counts on the GPU.
+  std::vector<uint64_t> UniformCounts;
+  /// For AMDGPU offload profiling: 1 bit per basic block indicating whether
+  /// the block is entered via a wave-uniform branch. Set during merge when
+  /// per-slot counters are reduced. If a counter value is a multiple of the
+  /// wave size, the branch is considered wave-uniform.
+  std::vector<uint8_t> UniformityBits;
+  uint16_t OffloadDeviceWaveSize = 0;
 
   InstrProfRecord() = default;
   InstrProfRecord(std::vector<uint64_t> Counts) : Counts(std::move(Counts)) {}
@@ -911,6 +921,8 @@ struct InstrProfRecord {
   InstrProfRecord(InstrProfRecord &&) = default;
   InstrProfRecord(const InstrProfRecord &RHS)
       : Counts(RHS.Counts), BitmapBytes(RHS.BitmapBytes),
+        UniformCounts(RHS.UniformCounts), UniformityBits(RHS.UniformityBits),
+        OffloadDeviceWaveSize(RHS.OffloadDeviceWaveSize),
         ValueData(RHS.ValueData
                       ? std::make_unique<ValueProfData>(*RHS.ValueData)
                       : nullptr) {}
@@ -918,6 +930,8 @@ struct InstrProfRecord {
   InstrProfRecord &operator=(const InstrProfRecord &RHS) {
     Counts = RHS.Counts;
     BitmapBytes = RHS.BitmapBytes;
+    UniformityBits = RHS.UniformityBits;
+    OffloadDeviceWaveSize = RHS.OffloadDeviceWaveSize;
     if (!RHS.ValueData) {
       ValueData = nullptr;
       return *this;
@@ -929,6 +943,17 @@ struct InstrProfRecord {
     return *this;
   }
 
+  /// Check if a basic block is entered via a wave-uniform branch.
+  /// Returns true if uniform (safe for PGO spill optimization) or if no
+  /// uniformity data is available (conservative default).
+  bool isBlockUniform(unsigned BlockIdx) const {
+    if (UniformityBits.empty())
+      return true; // No uniformity data, assume uniform (conservative)
+    if (BlockIdx / 8 >= UniformityBits.size())
+      return true; // Out of range, assume uniform
+    return (UniformityBits[BlockIdx / 8] >> (BlockIdx % 8)) & 1;
+  }
+
   /// Return the number of value profile kinds with non-zero number
   /// of profile sites.
   inline uint32_t getNumValueKinds() const;
@@ -953,8 +978,12 @@ struct InstrProfRecord {
 
   /// Merge the counts in \p Other into this one.
   /// Optionally scale merged counts by \p Weight.
+  /// If \p WaveSize is non-zero and Other has offload profiling slots,
+  /// compute uniformity bits based on whether counter values are multiples
+  /// of WaveSize.
   LLVM_ABI void merge(InstrProfRecord &Other, uint64_t Weight,
-                      function_ref<void(instrprof_error)> Warn);
+                      function_ref<void(instrprof_error)> Warn,
+                      unsigned WaveSize = 0);
 
   /// Scale up profile counts (including value profile data) by
   /// a factor of (N / D).
@@ -1079,6 +1108,14 @@ struct NamedInstrProfRecord : InstrProfRecord {
                        std::vector<uint8_t> BitmapBytes)
       : InstrProfRecord(std::move(Counts), std::move(BitmapBytes)), Name(Name),
         Hash(Hash) {}
+  NamedInstrProfRecord(StringRef Name, uint64_t Hash,
+                       std::vector<uint64_t> Counts,
+                       std::vector<uint8_t> BitmapBytes,
+                       std::vector<uint8_t> UniformityBits)
+      : InstrProfRecord(std::move(Counts), std::move(BitmapBytes)), Name(Name),
+        Hash(Hash) {
+    this->UniformityBits = std::move(UniformityBits);
+  }
 
   static bool hasCSFlagInHash(uint64_t FuncHash) {
     return ((FuncHash >> CS_FLAG_IN_FUNC_HASH) & 1);
@@ -1185,7 +1222,9 @@ enum ProfVersion {
   Version12 = 12,
   // In this version, the frontend PGO stable hash algorithm defaults to V4.
   Version13 = 13,
-  // The current version is 13.
+  // UniformityBits added for AMDGPU offload profiling divergence detection.
+  Version14 = 14,
+  // The current version is 14.
   CurrentVersion = INSTR_PROF_INDEX_VERSION
 };
 const uint64_t Version = ProfVersion::CurrentVersion;
diff --git a/llvm/include/llvm/ProfileData/InstrProfData.inc b/llvm/include/llvm/ProfileData/InstrProfData.inc
index 3117cb2fddf36..1f45eb0bc8b2c 100644
--- a/llvm/include/llvm/ProfileData/InstrProfData.inc
+++ b/llvm/include/llvm/ProfileData/InstrProfData.inc
@@ -78,6 +78,8 @@ INSTR_PROF_DATA(const uint64_t, llvm::Type::getInt64Ty(Ctx), FuncHash, \
                 ConstantInt::get(llvm::Type::getInt64Ty(Ctx), \
                 Inc->getHash()->getZExtValue()))
 INSTR_PROF_DATA(const IntPtrT, IntPtrTy, CounterPtr, RelativeCounterPtr)
+INSTR_PROF_DATA(const IntPtrT, IntPtrTy, UniformCounterPtr, \
+                RelativeUniformCounterPtr)
 INSTR_PROF_DATA(const IntPtrT, IntPtrTy, BitmapPtr, RelativeBitmapPtr)
 /* This is used to map function pointers for the indirect call targets to
  * function name hashes during the conversion from raw to merged profile
@@ -91,6 +93,10 @@ INSTR_PROF_DATA(const uint32_t, llvm::Type::getInt32Ty(Ctx), NumCounters, \
                 ConstantInt::get(llvm::Type::getInt32Ty(Ctx), NumCounters))
 INSTR_PROF_DATA(const uint16_t, Int16ArrayTy, NumValueSites[IPVK_Last+1], \
                 ConstantArray::get(Int16ArrayTy, Int16ArrayVals)) \
+INSTR_PROF_DATA(const uint16_t, llvm::Type::getInt16Ty(Ctx), \
+                OffloadDeviceWaveSize, \
+                ConstantInt::get(llvm::Type::getInt16Ty(Ctx), \
+                                 OffloadDeviceWaveSizeVal))
 INSTR_PROF_DATA(const uint32_t, llvm::Type::getInt32Ty(Ctx), NumBitmapBytes, \
                 ConstantInt::get(llvm::Type::getInt32Ty(Ctx), NumBitmapBytes))
 #undef INSTR_PROF_DATA
@@ -168,6 +174,12 @@ INSTR_PROF_GPU_SECT(const __llvm_profile_data *, llvm::PointerType::getUnqual( \
 INSTR_PROF_GPU_SECT(const __llvm_profile_data *, llvm::PointerType::getUnqual( \
                     Ctx), DataStop,                                            \
                     ConstantPointerNull::get(llvm::PointerType::getUnqual(Ctx)))
+INSTR_PROF_GPU_SECT(char *, llvm::PointerType::getUnqual(Ctx),                 \
+                    UniformCountersStart,                                      \
+                    ConstantPointerNull::get(llvm::PointerType::getUnqual(Ctx)))
+INSTR_PROF_GPU_SECT(char *, llvm::PointerType::getUnqual(Ctx),                 \
+                    UniformCountersStop,                                       \
+                    ConstantPointerNull::get(llvm::PointerType::getUnqual(Ctx)))
 INSTR_PROF_GPU_SECT(uint64_t *, llvm::PointerType::getUnqual(Ctx),             \
                     VersionVar,                                                \
                     ConstantPointerNull::get(llvm::PointerType::getUnqual(Ctx)))
@@ -192,6 +204,12 @@ INSTR_PROF_RAW_HEADER(uint64_t, NumCounters, NumCounters)
 INSTR_PROF_RAW_HEADER(uint64_t, PaddingBytesAfterCounters, PaddingBytesAfterCounters)
 INSTR_PROF_RAW_HEADER(uint64_t, NumBitmapBytes, NumBitmapBytes)
 INSTR_PROF_RAW_HEADER(uint64_t, PaddingBytesAfterBitmapBytes, PaddingBytesAfterBitmapBytes)
+INSTR_PROF_RAW_HEADER(uint64_t, NumUniformCounters, NumUniformCounters)
+INSTR_PROF_RAW_HEADER(uint64_t, PaddingBytesAfterUniformCounters, PaddingBytesAfterUniformCounters)
+INSTR_PROF_RAW_HEADER(uint64_t, UniformCountersDelta,
+                      UniformCountersBegin ? (uintptr_t)UniformCountersBegin -
+                                                 (uintptr_t)DataBegin
+                                           : 0)
 INSTR_PROF_RAW_HEADER(uint64_t, NamesSize,  NamesSize)
 INSTR_PROF_RAW_HEADER(uint64_t, CountersDelta,
                       (uintptr_t)CountersBegin - (uintptr_t)DataBegin)
@@ -356,6 +374,9 @@ INSTR_PROF_SECT_ENTRY(IPSK_data, \
 INSTR_PROF_SECT_ENTRY(IPSK_cnts, \
                       INSTR_PROF_QUOTE(INSTR_PROF_CNTS_COMMON), \
                       INSTR_PROF_CNTS_COFF, "__DATA,")
+INSTR_PROF_SECT_ENTRY(IPSK_ucnts, \
+                      INSTR_PROF_QUOTE(INSTR_PROF_UCNTS_COMMON), \
+                      INSTR_PROF_UCNTS_COFF, "__DATA,")
 INSTR_PROF_SECT_ENTRY(IPSK_bitmap, \
                       INSTR_PROF_QUOTE(INSTR_PROF_BITS_COMMON), \
                       INSTR_PROF_BITS_COFF, "__DATA,")
@@ -752,9 +773,9 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
         (uint64_t)'f' << 16 | (uint64_t)'R' << 8 | (uint64_t)129
 
 /* Raw profile format version (start from 1). */
-#define INSTR_PROF_RAW_VERSION 10
+#define INSTR_PROF_RAW_VERSION 11
 /* Indexed profile format version (start from 1). */
-#define INSTR_PROF_INDEX_VERSION 13
+#define INSTR_PROF_INDEX_VERSION 14
 /* Coverage mapping format version (start from 0). */
 #define INSTR_PROF_COVMAP_VERSION 6
 
@@ -803,6 +824,7 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
 #define INSTR_PROF_NAME_COMMON __llvm_prf_names
 #define INSTR_PROF_VNAME_COMMON __llvm_prf_vns
 #define INSTR_PROF_CNTS_COMMON __llvm_prf_cnts
+#define INSTR_PROF_UCNTS_COMMON __llvm_prf_ucnts
 #define INSTR_PROF_BITS_COMMON __llvm_prf_bits
 #define INSTR_PROF_VALS_COMMON __llvm_prf_vals
 #define INSTR_PROF_VNODES_COMMON __llvm_prf_vnds
@@ -820,6 +842,7 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
 #define INSTR_PROF_NAME_COFF ".lprfn$M"
 #define INSTR_PROF_VNAME_COFF ".lprfvn$M"
 #define INSTR_PROF_CNTS_COFF ".lprfc$M"
+#define INSTR_PROF_UCNTS_COFF ".lprfuc$M"
 #define INSTR_PROF_BITS_COFF ".lprfb$M"
 #define INSTR_PROF_VALS_COFF ".lprfv$M"
 #define INSTR_PROF_VNODES_COFF ".lprfnd$M"
diff --git a/llvm/include/llvm/ProfileData/InstrProfReader.h b/llvm/include/llvm/ProfileData/InstrProfReader.h
index 134195059f9e6..4d6351f6425ef 100644
--- a/llvm/include/llvm/ProfileData/InstrProfReader.h
+++ b/llvm/include/llvm/ProfileData/InstrProfReader.h
@@ -349,6 +349,7 @@ class RawInstrProfReader : public InstrProfReader {
   uint64_t Version;
   uint64_t CountersDelta;
   uint64_t BitmapDelta;
+  uint64_t UniformCountersDelta;
   uint64_t NamesDelta;
   const RawInstrProf::ProfileData<IntPtrT> *Data;
   const RawInstrProf::ProfileData<IntPtrT> *DataEnd;
@@ -358,6 +359,8 @@ class RawInstrProfReader : public InstrProfReader {
   const char *CountersEnd;
   const char *BitmapStart;
   const char *BitmapEnd;
+  const char *UniformCountersStart;
+  const char *UniformCountersEnd;
   const char *NamesStart;
   const char *NamesEnd;
   const char *VNamesStart = nullptr;
@@ -469,6 +472,7 @@ class RawInstrProfReader : public InstrProfReader {
   Error readFuncHash(NamedInstrProfRecord &Record);
   Error readRawCounts(InstrProfRecord &Record);
   Error readRawBitmapBytes(InstrProfRecord &Record);
+  Error readRawUniformCounters(InstrProfRecord &Record);
   Error readValueProfilingData(InstrProfRecord &Record);
   bool atEnd() const { return Data == DataEnd; }
 
@@ -482,6 +486,7 @@ class RawInstrProfReader : public InstrProfReader {
       // with respect to the next record.
       CountersDelta -= sizeof(*Data);
       BitmapDelta -= sizeof(*Data);
+      UniformCountersDelta -= sizeof(*Data);
     }
     Data++;
     ValueDataStart += CurValueDataSize;
diff --git a/llvm/include/llvm/ProfileData/InstrProfWriter.h b/llvm/include/llvm/ProfileData/InstrProfWriter.h
index 1b24425e68a9e..93212c19769e0 100644
--- a/llvm/include/llvm/ProfileData/InstrProfWriter.h
+++ b/llvm/include/llvm/ProfileData/InstrProfWriter.h
@@ -215,6 +215,7 @@ class InstrProfWriter {
     MemProfVersionRequested = Version;
   }
   void setMemProfFullSchema(bool Full) { MemProfFullSchema = Full; }
+
   // Compute the overlap b/w this object and Other. Program level result is
   // stored in Overlap and function level result is stored in FuncLevelOverlap.
   LLVM_ABI void overlapRecord(NamedInstrProfRecord &&Other,
diff --git a/llvm/lib/ProfileData/InstrProf.cpp b/llvm/lib/ProfileData/InstrProf.cpp
index d9ae1a6437bb1..a9ef7f64da95f 100644
--- a/llvm/lib/ProfileData/InstrProf.cpp
+++ b/llvm/lib/ProfileData/InstrProf.cpp
@@ -950,7 +950,8 @@ void InstrProfRecord::mergeValueProfData(
 }
 
 void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight,
-                            function_ref<void(instrprof_error)> Warn) {
+                            function_ref<void(instrprof_error)> Warn,
+                            unsigned WaveSize) {
   // If the number of counters doesn't match we either have bad data
   // or a hash collision.
   if (Counts.size() != Other.Counts.size()) {
@@ -975,7 +976,7 @@ void InstrProfRecord::merge(InstrProfRecord &Other, uint64_t Weight,
       setPseudoCount(PseudoWarm);
     return;
   }
-
+  OffloadDeviceWaveSize = Other.OffloadDeviceWaveSize;
   for (size_t I = 0, E = Other.Counts.size(); I < E; ++I) {
     bool Overflowed;
     uint64_t Value =
@@ -1715,7 +1716,7 @@ Expected<Header> Header::readFromBuffer(const unsigned char *Buffer) {
       IndexedInstrProf::ProfVersion::CurrentVersion)
     return make_error<InstrProfError>(instrprof_error::unsupported_version);
 
-  static_assert(IndexedInstrProf::ProfVersion::CurrentVersion == Version13,
+  static_assert(IndexedInstrProf::ProfVersion::CurrentVersion == Version14,
                 "Please update the reader as needed when a new field is added "
                 "or when indexed profile version gets bumped.");
 
@@ -1748,10 +1749,11 @@ size_t Header::size() const {
     // of the header, and byte offset of existing fields shouldn't change when
     // indexed profile version gets incremented.
     static_assert(
-        IndexedInstrProf::ProfVersion::CurrentVersion == Version13,
+        IndexedInstrProf::ProfVersion::CurrentVersion == Version14,
         "Please update the size computation below if a new field has "
         "been added to the header; for a version bump without new "
         "fields, add a case statement to fall through to the latest version.");
+  case 14ull: // UniformityBits added in record data, no header change
   case 13ull:
   case 12ull:
     return 72;
diff --git a/llvm/lib/ProfileData/InstrProfCorrelator.cpp b/llvm/lib/ProfileData/InstrProfCorrelator.cpp
index c571c338c8d32..c732382eb399a 100644
--- a/llvm/lib/ProfileData/InstrProfCorrelator.cpp
+++ b/llvm/lib/ProfileData/InstrProfCorrelator.cpp
@@ -311,6 +311,7 @@ void InstrProfCorrelatorImpl<IntPtrT>::addDataProbe(uint64_t NameRef,
       // In this mode, CounterPtr actually stores the section relative address
       // of the counter.
       maybeSwap<IntPtrT>(CounterOffset),
+      /*UniformCounterPtr=*/maybeSwap<IntPtrT>(0),
       // TODO: MC/DC is not yet supported.
       /*BitmapOffset=*/maybeSwap<IntPtrT>(0),
       maybeSwap<IntPtrT>(FunctionPtr),
@@ -318,6 +319,7 @@ void InstrProfCorrelatorImpl<IntPtrT>::addDataProbe(uint64_t NameRef,
       /*ValuesPtr=*/maybeSwap<IntPtrT>(0),
       maybeSwap<uint32_t>(NumCounters),
       /*NumValueSites=*/{maybeSwap<uint16_t>(0), maybeSwap<uint16_t>(0)},
+      /*OffloadDeviceWaveSize=*/maybeSwap<uint16_t>(0),
       // TODO: MC/DC is not yet supported.
       /*NumBitmapBytes=*/maybeSwap<uint32_t>(0),
   });
diff --git a/llvm/lib/ProfileData/InstrProfReader.cpp b/llvm/lib/ProfileData/InstrProfReader.cpp
index 8147ee8d0e816..b0bf80c0a7a6e 100644
--- a/llvm/lib/ProfileData/InstrProfReader.cpp
+++ b/llvm/lib/ProfileData/InstrProfReader.cpp
@@ -607,6 +607,7 @@ Error RawInstrProfReader<IntPtrT>::readHeader(
 
   CountersDelta = swap(Header.CountersDelta);
   BitmapDelta = swap(Header.BitmapDelta);
+  UniformCountersDelta = swap(Header.UniformCountersDelta);
   NamesDelta = swap(Header.NamesDelta);
   auto NumData = swap(Header.NumData);
   auto PaddingBytesBeforeCounters = swap(Header.PaddingBytesBeforeCounters);
@@ -614,6 +615,9 @@ Error RawInstrProfReader<IntPtrT>::readHeader(
   auto PaddingBytesAfterCounters = swap(Header.PaddingBytesAfterCounters);
   auto NumBitmapBytes = swap(Header.NumBitmapBytes);
   auto PaddingBytesAfterBitmapBytes = swap(Header.PaddingBytesAfterBitmapBytes);
+  auto NumUniformCounters = swap(Header.NumUniformCounters);
+  auto PaddingBytesAfterUniformCounters =
+      swap(Header.PaddingBytesAfterUniformCounters);
   auto NamesSize = swap(Header.NamesSize);
   auto VTableNameSize = swap(Header.VNamesSize);
   auto NumVTables = swap(Header.NumVTables);
@@ -626,14 +630,17 @@ Error RawInstrProfReader<IntPtrT>::readHeader(
   auto VTableSectionSize =
       NumVTables * sizeof(RawInstrProf::VTableProfileData<IntPtrT>);
   auto PaddingBytesAfterVTableProfData = getNumPaddingBytes(VTableSectionSize);
+  auto UniformCountersSectionSize = NumUniformCounters * sizeof(uint64_t);
 
   // Profile data starts after profile header and binary ids if exist.
   ptrdiff_t DataOffset = sizeof(RawInstrProf::Header) + BinaryIdSize;
   ptrdiff_t CountersOffset = DataOffset + DataSize + PaddingBytesBeforeCounters;
   ptrdiff_t BitmapOffset =
       CountersOffset + CountersSize + PaddingBytesAfterCounters;
-  ptrdiff_t NamesOffset =
+  ptrdiff_t UniformCountersOffset =
       BitmapOffset + NumBitmapBytes + PaddingBytesAfterBitmapBytes;
+  ptrdiff_t NamesOffset = UniformCountersOffset + UniformCountersSectionSize +
+                          PaddingBytesAfterUniformCounters;
   ptrdiff_t VTableProfDataOffset =
       NamesOffset + NamesSize + PaddingBytesAfterNames;
   ptrdiff_t VTableNameOffset = VTableProfDataOffset + VTableSectionSize +
@@ -694,6 +701,8 @@ Error RawInstrProfReader<IntPtrT>::readHeader(
   CountersEnd = CountersStart + CountersSize;
   BitmapStart = Start + BitmapOffset;
   BitmapEnd = BitmapStart + NumBitmapBytes;
+  UniformCountersStart = Start + UniformCountersOffset;
+  UniformCountersEnd = UniformCountersStart + UniformCountersSectionSize;
   ValueDataStart = reinterpret_cast<const uint8_t *>(Start + ValueDataOffset);
 
   std::unique_ptr<InstrProfSymtab> NewSymtab = std::make_unique<InstrProfSymtab>();
@@ -827,6 +836,52 @@ Error RawInstrProfReader<IntPtrT>::readRawBitmapBytes(InstrProfRecord &Record) {
   return success();
 }
 
+template <class IntPtrT>
+Error RawInstrProfReader<IntPtrT>::readRawUniformCounters(
+    InstrProfRecord &Record) {
+  Record.UniformCounts.clear();
+
+  if (UniformCountersStart == UniformCountersEnd)
+    return success();
+
+  uint32_t NumCounters = swap(Data->NumCounters);
+
+  ptrdiff_t UniformCounterOffset =
+      swap(Data->UniformCounterPtr) - UniformCountersDelta;
+  if (UniformCounterOffset < 0)
+    return error(instrprof_error::malformed,
+                 ("uniform counter offset " + Twine(UniformCounterOffset) +
+                  " is negative")
+                     .str());
+
+  if (UniformCounterOffset >= UniformCountersEnd - UniformCountersStart)
+    return error(instrprof_error::malformed,
+                 ("uniform counter offset " + Twine(UniformCounterOffset) +
+                  " is greater than the maximum uniform counter offset " +
+                  Twine(UniformCountersEnd - UniformCountersStart - 1))
+                     .str());
+
+  uint64_t MaxNumCounters =
+      (UniformCountersEnd - (UniformCountersStart + UniformCounterOffset)) /
+      sizeof(uint64_t);
+  if (NumCounters > MaxNumCounters)
+    return error(instrprof_error::malformed,
+                 ("number of uniform counters " + Twine(NumCounters) +
+                  " is greater than the maximum number of uniform counters " +
+                  Twine(MaxNumCounters))
+                     .str());
+
+  Record.UniformCounts.reserve(NumCounters);
+  for (uint32_t I = 0; I < NumCounters; I++) {
+    const char *Ptr =
+        UniformCountersStart + UniformCounterOffset + I * sizeof(uint64_t);
+    uint64_t CounterValue = swap(*reinterpret_cast<const uint64_t *>(Ptr));
+    Record.UniformCounts.push_back(CounterValue);
+  }
+
+  return success();
+}
+
 template <class IntPtrT>
 Error RawInstrProfReader<IntPtrT>::readValueProfilingData(
     InstrProfRecord &Record) {
@@ -873,6 +928,8 @@ Error RawInstrProfReader<IntPtrT>::readNextRecord(NamedInstrProfRecord &Record)
   if (Error E = readFuncHash(Record))
     return error(std::move(E));
 
+  Record.OffloadDeviceWaveSize = swap(Data->OffloadDeviceWaveSize);
+
   // Read raw counts and set Record.
   if (Error E = readRawCounts(Record))
     return error(std::move(E));
@@ -881,6 +938,10 @@ Error RawInstrProfReader<IntPtrT>::readNextRecord(NamedInstrProfRecord &Record)
   if (Error E = readRawBitmapBytes(Record))
     return error(std::move(E));
 
+  // Read raw uniform counters and set Record.
+  if (Error E = readRawUniformCounters(Record))
+    return error(std::move(E));
+
   // Read value data and set Record.
   if (Error E = readValueProfilingData(Record))
     return error(std::move(E));
@@ -945,11 +1006,12 @@ data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
   DataBuffer.clear();
   std::vector<uint64_t> CounterBuffer;
   std::vector<uint8_t> BitmapByteBuffer;
+  std::vector<uint8_t> UniformityBitsBuffer;
 
   const unsigned char *End = D + N;
   while (D < End) {
     // Read hash.
-    if (D + sizeof(uint64_t) >= End)
+    if (D + sizeof(uint64_t) > End)
       return data_type();
     uint64_t Hash = endian::readNext<uint64_t, llvm::endianness::little>(D);
 
@@ -977,18 +1039,51 @@ data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
       if (D + sizeof(uint64_t) > End)
         return data_type();
       BitmapBytes = endian::readNext<uint64_t, llvm::endianness::little>(D);
-      // Read bitmap byte values.
-      if (D + BitmapBytes * sizeof(uint8_t) > End)
-        return data_type();
       BitmapByteBuffer.clear();
       BitmapByteBuffer.reserve(BitmapBytes);
-      for (uint64_t J = 0; J < BitmapBytes; ++J)
-        BitmapByteBuffer.push_back(static_cast<uint8_t>(
-            endian::readNext<uint64_t, llvm::endianness::little>(D)));
+
+      if (GET_VERSION(FormatVersion) >=
+          IndexedInstrProf::ProfVersion::Version14) {
+        // Version 14+: bitmap bytes stored as uint8_t with padding.
+        uint64_t PaddedSize = alignTo(BitmapBytes, sizeof(uint64_t));
+        if (D + PaddedSize > End)
+          return data_type();
+        for (uint64_t J = 0; J < BitmapBytes; ++J)
+          BitmapByteBuffer.push_back(
+              endian::readNext<uint8_t, llvm::endianness::little>(D));
+        for (uint64_t J = BitmapBytes; J < PaddedSize; ++J)
+          (void)endian::readNext<uint8_t, llvm::endianness::little>(D);
+
+        // Read uniformity bits (AMDGPU offload profiling).
+        uint64_t UniformityBitsSize = 0;
+        if (D + sizeof(uint64_t) > End)
+          return data_type();
+        UniformityBitsSize =
+            endian::readNext<uint64_t, llvm::endianness::little>(D);
+        uint64_t PaddedUniformitySize =
+            alignTo(UniformityBitsSize, sizeof(uint64_t));
+        if (D + PaddedUniformitySize > End)
+          return data_type();
+        UniformityBitsBuffer.clear();
+        UniformityBitsBuffer.reserve(UniformityBitsSize);
+        for (uint64_t J = 0; J < UniformityBitsSize; ++J)
+          UniformityBitsBuffer.push_back(
+              endian::readNext<uint8_t, llvm::endianness::little>(D));
+        for (uint64_t J = UniformityBitsSize; J < PaddedUniformitySize; ++J)
+          (void)endian::readNext<uint8_t, llvm::endianness::little>(D);
+      } else {
+        // Version 11-13: each bitmap byte stored as a uint64_t.
+        if (D + BitmapBytes * sizeof(uint64_t) > End)
+          return data_type();
+        for (uint64_t J = 0; J < BitmapBytes; ++J)
+          BitmapByteBuffer.push_back(static_cast<uint8_t>(
+              endian::readNext<uint64_t, llvm::endianness::little>(D)));
+      }
     }
 
     DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer),
-                            std::move(BitmapByteBuffer));
+                            std::move(BitmapByteBuffer),
+                            std::move(UniformityBitsBuffer));
 
     // Read value profiling data.
     if (GET_VERSION(FormatVersion) > IndexedInstrProf::ProfVersion::Version2 &&
diff --git a/llvm/lib/ProfileData/InstrProfWriter.cpp b/llvm/lib/ProfileData/InstrProfWriter.cpp
index 0f15ca8ff6df7..96611612184e7 100644
--- a/llvm/lib/ProfileData/InstrProfWriter.cpp
+++ b/llvm/lib/ProfileData/InstrProfWriter.cpp
@@ -51,6 +51,7 @@ class InstrProfRecordWriterTrait {
   llvm::endianness ValueProfDataEndianness = llvm::endianness::little;
   InstrProfSummaryBuilder *SummaryBuilder;
   InstrProfSummaryBuilder *CSSummaryBuilder;
+  bool WritePrevVersion = false;
 
   InstrProfRecordWriterTrait() = default;
 
@@ -58,7 +59,7 @@ class InstrProfRecordWriterTrait {
     return IndexedInstrProf::ComputeHash(K);
   }
 
-  static std::pair<offset_type, offset_type>
+  std::pair<offset_type, offset_type>
   EmitKeyDataLength(raw_ostream &Out, key_type_ref K, data_type_ref V) {
     using namespace support;
 
@@ -74,7 +75,16 @@ class InstrProfRecordWriterTrait {
       M += sizeof(uint64_t); // The size of the Counts vector
       M += ProfRecord.Counts.size() * sizeof(uint64_t);
       M += sizeof(uint64_t); // The size of the Bitmap vector
-      M += ProfRecord.BitmapBytes.size() * sizeof(uint64_t);
+      if (WritePrevVersion) {
+        // Version 13: each bitmap byte stored as a uint64_t.
+        M += ProfRecord.BitmapBytes.size() * sizeof(uint64_t);
+      } else {
+        // Version 14+: bitmap bytes as uint8_t with padding, plus
+        // uniformity bits.
+        M += alignTo(ProfRecord.BitmapBytes.size(), sizeof(uint64_t));
+        M += sizeof(uint64_t); // The size of the UniformityBits vector
+        M += alignTo(ProfRecord.UniformityBits.size(), sizeof(uint64_t));
+      }
 
       // Value data
       M += ValueProfData::getSize(ProfileData.second);
@@ -88,7 +98,8 @@ class InstrProfRecordWriterTrait {
     Out.write(K.data(), N);
   }
 
-  void EmitData(raw_ostream &Out, key_type_ref, data_type_ref V, offset_type) {
+  void EmitData(raw_ostream &Out, key_type_ref K, data_type_ref V,
+                offset_type) {
     using namespace support;
 
     endian::Writer LE(Out, llvm::endianness::little);
@@ -105,8 +116,27 @@ class InstrProfRecordWriterTrait {
         LE.write<uint64_t>(I);
 
       LE.write<uint64_t>(ProfRecord.BitmapBytes.size());
-      for (uint64_t I : ProfRecord.BitmapBytes)
-        LE.write<uint64_t>(I);
+      if (WritePrevVersion) {
+        // Version 13: each bitmap byte stored as a uint64_t.
+        for (uint8_t I : ProfRecord.BitmapBytes)
+          LE.write<uint64_t>(I);
+      } else {
+        // Version 14+: bitmap bytes as uint8_t with padding.
+        for (uint8_t I : ProfRecord.BitmapBytes)
+          LE.write<uint8_t>(I);
+        for (size_t I = ProfRecord.BitmapBytes.size();
+             I < alignTo(ProfRecord.BitmapBytes.size(), sizeof(uint64_t)); ++I)
+          LE.write<uint8_t>(0);
+
+        // Write uniformity bits (AMDGPU offload profiling).
+        LE.write<uint64_t>(ProfRecord.UniformityBits.size());
+        for (uint8_t I : ProfRecord.UniformityBits)
+          LE.write<uint8_t>(I);
+        for (size_t I = ProfRecord.UniformityBits.size();
+             I < alignTo(ProfRecord.UniformityBits.size(), sizeof(uint64_t));
+             ++I)
+          LE.write<uint8_t>(0);
+      }
 
       // Write value data
       std::unique_ptr<ValueProfData> VDataPtr =
@@ -209,7 +239,7 @@ void InstrProfWriter::addRecord(StringRef Name, uint64_t Hash,
       Dest.scale(Weight, 1, MapWarn);
   } else {
     // We're updating a function we've seen before.
-    Dest.merge(I, Weight, MapWarn);
+    Dest.merge(I, Weight, MapWarn, I.OffloadDeviceWaveSize);
   }
 
   Dest.sortValueData();
@@ -524,6 +554,7 @@ Error InstrProfWriter::writeImpl(ProfOStream &OS) {
   InfoObj->SummaryBuilder = &ISB;
   InstrProfSummaryBuilder CSISB(ProfileSummaryBuilder::DefaultCutoffs);
   InfoObj->CSSummaryBuilder = &CSISB;
+  InfoObj->WritePrevVersion = WritePrevVersion;
 
   // Populate the hash table generator.
   SmallVector<std::pair<StringRef, const ProfilingData *>> OrderedData;
@@ -542,7 +573,7 @@ Error InstrProfWriter::writeImpl(ProfOStream &OS) {
   // The WritePrevVersion handling will either need to be removed or updated
   // if the version is advanced beyond 12.
   static_assert(IndexedInstrProf::ProfVersion::CurrentVersion ==
-                IndexedInstrProf::ProfVersion::Version13);
+                IndexedInstrProf::ProfVersion::Version14);
   if (static_cast<bool>(ProfileKind & InstrProfKind::IRInstrumentation))
     Header.Version |= VARIANT_MASK_IR_PROF;
   if (static_cast<bool>(ProfileKind & InstrProfKind::ContextSensitive))
diff --git a/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp b/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp
index 8e4ba41919768..58aac27f95225 100644
--- a/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp
+++ b/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp
@@ -164,6 +164,14 @@ cl::opt<bool> SpeculativeCounterPromotionToLoop(
              " update can be further/iteratively promoted into an acyclic "
              " region."));
 
+static cl::opt<unsigned> OffloadPGOSampling(
+    "offload-pgo-sampling",
+    cl::desc("Log2 of the sampling period for offload PGO instrumentation. "
+             "Only 1 in every 2^N blocks is instrumented. "
+             "0 = all blocks, 1 = 50%, 2 = 25%, 3 = 12.5% (default). "
+             "Higher values reduce overhead at the cost of sparser profiles."),
+    cl::init(3));
+
 cl::opt<bool> IterativeCounterPromotion(
     "iterative-counter-promotion", cl::init(true),
     cl::desc("Allow counter promotion across the whole loop nest."));
@@ -269,6 +277,8 @@ class InstrLowerer final {
   struct PerFunctionProfileData {
     uint32_t NumValueSites[IPVK_Last + 1] = {};
     GlobalVariable *RegionCounters = nullptr;
+    GlobalVariable *UniformCounters =
+        nullptr; // Per-block uniform-entry counters
     GlobalVariable *DataVar = nullptr;
     GlobalVariable *RegionBitmaps = nullptr;
     uint32_t NumBitmapBytes = 0;
@@ -298,6 +308,18 @@ class InstrLowerer final {
 
   int64_t TotalCountersPromoted = 0;
 
+  // Per-function cache of invariant values for GPU PGO instrumentation.
+  // Computed once at the function entry and reused across all instrumentation
+  // points to avoid redundant IR and help the optimizer.
+  struct GPUPGOInvariants {
+    Value *Matched = nullptr;
+    bool WaveSizeStored = false;
+  };
+  DenseMap<Function *, GPUPGOInvariants> GPUInvariantsCache;
+
+  /// Emit invariant PGO values at the function entry block and cache them.
+  GPUPGOInvariants &getOrCreateGPUInvariants(Function *F);
+
   /// Lower instrumentation intrinsics in the function. Returns true if there
   /// any lowering.
   bool lowerIntrinsics(Function *F);
@@ -354,6 +376,10 @@ class InstrLowerer final {
   /// referring to them will also be created.
   GlobalVariable *getOrCreateRegionCounters(InstrProfCntrInstBase *Inc);
 
+  /// Get the uniform entry counters for GPU divergence tracking.
+  /// These counters track how often blocks are entered with all lanes active.
+  GlobalVariable *getOrCreateUniformCounters(InstrProfCntrInstBase *Inc);
+
   /// Create the region counters.
   GlobalVariable *createRegionCounters(InstrProfCntrInstBase *Inc,
                                        StringRef Name,
@@ -1197,26 +1223,125 @@ void InstrLowerer::lowerTimestamp(
   TimestampInstruction->eraseFromParent();
 }
 
+InstrLowerer::GPUPGOInvariants &
+InstrLowerer::getOrCreateGPUInvariants(Function *F) {
+  auto It = GPUInvariantsCache.find(F);
+  if (It != GPUInvariantsCache.end())
+    return It->second;
+
+  LLVMContext &Context = M.getContext();
+  auto *Int32Ty = Type::getInt32Ty(Context);
+
+  BasicBlock &EntryBB = F->getEntryBlock();
+  IRBuilder<> Builder(&*EntryBB.getFirstInsertionPt());
+
+  Value *Matched = ConstantInt::getTrue(Context);
+  if (OffloadPGOSampling > 0) {
+    FunctionCallee IsSampledFn =
+        M.getOrInsertFunction(RTLIB::RuntimeLibcallsInfo::getLibcallImplName(
+                                  RTLIB::impl___llvm_profile_sampling_gpu),
+                              Int32Ty, Int32Ty);
+    Value *SampledInt = Builder.CreateCall(
+        IsSampledFn, {ConstantInt::get(Int32Ty, OffloadPGOSampling)},
+        "pgo.sampled");
+    Matched = Builder.CreateICmpNE(SampledInt, ConstantInt::get(Int32Ty, 0),
+                                   "pgo.matched");
+  }
+
+  auto &Inv = GPUInvariantsCache[F];
+  Inv.Matched = Matched;
+  return Inv;
+}
+
 void InstrLowerer::lowerIncrement(InstrProfIncrementInst *Inc) {
-  auto *Addr = getCounterAddress(Inc);
   IRBuilder<> Builder(Inc);
   if (isGPUProfTarget(M)) {
-    auto *Int64Ty = Builder.getInt64Ty();
-    auto *PtrTy = Builder.getPtrTy();
-    auto *CalleeTy = FunctionType::get(Type::getVoidTy(M.getContext()),
+    Function *F = Inc->getFunction();
+    auto &Inv = getOrCreateGPUInvariants(F);
+
+    LLVMContext &Context = M.getContext();
+    auto *Int64Ty = Type::getInt64Ty(Context);
+    auto *PtrTy = PointerType::getUnqual(Context);
+
+    auto *Addr = getCounterAddress(Inc);
+
+    // Store the device wave/warp size into the profile data struct once per
+    // function. The size lets the host-side merge classify wave-uniform
+    // counters. AMDGPU folds llvm.amdgcn.wavefrontsize to the subtarget's
+    // constant; other GPUs use their fixed warp size.
+    if (!Inv.WaveSizeStored) {
+      Inv.WaveSizeStored = true;
+      GlobalVariable *NamePtr = Inc->getName();
+      auto &PD = ProfileDataMap[NamePtr];
+      if (PD.DataVar) {
+        IRBuilder<> EntryBuilder(&*F->getEntryBlock().getFirstInsertionPt());
+        Value *WaveSize16 = nullptr;
+        // Look the intrinsic up by name so this target-agnostic pass does not
+        // pull in IntrinsicsAMDGPU.h. AMDGPU folds the intrinsic to the
+        // subtarget's wavefront size; other GPUs fall back to a 32-lane warp.
+        if (TT.isAMDGPU()) {
+          Intrinsic::ID WaveSizeID =
+              Intrinsic::lookupIntrinsicID("llvm.amdgcn.wavefrontsize");
+          if (WaveSizeID != Intrinsic::not_intrinsic) {
+            Function *WaveSizeFn =
+                Intrinsic::getOrInsertDeclaration(&M, WaveSizeID);
+            Value *WaveSize = EntryBuilder.CreateCall(WaveSizeFn);
+            WaveSize16 = EntryBuilder.CreateTrunc(
+                WaveSize, Type::getInt16Ty(Context), "wavesize.i16");
+          }
+        }
+        if (!WaveSize16)
+          WaveSize16 = ConstantInt::get(Type::getInt16Ty(Context), 32);
+        Value *WaveSizeAddr = EntryBuilder.CreateStructGEP(
+            PD.DataVar->getValueType(), PD.DataVar, 9, "profd.wavesize");
+        EntryBuilder.CreateStore(WaveSize16, WaveSizeAddr);
+      }
+    }
+
+    GlobalVariable *UniformCounters = getOrCreateUniformCounters(Inc);
+    Value *UniformAddrArg = ConstantPointerNull::get(PtrTy);
+    if (UniformCounters) {
+      Value *UniformIndices[] = {Builder.getInt32(0), Inc->getIndex()};
+      Value *UniformAddr = Builder.CreateInBoundsGEP(
+          UniformCounters->getValueType(), UniformCounters, UniformIndices,
+          "unifctr.addr");
+      UniformAddrArg =
+          Builder.CreatePointerBitCastOrAddrSpaceCast(UniformAddr, PtrTy);
+    }
+    Value *CastAddr = Builder.CreatePointerBitCastOrAddrSpaceCast(Addr, PtrTy);
+    Value *StepI64 =
+        Builder.CreateZExtOrTrunc(Inc->getStep(), Int64Ty, "step.i64");
+
+    auto *CalleeTy = FunctionType::get(Type::getVoidTy(Context),
                                        {PtrTy, PtrTy, Int64Ty}, false);
     FunctionCallee Callee =
         M.getOrInsertFunction(RTLIB::RuntimeLibcallsInfo::getLibcallImplName(
                                   RTLIB::impl___llvm_profile_instrument_gpu),
                               CalleeTy);
-    Value *CastAddr = Builder.CreatePointerBitCastOrAddrSpaceCast(Addr, PtrTy);
-    Value *Uniform =
-        ConstantPointerNull::get(PointerType::getUnqual(M.getContext()));
-    Value *StepI64 =
-        Builder.CreateZExtOrTrunc(Inc->getStep(), Int64Ty, "step.i64");
-    Builder.CreateCall(Callee, {CastAddr, Uniform, StepI64});
-  } else if (Options.Atomic || AtomicCounterUpdateAll ||
-             (Inc->getIndex()->isNullValue() && AtomicFirstCounter)) {
+
+    if (OffloadPGOSampling > 0) {
+      BasicBlock *CurBB = Builder.GetInsertBlock();
+      BasicBlock *ContBB =
+          CurBB->splitBasicBlock(BasicBlock::iterator(Inc), "po_cont");
+      BasicBlock *ThenBB = BasicBlock::Create(Context, "po_then", F);
+
+      CurBB->getTerminator()->eraseFromParent();
+      IRBuilder<> HeadBuilder(CurBB);
+      HeadBuilder.CreateCondBr(Inv.Matched, ThenBB, ContBB);
+
+      IRBuilder<> ThenBuilder(ThenBB);
+      ThenBuilder.CreateCall(Callee, {CastAddr, UniformAddrArg, StepI64});
+      ThenBuilder.CreateBr(ContBB);
+    } else {
+      Builder.CreateCall(Callee, {CastAddr, UniformAddrArg, StepI64});
+    }
+    Inc->eraseFromParent();
+    return;
+  }
+
+  auto *Addr = getCounterAddress(Inc);
+  if (Options.Atomic || AtomicCounterUpdateAll ||
+      (Inc->getIndex()->isNullValue() && AtomicFirstCounter)) {
     Builder.CreateAtomicRMW(AtomicRMWInst::Add, Addr, Inc->getStep(),
                             MaybeAlign(), AtomicOrdering::Monotonic);
   } else {
@@ -1753,12 +1878,54 @@ InstrLowerer::getOrCreateRegionCounters(InstrProfCntrInstBase *Inc) {
     CompilerUsedVars.push_back(PD.RegionCounters);
   }
 
+  // Create uniform counters before the data variable so that
+  // UniformCounterPtr can reference them in createDataVariable().
+  getOrCreateUniformCounters(Inc);
+
   // Create the data variable (if it doesn't already exist).
   createDataVariable(Inc);
 
   return PD.RegionCounters;
 }
 
+GlobalVariable *
+InstrLowerer::getOrCreateUniformCounters(InstrProfCntrInstBase *Inc) {
+  // Uniform counters are only meaningful for GPU profile targets.
+  if (!isGPUProfTarget(M))
+    return nullptr;
+
+  GlobalVariable *NamePtr = Inc->getName();
+  auto &PD = ProfileDataMap[NamePtr];
+  if (PD.UniformCounters)
+    return PD.UniformCounters;
+
+  // Ensure RegionCounters exists first (we need the same size)
+  getOrCreateRegionCounters(Inc);
+
+  uint64_t NumCounters = Inc->getNumCounters()->getZExtValue();
+
+  LLVMContext &Ctx = M.getContext();
+  ArrayType *CounterTy = ArrayType::get(Type::getInt64Ty(Ctx), NumCounters);
+
+  bool Renamed;
+  std::string VarName = getVarName(Inc, "__llvm_prf_unifcnt_", Renamed);
+
+  auto *GV = new GlobalVariable(M, CounterTy, false, NamePtr->getLinkage(),
+                                Constant::getNullValue(CounterTy), VarName);
+  GV->setAlignment(Align(8));
+
+  GV->setSection(getInstrProfSectionName(IPSK_ucnts, TT.getObjectFormat()));
+
+  GV->setComdat(M.getOrInsertComdat(VarName));
+  GV->setLinkage(GlobalValue::LinkOnceODRLinkage);
+  GV->setVisibility(GlobalValue::ProtectedVisibility);
+
+  PD.UniformCounters = GV;
+  CompilerUsedVars.push_back(GV);
+
+  return PD.UniformCounters;
+}
+
 void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
   // When debug information is correlated to profile data, a data variable
   // is not needed.
@@ -1823,6 +1990,7 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
   uint64_t NumCounters = Inc->getNumCounters()->getZExtValue();
 
   Constant *CounterPtr = PD.RegionCounters;
+  Constant *UniformCounterPtr = PD.UniformCounters;
 
   uint64_t NumBitmapBytes = PD.NumBitmapBytes;
 
@@ -1838,6 +2006,8 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
     Int16ArrayVals[Kind] = ConstantInt::get(Int16Ty, PD.NumValueSites[Kind]);
 
+  uint16_t OffloadDeviceWaveSizeVal = 0;
+
   if (isGPUProfTarget(M)) {
     // For GPU targets, weak functions need weak linkage for their profile data
     // aliases to allow linker deduplication across TUs
@@ -1875,6 +2045,7 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
       new GlobalVariable(M, DataTy, false, Linkage, nullptr, DataVarName);
 
   Constant *RelativeCounterPtr;
+  Constant *RelativeUniformCounterPtr = ConstantInt::get(IntPtrTy, 0);
   GlobalVariable *BitmapPtr = PD.RegionBitmaps;
   Constant *RelativeBitmapPtr = ConstantInt::get(IntPtrTy, 0);
   InstrProfSectKind DataSectionKind;
@@ -1885,6 +2056,9 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
     RelativeCounterPtr = ConstantExpr::getPtrToInt(CounterPtr, IntPtrTy);
     if (BitmapPtr != nullptr)
       RelativeBitmapPtr = ConstantExpr::getPtrToInt(BitmapPtr, IntPtrTy);
+    if (UniformCounterPtr != nullptr)
+      RelativeUniformCounterPtr =
+          ConstantExpr::getPtrToInt(UniformCounterPtr, IntPtrTy);
   } else if (TT.isNVPTX()) {
     // The NVPTX target cannot handle self-referencing constant expressions in
     // global initializers at all. Use absolute pointers and have the runtime
@@ -1902,6 +2076,10 @@ void InstrLowerer::createDataVariable(InstrProfCntrInstBase *Inc) {
       RelativeBitmapPtr =
           ConstantExpr::getSub(ConstantExpr::getPtrToInt(BitmapPtr, IntPtrTy),
                                ConstantExpr::getPtrToInt(Data, IntPtrTy));
+    if (UniformCounterPtr != nullptr)
+      RelativeUniformCounterPtr = ConstantExpr::getSub(
+          ConstantExpr::getPtrToInt(UniformCounterPtr, IntPtrTy),
+          ConstantExpr::getPtrToInt(Data, IntPtrTy));
   }
 
   Constant *DataVals[] = {
diff --git a/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp b/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
index b6d07aa821e7f..b30121a87d483 100644
--- a/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
+++ b/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
@@ -118,6 +118,7 @@
 #include <algorithm>
 #include <cassert>
 #include <cstdint>
+#include <cstdlib>
 #include <memory>
 #include <numeric>
 #include <optional>
@@ -379,9 +380,15 @@ class FunctionInstrumenter final {
   // values. Supporting other values is relatively straight-forward - just
   // another counter range within the context.
   bool isValueProfilingDisabled() const {
+    // Value profiling is disabled for GPU targets because the device-side
+    // profiling runtime does not yet implement
+    // __llvm_profile_instrument_target. The existing compiler-rt implementation
+    // uses a linked-list with locks and eviction policy that is not efficient
+    // for massively parallel GPU execution. A GPU-optimized implementation is
+    // left as future work.
     return DisableValueProfiling ||
            InstrumentationType == PGOInstrumentationType::CTXPROF ||
-           M.getTargetTriple().isGPU();
+           isGPUProfTarget(M);
   }
 
   bool shouldInstrumentEntryBB() const {
@@ -1199,6 +1206,9 @@ class PGOUseFunc {
   // Annotate the irreducible loop header weights.
   void annotateIrrLoopHeaderWeights();
 
+  // Annotate per-block uniformity info for offload profiling.
+  void setBlockUniformityAttribute();
+
   // The hotness of the function from the profile count.
   enum FuncFreqAttr { FFA_Normal, FFA_Cold, FFA_Hot };
 
@@ -1305,11 +1315,25 @@ bool PGOUseFunc::setInstrumentedCounts(
 
   setupBBInfoEdges(FuncInfo);
 
-  unsigned NumCounters =
-      InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
+  unsigned NumInstrumentedBBs = InstrumentBBs.size();
+  unsigned NumSelects = FuncInfo.SIVisitor.getNumOfSelectInsts();
+  unsigned NumCounters = NumInstrumentedBBs + NumSelects;
   // The number of counters here should match the number of counters
   // in profile. Return if they mismatch.
   if (NumCounters != CountFromProfile.size()) {
+    LLVM_DEBUG({
+      dbgs() << "PGO COUNTER MISMATCH for function " << F.getName() << ":\n";
+      dbgs() << "  Expected counters: " << NumCounters << "\n";
+      dbgs() << "    - From instrumented edges: " << NumInstrumentedBBs << "\n";
+      for (size_t i = 0; i < InstrumentBBs.size(); ++i) {
+        dbgs() << "      " << i << ": ";
+        InstrumentBBs[i]->printAsOperand(dbgs(), false);
+        dbgs() << "\n";
+      }
+      dbgs() << "    - From select instructions: " << NumSelects << "\n";
+      dbgs() << "  Actual counters from profile: " << CountFromProfile.size()
+             << "\n";
+    });
     return false;
   }
   auto *FuncEntry = &*F.begin();
@@ -1317,6 +1341,7 @@ bool PGOUseFunc::setInstrumentedCounts(
   // Set the profile count to the Instrumented BBs.
   uint32_t I = 0;
   for (BasicBlock *InstrBB : InstrumentBBs) {
+
     uint64_t CountValue = CountFromProfile[I++];
     PGOUseBBInfo &Info = getBBInfo(InstrBB);
     // If we reach here, we know that we have some nonzero count
@@ -1762,6 +1787,42 @@ void PGOUseFunc::annotateIrrLoopHeaderWeights() {
   }
 }
 
+void PGOUseFunc::setBlockUniformityAttribute() {
+  if (ProfileRecord.UniformityBits.empty())
+    return;
+
+  // Annotate uniformity on each instrumented IR basic block so later codegen
+  // passes (MachineFunction) can consume it without relying on fragile block
+  // numbering heuristics.
+  //
+  // Metadata kind: LLVMContext::MD_block_uniformity_profile
+  // Payload: i1 (true = uniform, false = divergent)
+
+  std::vector<BasicBlock *> InstrumentBBs;
+  FuncInfo.getInstrumentBBs(InstrumentBBs);
+
+  LLVMContext &Ctx = F.getContext();
+  Type *Int1Ty = Type::getInt1Ty(Ctx);
+
+  for (size_t I = 0, E = InstrumentBBs.size(); I < E; ++I) {
+    BasicBlock *BB = InstrumentBBs[I];
+    if (!BB || !BB->getTerminator())
+      continue;
+    bool IsUniform = ProfileRecord.isBlockUniform(I);
+    auto *MD = MDNode::get(
+        Ctx, ConstantAsMetadata::get(ConstantInt::get(Int1Ty, IsUniform)));
+    BB->getTerminator()->setMetadata(LLVMContext::MD_block_uniformity_profile,
+                                     MD);
+  }
+
+  LLVM_DEBUG({
+    dbgs() << "PGO: Set block uniformity profile for " << F.getName() << ": ";
+    for (size_t I = 0, E = InstrumentBBs.size(); I < E; ++I)
+      dbgs() << (ProfileRecord.isBlockUniform(I) ? 'U' : 'D');
+    dbgs() << "\n";
+  });
+}
+
 void SelectInstVisitor::instrumentOneSelectInst(SelectInst &SI) {
   Module *M = F.getParent();
   IRBuilder<> Builder(&SI);
@@ -2273,6 +2334,7 @@ static bool annotateAllFunctions(
     Func.setBranchWeights();
     Func.annotateValueSites();
     Func.annotateIrrLoopHeaderWeights();
+    Func.setBlockUniformityAttribute();
     PGOUseFunc::FuncFreqAttr FreqAttr = Func.getFuncFreqAttr();
     if (FreqAttr == PGOUseFunc::FFA_Cold)
       ColdFunctions.push_back(&F);
@@ -2410,11 +2472,12 @@ void llvm::setProfMetadata(Instruction *TI, ArrayRef<uint64_t> EdgeCounts,
   LLVM_DEBUG(dbgs() << "Weight is: "; for (const auto &W
                                            : Weights) {
     dbgs() << W << " ";
-  } dbgs() << "\n";);
+  } dbgs() << "\n");
 
   misexpect::checkExpectAnnotations(*TI, Weights, /*IsFrontend=*/false);
 
   setBranchWeights(*TI, Weights, /*IsExpected=*/false);
+
   if (EmitBranchProbability) {
     std::string BrCondStr = getBranchCondString(TI);
     if (BrCondStr.empty())
diff --git a/llvm/test/CodeGen/AArch64/global-merge-profile-sections.ll b/llvm/test/CodeGen/AArch64/global-merge-profile-sections.ll
index 7108d82ddaa2e..37e5295c91aee 100644
--- a/llvm/test/CodeGen/AArch64/global-merge-profile-sections.ll
+++ b/llvm/test/CodeGen/AArch64/global-merge-profile-sections.ll
@@ -5,7 +5,7 @@ $__profc_begin = comdat nodeduplicate
 $__profc_end = comdat nodeduplicate
 
 @__profc_begin = private global [2 x i64] zeroinitializer, section "__llvm_prf_cnts", comdat, align 8
- at __profd_begin = private global { i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i32 } { i64 -1301828029439649651, i64 172590168, i64 sub (i64 ptrtoint (ptr @__profc_begin to i64), i64 ptrtoint (ptr @__profd_begin to i64)), i64 0, ptr null, ptr null, i32 2, [3 x i16] zeroinitializer, i32 0 }, section "__llvm_prf_data", comdat($__profc_begin), align 8
+ at __profd_begin = private global { i64, i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i16, i32 } { i64 -1301828029439649651, i64 172590168, i64 sub (i64 ptrtoint (ptr @__profc_begin to i64), i64 ptrtoint (ptr @__profd_begin to i64)), i64 0, i64 0, ptr null, ptr null, i32 2, [3 x i16] zeroinitializer, i16 0, i32 0 }, section "__llvm_prf_data", comdat($__profc_begin), align 8
 @__profc_end = private global [2 x i64] zeroinitializer, section "__llvm_prf_cnts", comdat, align 8
- at __profd_end = private global { i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i32 } { i64 3274037854792712831, i64 172590168, i64 sub (i64 ptrtoint (ptr @__profc_end to i64), i64 ptrtoint (ptr @__profd_end to i64)), i64 0, ptr null, ptr null, i32 2, [3 x i16] zeroinitializer, i32 0 }, section "__llvm_prf_data", comdat($__profc_end), align 8
+ at __profd_end = private global { i64, i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i16, i32 } { i64 3274037854792712831, i64 172590168, i64 sub (i64 ptrtoint (ptr @__profc_end to i64), i64 ptrtoint (ptr @__profd_end to i64)), i64 0, i64 0, ptr null, ptr null, i32 2, [3 x i16] zeroinitializer, i16 0, i32 0 }, section "__llvm_prf_data", comdat($__profc_end), align 8
 
diff --git a/llvm/test/Instrumentation/InstrProfiling/amdgpu-3d-grid.ll b/llvm/test/Instrumentation/InstrProfiling/amdgpu-3d-grid.ll
new file mode 100644
index 0000000000000..f775b9756e878
--- /dev/null
+++ b/llvm/test/Instrumentation/InstrProfiling/amdgpu-3d-grid.ll
@@ -0,0 +1,23 @@
+;; Test that AMDGPU PGO instrumentation generates contiguous counter arrays
+;; and profile section symbols with CUID-based naming. The __llvm_profile_sampling_gpu
+;; library function handles 3D block linearization internally.
+
+; RUN: opt %s -mtriple=amdgcn-amd-amdhsa -passes=instrprof -S | FileCheck %s
+
+ at __hip_cuid_abcdef789 = addrspace(1) global i8 0
+ at __profn_kernel_3d = private constant [9 x i8] c"kernel_3d"
+
+define amdgpu_kernel void @kernel_3d() {
+  call void @llvm.instrprof.increment(ptr @__profn_kernel_3d, i64 12345, i32 1, i32 0)
+  ret void
+}
+
+declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
+
+;; Per-function comdat counters (3D grid linearization is handled in the runtime library)
+; CHECK: @__profc_kernel_3d = linkonce_odr protected addrspace(1) global [1 x i64]
+; CHECK: @__llvm_prf_unifcnt_kernel_3d = linkonce_odr protected addrspace(1) global [1 x i64]
+
+;; Check sampling guard calls library function
+; CHECK: call i32 @__llvm_profile_sampling_gpu(i32 3)
+; CHECK: call void @__llvm_profile_instrument_gpu(ptr addrspacecast (ptr addrspace(1) @__profc_kernel_3d to ptr), ptr addrspacecast (ptr addrspace(1) @__llvm_prf_unifcnt_kernel_3d to ptr), i64 1)
diff --git a/llvm/test/Instrumentation/InstrProfiling/amdgpu-profc-arrays.ll b/llvm/test/Instrumentation/InstrProfiling/amdgpu-contiguous-counters.ll
similarity index 52%
rename from llvm/test/Instrumentation/InstrProfiling/amdgpu-profc-arrays.ll
rename to llvm/test/Instrumentation/InstrProfiling/amdgpu-contiguous-counters.ll
index eab78fb3591b1..8d48372a12de6 100644
--- a/llvm/test/Instrumentation/InstrProfiling/amdgpu-profc-arrays.ll
+++ b/llvm/test/Instrumentation/InstrProfiling/amdgpu-contiguous-counters.ll
@@ -1,13 +1,17 @@
 ;; Per-kernel __profc_* arrays land in section __llvm_prf_cnts with one slot
-;; per counter, and counter increments lower to __llvm_profile_instrument_gpu
-;; calls whose pointer argument is a GEP into the per-kernel array.
+;; per counter, and a parallel __llvm_prf_unifcnt_* array lands in
+;; __llvm_prf_ucnts. Counter increments lower to __llvm_profile_instrument_gpu
+;; calls whose first pointer argument is a GEP into the per-kernel counter
+;; array and whose second is a GEP into the matching uniform-counter array.
+;; Sampling is disabled so the increment lowers to a direct call.
 
-; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=instrprof < %s | FileCheck %s
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -offload-pgo-sampling=0 -passes=instrprof < %s | FileCheck %s
 
 @__profn_kernel1 = private constant [7 x i8] c"kernel1"
 @__profn_kernel2 = private constant [7 x i8] c"kernel2"
 
 ; CHECK: @__profc_kernel1 = linkonce_odr protected addrspace(1) global [2 x i64] zeroinitializer, section "__llvm_prf_cnts"
+; CHECK: @__llvm_prf_unifcnt_kernel1 = linkonce_odr protected addrspace(1) global [2 x i64] zeroinitializer, section "__llvm_prf_ucnts"
 ; CHECK: @__profc_kernel2 = linkonce_odr protected addrspace(1) global [1 x i64] zeroinitializer, section "__llvm_prf_cnts"
 
 define amdgpu_kernel void @kernel1() {
@@ -23,4 +27,6 @@ define amdgpu_kernel void @kernel2() {
 
 declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
 
-; CHECK: call void @__llvm_profile_instrument_gpu(ptr addrspacecast (ptr addrspace(1) getelementptr inbounds ([2 x i64], ptr addrspace(1) @__profc_kernel1, i32 0, i32 1) to ptr), ptr null, i64 1)
+;; Second counter slot uses a GEP into the per-kernel counter array and the
+;; matching uniform-counter array.
+; CHECK: call void @__llvm_profile_instrument_gpu(ptr addrspacecast (ptr addrspace(1) getelementptr inbounds ([2 x i64], ptr addrspace(1) @__profc_kernel1, i32 0, i32 1) to ptr), ptr addrspacecast (ptr addrspace(1) getelementptr inbounds ([2 x i64], ptr addrspace(1) @__llvm_prf_unifcnt_kernel1, i32 0, i32 1) to ptr), i64 1)
diff --git a/llvm/test/Instrumentation/InstrProfiling/amdgpu-instrumentation.ll b/llvm/test/Instrumentation/InstrProfiling/amdgpu-instrumentation.ll
index efe53ab1ebdfb..963e141934df1 100644
--- a/llvm/test/Instrumentation/InstrProfiling/amdgpu-instrumentation.ll
+++ b/llvm/test/Instrumentation/InstrProfiling/amdgpu-instrumentation.ll
@@ -1,8 +1,16 @@
-;; Test basic AMDGPU PGO instrumentation lowering.
-;; Verifies that each instrumentation point lowers directly to a call to
-;; __llvm_profile_instrument_gpu with a null uniform-counter argument.
+;; Test AMDGPU PGO instrumentation lowering with multiple basic blocks.
+;; Verifies:
+;; 1. Sampling decision is computed once in the entry block.
+;; 2. Each instrumentation point calls __llvm_profile_instrument_gpu behind the
+;;    sampling guard branch.
+;; 3. No-sampling mode (sampling=0) calls __llvm_profile_instrument_gpu unconditionally.
 
-; RUN: opt %s -mtriple=amdgcn-amd-amdhsa -passes=instrprof -S | FileCheck %s
+; RUN: opt %s -mtriple=amdgcn-amd-amdhsa -passes=instrprof \
+; RUN:   -offload-pgo-sampling=3 -S \
+; RUN:   | FileCheck %s --check-prefix=SAMPLED
+; RUN: opt %s -mtriple=amdgcn-amd-amdhsa -passes=instrprof \
+; RUN:   -offload-pgo-sampling=0 -S \
+; RUN:   | FileCheck %s --check-prefix=NOSAMPLE
 
 @__hip_cuid_test01 = addrspace(1) global i8 0
 @__profn_test_kernel = private constant [11 x i8] c"test_kernel"
@@ -24,9 +32,26 @@ if.end:
 
 declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
 
-; CHECK-LABEL: define {{.*}} @test_kernel
-; CHECK-NOT: @__llvm_profile_sampling_gpu
-; CHECK: call void @__llvm_profile_instrument_gpu(
-; CHECK-SAME: ptr addrspacecast (ptr addrspace(1) @__profc_test_kernel to ptr), ptr null, i64 1)
-; CHECK: call void @__llvm_profile_instrument_gpu(
-; CHECK-SAME: ptr addrspacecast (ptr addrspace(1) getelementptr inbounds ([4 x i64], ptr addrspace(1) @__profc_test_kernel, i32 0, i32 1) to ptr), ptr null, i64 1)
+;; ---- Sampled mode (sampling=3) ----
+
+;; Entry block: sampling decision computed once
+; SAMPLED-LABEL: define {{.*}} @test_kernel
+; SAMPLED: entry:
+; SAMPLED: %pgo.sampled = call i32 @__llvm_profile_sampling_gpu(i32 3)
+; SAMPLED: %pgo.matched = icmp ne i32 %pgo.sampled, 0
+; SAMPLED: br i1 %pgo.matched, label %po_then, label %po_cont
+
+;; Second instrumentation point reuses same sampling decision
+; SAMPLED: br i1 %pgo.matched, label %po_then{{[0-9]+}}, label %po_cont{{[0-9]+}}
+
+;; Both instrumentation points call the library function
+; SAMPLED: call void @__llvm_profile_instrument_gpu(
+; SAMPLED: call void @__llvm_profile_instrument_gpu(
+
+;; ---- No-sampling mode (sampling=0) ----
+
+;; No sampling guard — direct call
+; NOSAMPLE-LABEL: define {{.*}} @test_kernel
+; NOSAMPLE: entry:
+; NOSAMPLE-NOT: @__llvm_profile_sampling_gpu
+; NOSAMPLE: call void @__llvm_profile_instrument_gpu(
diff --git a/llvm/test/Instrumentation/InstrProfiling/amdgpu-uniform-counters.ll b/llvm/test/Instrumentation/InstrProfiling/amdgpu-uniform-counters.ll
new file mode 100644
index 0000000000000..ffdd41043bdbe
--- /dev/null
+++ b/llvm/test/Instrumentation/InstrProfiling/amdgpu-uniform-counters.ll
@@ -0,0 +1,22 @@
+;; Test that AMDGPU targets generate uniform counter arrays alongside regular
+;; counters. The uniform counter is passed to __llvm_profile_instrument_gpu which
+;; updates it when all lanes in the wave are active.
+
+; RUN: opt %s -mtriple=amdgcn-amd-amdhsa -passes=instrprof -S | FileCheck %s
+
+ at __hip_cuid_test123 = addrspace(1) global i8 0
+ at __profn_test_kernel = private constant [11 x i8] c"test_kernel"
+
+define amdgpu_kernel void @test_kernel() {
+  call void @llvm.instrprof.increment(ptr @__profn_test_kernel, i64 12345, i32 1, i32 0)
+  ret void
+}
+
+declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
+
+;; Per-function counter + uniform-counter globals (comdat)
+; CHECK: @__profc_test_kernel = linkonce_odr protected addrspace(1) global [1 x i64]
+; CHECK: @__llvm_prf_unifcnt_test_kernel = linkonce_odr protected addrspace(1) global [1 x i64]
+
+;; __llvm_profile_instrument_gpu receives counter and uniform-counter bases
+; CHECK: call void @__llvm_profile_instrument_gpu(ptr addrspacecast (ptr addrspace(1) @__profc_test_kernel to ptr), ptr addrspacecast (ptr addrspace(1) @__llvm_prf_unifcnt_test_kernel to ptr), i64 1)
diff --git a/llvm/test/Instrumentation/InstrProfiling/amdgpu-wave32.ll b/llvm/test/Instrumentation/InstrProfiling/amdgpu-wave32.ll
new file mode 100644
index 0000000000000..9cdf733da1c6f
--- /dev/null
+++ b/llvm/test/Instrumentation/InstrProfiling/amdgpu-wave32.ll
@@ -0,0 +1,35 @@
+;; Test that AMDGPU PGO instrumentation generates library calls for Wave32.
+;; Verifies sampling guard and __llvm_profile_instrument_gpu call with correct
+;; addrspace(1) counter pointers.
+
+; RUN: opt %s -mtriple=amdgcn-amd-amdhsa -passes=instrprof -S | FileCheck %s
+
+ at __hip_cuid_abcdef456 = addrspace(1) global i8 0
+ at __profn_kernel_w32 = private constant [10 x i8] c"kernel_w32"
+
+define amdgpu_kernel void @kernel_w32() #0 {
+  call void @llvm.instrprof.increment(ptr @__profn_kernel_w32, i64 12345, i32 1, i32 0)
+  ret void
+}
+
+declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
+
+attributes #0 = { "target-cpu"="gfx1100" }
+
+;; Per-function comdat counters + uniform counters
+; CHECK: @__profc_kernel_w32 = linkonce_odr protected addrspace(1) global [1 x i64]
+; CHECK: @__llvm_prf_unifcnt_kernel_w32 = linkonce_odr protected addrspace(1) global [1 x i64]
+; CHECK: @__profd_kernel_w32 = linkonce_odr protected addrspace(1) global { {{.*}} i16 0, i32 0 }
+
+;; Check wave size stored via intrinsic
+; CHECK: %wavesize.i16 = trunc i32 %{{.*}} to i16
+; CHECK: store i16 %wavesize.i16, ptr addrspace(1) getelementptr inbounds {{.*}} @__profd_kernel_w32
+
+;; Check sampling guard (default sampling=3)
+; CHECK: %pgo.sampled = call i32 @__llvm_profile_sampling_gpu(i32 3)
+; CHECK: %pgo.matched = icmp ne i32 %pgo.sampled, 0
+; CHECK: br i1 %pgo.matched, label %po_then, label %po_cont
+
+;; Check library call
+; CHECK: po_then:
+; CHECK: call void @__llvm_profile_instrument_gpu(ptr addrspacecast (ptr addrspace(1) @__profc_kernel_w32 to ptr), ptr addrspacecast (ptr addrspace(1) @__llvm_prf_unifcnt_kernel_w32 to ptr), i64 1)
diff --git a/llvm/test/Instrumentation/InstrProfiling/amdgpu-wave64.ll b/llvm/test/Instrumentation/InstrProfiling/amdgpu-wave64.ll
new file mode 100644
index 0000000000000..0a776ba5da687
--- /dev/null
+++ b/llvm/test/Instrumentation/InstrProfiling/amdgpu-wave64.ll
@@ -0,0 +1,34 @@
+;; Test that AMDGPU PGO instrumentation generates library calls for Wave64.
+;; Wave64 targets (e.g., gfx908) should embed wave size 64 in profile data.
+
+; RUN: opt %s -mtriple=amdgcn-amd-amdhsa -passes=instrprof -S | FileCheck %s
+
+ at __hip_cuid_abcdef123 = addrspace(1) global i8 0
+ at __profn_kernel_w64 = private constant [10 x i8] c"kernel_w64"
+
+define amdgpu_kernel void @kernel_w64() #0 {
+  call void @llvm.instrprof.increment(ptr @__profn_kernel_w64, i64 12345, i32 1, i32 0)
+  ret void
+}
+
+declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
+
+attributes #0 = { "target-cpu"="gfx908" }
+
+;; Per-function comdat counters
+; CHECK: @__profc_kernel_w64 = linkonce_odr protected addrspace(1) global [1 x i64]
+; CHECK: @__llvm_prf_unifcnt_kernel_w64 = linkonce_odr protected addrspace(1) global [1 x i64]
+; CHECK: @__profd_kernel_w64 = linkonce_odr protected addrspace(1) global { {{.*}} i16 0, i32 0 }
+
+;; Check wave size stored via intrinsic
+; CHECK: %wavesize.i16 = trunc i32 %{{.*}} to i16
+; CHECK: store i16 %wavesize.i16, ptr addrspace(1) getelementptr inbounds {{.*}} @__profd_kernel_w64
+
+;; Check sampling guard
+; CHECK: %pgo.sampled = call i32 @__llvm_profile_sampling_gpu(i32 3)
+; CHECK: %pgo.matched = icmp ne i32 %pgo.sampled, 0
+; CHECK: br i1 %pgo.matched, label %po_then, label %po_cont
+
+;; Check library call
+; CHECK: po_then:
+; CHECK: call void @__llvm_profile_instrument_gpu(ptr addrspacecast (ptr addrspace(1) @__profc_kernel_w64 to ptr), ptr addrspacecast (ptr addrspace(1) @__llvm_prf_unifcnt_kernel_w64 to ptr), i64 1)
diff --git a/llvm/test/Instrumentation/InstrProfiling/coverage.ll b/llvm/test/Instrumentation/InstrProfiling/coverage.ll
index 08cbcaa962b76..695a8829fdf75 100644
--- a/llvm/test/Instrumentation/InstrProfiling/coverage.ll
+++ b/llvm/test/Instrumentation/InstrProfiling/coverage.ll
@@ -5,12 +5,12 @@ target triple = "aarch64-unknown-linux-gnu"
 
 @__profn_foo = private constant [3 x i8] c"foo"
 ; CHECK: @__profc_foo = private global [1 x i8] c"\FF", section "__llvm_prf_cnts", comdat, align 1
-; CHECK: @__profd_foo = private global { i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i32 } { i64 {{.*}}, i64 {{.*}}, i64 sub (i64 ptrtoint (ptr @__profc_foo to i64)
-; BINARY: @__profd_foo = private global { i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i32 } { i64 {{.*}}, i64 {{.*}}, i64 ptrtoint (ptr @__profc_foo to i64),
+; CHECK: @__profd_foo = private global { i64, i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i16, i32 } { i64 {{.*}}, i64 {{.*}}, i64 sub (i64 ptrtoint (ptr @__profc_foo to i64)
+; BINARY: @__profd_foo = private global { i64, i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i16, i32 } { i64 {{.*}}, i64 {{.*}}, i64 ptrtoint (ptr @__profc_foo to i64),
 @__profn_bar = private constant [3 x i8] c"bar"
 ; CHECK: @__profc_bar = private global [1 x i8] c"\FF", section "__llvm_prf_cnts", comdat, align 1
-; CHECK: @__profd_bar = private global { i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i32 } { i64 {{.*}}, i64 {{.*}}, i64 sub (i64 ptrtoint (ptr @__profc_bar to i64)
-; BINARY: @__profd_bar = private global { i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i32 } { i64 {{.*}}, i64 {{.*}}, i64 ptrtoint (ptr @__profc_bar to i64),
+; CHECK: @__profd_bar = private global { i64, i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i16, i32 } { i64 {{.*}}, i64 {{.*}}, i64 sub (i64 ptrtoint (ptr @__profc_bar to i64)
+; BINARY: @__profd_bar = private global { i64, i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i16, i32 } { i64 {{.*}}, i64 {{.*}}, i64 ptrtoint (ptr @__profc_bar to i64),
 
 ; CHECK: @__llvm_prf_nm = {{.*}} section "__llvm_prf_names"
 ; BINARY: @__llvm_prf_nm ={{.*}} section "__llvm_covnames"
diff --git a/llvm/test/Instrumentation/InstrProfiling/inline-data-var-create.ll b/llvm/test/Instrumentation/InstrProfiling/inline-data-var-create.ll
index 456103164378e..8644c320f867b 100644
--- a/llvm/test/Instrumentation/InstrProfiling/inline-data-var-create.ll
+++ b/llvm/test/Instrumentation/InstrProfiling/inline-data-var-create.ll
@@ -7,17 +7,18 @@
 
 target triple = "x86_64-unknown-linux-gnu"
 
-; INLINEFIRST: @__profd_foo = private global{{.*}}zeroinitializer, i32 21
-; INLINEFIRST: @__profd_bar = private global{{.*}}zeroinitializer, i32 23
-; INLINEFIRST: @__profd_foobar = private global{{.*}}zeroinitializer, i32 99
-
-; INLINEAFTER: @__profd_foobar = private global{{.*}}zeroinitializer, i32 99
-; INLINEAFTER: @__profd_foo = private global{{.*}}zeroinitializer, i32 21
-; INLINEAFTER: @__profd_bar = private global{{.*}}zeroinitializer, i32 23
-
-; NOINLINE: @__profd_foobar = private global{{.*}}zeroinitializer, i32 99
-; NOINLINE: @__profd_foo = private global{{.*}}zeroinitializer, i32 21
-; NOINLINE: @__profd_bar = private global{{.*}}zeroinitializer, i32 23
+;; Note: struct layout is { ..., i32, [3 x i16], i16, i32 } where the i16 is OffloadDeviceWaveSize
+; INLINEFIRST: @__profd_foo = private global{{.*}}i16 0, i32 21
+; INLINEFIRST: @__profd_bar = private global{{.*}}i16 0, i32 23
+; INLINEFIRST: @__profd_foobar = private global{{.*}}i16 0, i32 99
+
+; INLINEAFTER: @__profd_foobar = private global{{.*}}i16 0, i32 99
+; INLINEAFTER: @__profd_foo = private global{{.*}}i16 0, i32 21
+; INLINEAFTER: @__profd_bar = private global{{.*}}i16 0, i32 23
+
+; NOINLINE: @__profd_foobar = private global{{.*}}i16 0, i32 99
+; NOINLINE: @__profd_foo = private global{{.*}}i16 0, i32 21
+; NOINLINE: @__profd_bar = private global{{.*}}i16 0, i32 23
 
 declare void @llvm.instrprof.increment(ptr %0, i64 %1, i32 %2, i32 %3)
 declare void @llvm.instrprof.mcdc.parameters(ptr %0, i64 %1, i32 %2)
diff --git a/llvm/test/Instrumentation/InstrProfiling/platform.ll b/llvm/test/Instrumentation/InstrProfiling/platform.ll
index 9c76a5caf2a51..5f7eb47d2b611 100644
--- a/llvm/test/Instrumentation/InstrProfiling/platform.ll
+++ b/llvm/test/Instrumentation/InstrProfiling/platform.ll
@@ -9,26 +9,35 @@
 ; RUN: opt < %s -mtriple=x86_64-pc-windows -passes=instrprof -S | FileCheck %s -check-prefix=WINDOWS
 ; RUN: opt < %s -mtriple=powerpc64-ibm-aix-xcoff -passes=instrprof -S | FileCheck %s -check-prefix=AIX
 ; RUN: opt < %s -mtriple=arm-elf -passes=instrprof -S | FileCheck %s -check-prefix=BAREMETAL
+; RUN: opt < %s -mtriple=amdgcn-amd-amdhsa -passes=instrprof -S | FileCheck %s -check-prefix=AMDGPU
 
 @__profn_foo = private constant [3 x i8] c"foo"
 ; MACHO-NOT: __profn_foo
 ; ELF-NOT: __profn_foo
 ; WINDOWS-NOT: __profn_foo
 ; AIX-NOT: __profn_foo
+; AMDGPU-NOT: __profn_foo
 
 ; MACHO: @__profc_foo = private global [1 x i64] zeroinitializer, section "__DATA,__llvm_prf_cnts", align 8
 ; ELF: @__profc_foo = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", comdat, align 8
 ; WINDOWS: @__profc_foo = private global [1 x i64] zeroinitializer, section ".lprfc$M", align 8
 ; AIX: @__profc_foo = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", align 8
+;; AMDGPU without CUID: COMDAT counters + uniform counter slot (HIP-style GPU profiling).
+; AMDGPU: $__llvm_prf_unifcnt_foo = comdat any
+; AMDGPU: @__profc_foo = linkonce_odr protected addrspace(1) global [{{[0-9]+}} x i64] zeroinitializer, section "__llvm_prf_cnts", comdat, align 8
+; AMDGPU: @__llvm_prf_unifcnt_foo = linkonce_odr protected addrspace(1) global [{{[0-9]+}} x i64] zeroinitializer, section "__llvm_prf_ucnts", comdat, align 8
 
 ; MACHO: @__profd_foo = private {{.*}}, section "__DATA,__llvm_prf_data,regular,live_support", align 8
 ; ELF: @__profd_foo = private {{.*}}, section "__llvm_prf_data", comdat($__profc_foo), align 8
 ; WINDOWS: @__profd_foo = private global {{.*}}, section ".lprfd$M", align 8
 ; AIX: @__profd_foo = private {{.*}}, section "__llvm_prf_data", align 8
+;; AMDGPU without CUID uses per-function data (not alias)
+; AMDGPU: @__profd_foo = linkonce_odr protected addrspace(1) global {{.*}}, section "__llvm_prf_data", comdat($__profc_foo), align 8
 
 ; ELF: @__llvm_prf_nm = private constant [{{.*}} x i8] c"{{.*}}", section "{{.*}}__llvm_prf_names"{{.*}}, align 1
 ; WINDOWS: @__llvm_prf_nm = private constant [{{.*}} x i8] c"{{.*}}", section "{{.*}}lprfn$M", align 1
 ; AIX: @__llvm_prf_nm = private constant [{{.*}} x i8] c"{{.*}}", section "{{.*}}__llvm_prf_names", align 1
+; AMDGPU: @__llvm_prf_nm = private addrspace(1) constant [{{.*}} x i8] c"{{.*}}", section "__llvm_prf_names", align 1
 
 define void @foo() {
   call void @llvm.instrprof.increment(ptr @__profn_foo, i64 0, i32 1, i32 0)
@@ -37,6 +46,9 @@ define void @foo() {
 
 declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
 
+;; AMDGPU without CUID uses standard per-function allocation (for OpenMP compatibility)
+;; Start/stop symbols behavior is platform-specific
+
 ;; Emit registration functions for platforms that don't find the
 ;; symbols by their sections.
 
@@ -48,6 +60,7 @@ declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
 ; WINDOWS-NOT: define internal void @__llvm_profile_register_functions
 ; AIX-NOT: define internal void @__llvm_profile_register_functions
 ; BAREMETAL-NOT: define internal void @__llvm_profile_register_functions
+; AMDGPU-NOT: define internal void @__llvm_profile_register_functions
 
 ;; PR38340: When dynamic registration is used, we had a bug where we'd register
 ;; something that's not a __profd_* variable.
@@ -60,3 +73,4 @@ declare void @llvm.instrprof.increment(ptr, i64, i32, i32)
 ; WINDOWS-NOT: define internal void @__llvm_profile_init
 ; AIX-NOT: define internal void @__llvm_profile_init
 ; BAREMETAL-NOT: define internal void @__llvm_profile_init
+; AMDGPU-NOT: define internal void @__llvm_profile_init
diff --git a/llvm/test/Transforms/PGOProfile/Inputs/thinlto_indirect_call_promotion.profraw b/llvm/test/Transforms/PGOProfile/Inputs/thinlto_indirect_call_promotion.profraw
index 3daa98f937b691880ffff203c9426bfacddf749d..596ba3236108eedb5cf7546418d99f9d24f2c7d7 100644
GIT binary patch
delta 207
zcmZ3$a)X7lu_!ISs37M*_e4(ni4GiKB*7ODqoTHZ<^TT at H*Vc at 37eQ^#m4{z1}wKh
zBFnVo7W{_*m<)_|fT(z;rIvt0MF3a at 1H*+SAhRaxF)GO^;83~%taNf3<17V;xeN at 9
IC at w+p0h4(_hX4Qo

delta 174
zcmcb?vVeuNu_!ISs37M**F;YHi3tLH8J8YEkzV!x|K>Vr=_jER>#W!UR)JWP`52`p
zJ`fOF at gD*hpdbMzt0h-}CRG5IVqln|Cj&Adj!{X>0!_&Qi0*levlJi(F)%Pf4Te~Y
G-~#}Nbw4lw

diff --git a/llvm/test/Transforms/PGOProfile/comdat_internal.ll b/llvm/test/Transforms/PGOProfile/comdat_internal.ll
index 1bad0db1b4762..c948ad84392a0 100644
--- a/llvm/test/Transforms/PGOProfile/comdat_internal.ll
+++ b/llvm/test/Transforms/PGOProfile/comdat_internal.ll
@@ -13,9 +13,9 @@ $foo = comdat any
 ; CHECK: @__llvm_profile_raw_version = hidden constant i64 {{[0-9]+}}, comdat
 ; CHECK-NOT: __profn__stdin__foo
 ; CHECK: @__profc__stdin__foo.[[#FOO_HASH]] = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", comdat, align 8
-; CHECK: @__profd__stdin__foo.[[#FOO_HASH]] = private global { i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i32 } { i64 {{.*}}, i64 [[#FOO_HASH]], i64 sub (i64 ptrtoint (ptr @__profc__stdin__foo.742261418966908927 to i64), i64 ptrtoint (ptr @__profd__stdin__foo.742261418966908927 to i64)), i64 0, ptr null
+; CHECK: @__profd__stdin__foo.[[#FOO_HASH]] = private global { i64, i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i16, i32 } { i64 {{.*}}, i64 [[#FOO_HASH]], i64 sub (i64 ptrtoint (ptr @__profc__stdin__foo.742261418966908927 to i64), i64 ptrtoint (ptr @__profd__stdin__foo.742261418966908927 to i64)), i64 0, i64 0, ptr null
 ; CHECK-NOT: @foo
-; CHECK-SAME: , ptr null, i32 1, [3 x i16] zeroinitializer, i32 0 }, section "__llvm_prf_data", comdat($__profc__stdin__foo.[[#FOO_HASH]]), align 8
+; CHECK-SAME: , ptr null, i32 1, [3 x i16] zeroinitializer, i16 0, i32 0 }, section "__llvm_prf_data", comdat($__profc__stdin__foo.[[#FOO_HASH]]), align 8
 ; CHECK: @__llvm_prf_nm
 ; CHECK: @llvm.compiler.used
 
diff --git a/llvm/test/Transforms/PGOProfile/instrprof_burst_sampling_fast.ll b/llvm/test/Transforms/PGOProfile/instrprof_burst_sampling_fast.ll
index 56d8364d8f543..7261cc24b0dcf 100644
--- a/llvm/test/Transforms/PGOProfile/instrprof_burst_sampling_fast.ll
+++ b/llvm/test/Transforms/PGOProfile/instrprof_burst_sampling_fast.ll
@@ -14,7 +14,7 @@ $__llvm_profile_raw_version = comdat any
 
 ; SAMPLE-VAR: @__llvm_profile_sampling = thread_local global i16 0, comdat
 ; SAMPLE-VAR: @__profc_f = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", comdat, align 8
-; SAMPLE-VAR: @__profd_f = private global { i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i32 } { i64 -3706093650706652785, i64 12884901887, i64 sub (i64 ptrtoint (ptr @__profc_f to i64), i64 ptrtoint (ptr @__profd_f to i64)), i64 0, ptr @f.local, ptr null, i32 1, [3 x i16] zeroinitializer, i32 0 }, section "__llvm_prf_data", comdat($__profc_f), align 8
+; SAMPLE-VAR: @__profd_f = private global { i64, i64, i64, i64, i64, ptr, ptr, i32, [3 x i16], i16, i32 } { i64 -3706093650706652785, i64 12884901887, i64 sub (i64 ptrtoint (ptr @__profc_f to i64), i64 ptrtoint (ptr @__profd_f to i64)), i64 0, i64 0, ptr @f.local, ptr null, i32 1, [3 x i16] zeroinitializer, i16 0, i32 0 }, section "__llvm_prf_data", comdat($__profc_f), align 8
 ; SAMPLE-VAR: @__llvm_prf_nm = private constant {{.*}}, section "__llvm_prf_names", align 1
 ; SAMPLE-VAR: @llvm.compiler.used = appending global [2 x ptr] [ptr @__llvm_profile_sampling, ptr @__profd_f], section "llvm.metadata"
 ; SAMPLE-VAR: @llvm.used = appending global [1 x ptr] [ptr @__llvm_prf_nm], section "llvm.metadata"
diff --git a/llvm/test/Transforms/PGOProfile/vtable_profile.ll b/llvm/test/Transforms/PGOProfile/vtable_profile.ll
index aae1e2d8b4e49..0c554db05cfb4 100644
--- a/llvm/test/Transforms/PGOProfile/vtable_profile.ll
+++ b/llvm/test/Transforms/PGOProfile/vtable_profile.ll
@@ -49,7 +49,7 @@ target triple = "x86_64-unknown-linux-gnu"
 @llvm.compiler.used = appending global [1 x ptr] [ptr @_ZTV5Base1], section "llvm.metadata"
 
 ; GEN: __llvm_profile_raw_version = comdat any
-; GEN: __llvm_profile_raw_version = hidden constant i64 72057594037927946, comdat
+; GEN: __llvm_profile_raw_version = hidden constant i64 72057594037927947, comdat
 ; GEN: __profn__Z4funci = private constant [8 x i8] c"_Z4funci"
 
 ; LOWER: $__profvt__ZTV7Derived = comdat nodeduplicate
diff --git a/llvm/test/tools/llvm-profdata/Inputs/c-general.profraw b/llvm/test/tools/llvm-profdata/Inputs/c-general.profraw
index a3e884343942ebc70ba95ab4ee006630b6816d80..dec90151a79352bf898c8c9d1d2d23eaa4bf5eed 100644
GIT binary patch
delta 197
zcmeys|3ZMXu_!ISs37M*_e4(ni4GiKB*Ev<^}llG%K!gm%^2F;!Y0QtOHFJr0Lg8T
zXJQeUI8h$N{vgl9ZUA9VHei%zVt{c97!??2KqWQ+CD>j-IUj%$8j~+Fg0xRIV3KFz
n34rkmm=qWbV4Mv=DGmo1?*mXu0>&|5R$#mVF>Z4Ka{~(iCWJse

delta 160
zcmaDM at PVJRu_!ISs37M**F;YHi3tLH4>VUWO0D|;f9~Aye@{av)_G05AkX9?G4Z4P
zWC=!jCUpl8D*?#5-~nQ-U{qk-0b+ar3T+dZypTzLvIUa at V+5Ghz$DMKrUT480TkQ-
nW(hFMGgW{ECteVkoWU%>m;n}Fz^uS%0Afv?z_IxQa|H_kt<yS_

diff --git a/llvm/test/tools/llvm-profdata/Inputs/compressed.profraw b/llvm/test/tools/llvm-profdata/Inputs/compressed.profraw
index e3f77e870d4d20828119348e70eb44e6d39e0ec0..778e80fce2691a35d9654748763b0aa8c233ec13 100644
GIT binary patch
delta 197
zcmX at Wze9ktu_!ISs37M*_e4(ni4GiKB*EwKYmREqs{j9Gk0sqI4xJpwEH$yg03^3T
zo{2?Z;zW56`-40ay8(nf*?>`=i2=qbU{qk70hQPQlwf-S<$M52XiUDy2+}^;fJvT-
nCjiDPU{YW#fN?eer8pd5ybnMr2^hzKS%L8e#JJ4`%mFL_d2B$e

delta 159
zcmdlXaDbn)u_!ISs37M**F;YHi3tL{4=Vq*Em-yc|6E_D%J_-3-V-m#Gr34i{3t(J
zf>EAH-2uc(0J1K4fLJRS6&QDb7$1N_+XN;rWRjn3!KA<#0cJHY$uq6#05eYj1vh|M
m0?hJE6=1=M7X&6}FbgndfW;RuD=->>SQ963Z2rLP!2$q9VL7b;

diff --git a/llvm/test/tools/llvm-profdata/binary-ids-padding.test b/llvm/test/tools/llvm-profdata/binary-ids-padding.test
index cc3e6c38e6907..f20161f9886e3 100644
--- a/llvm/test/tools/llvm-profdata/binary-ids-padding.test
+++ b/llvm/test/tools/llvm-profdata/binary-ids-padding.test
@@ -18,7 +18,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201rforpl\377' > %t.profraw
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t.profraw
 // There will be 2 20-byte binary IDs, so the total Binary IDs size will be 64 bytes.
 //   2 * 8  binary ID sizes
 // + 2 * 20 binary IDs (of size 20)
@@ -32,6 +32,9 @@ RUN: printf '\3\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\20\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\4\0\1\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -64,12 +67,14 @@ RUN: printf '\0\0\4\0\1\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 
 RUN: printf '\067\265\035\031\112\165\023\344' >> %t.profraw
 RUN: printf '\02\0\0\0\0\0\0\0' >> %t.profraw
-RUN: printf '\310\377\3\0\1\0\0\0' >> %t.profraw
+RUN: printf '\270\377\3\0\1\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -81,8 +86,8 @@ RUN: printf '\067\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\101\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\7\0foo\1bar\0\0\0\0\0\0\0' >> %t.profraw
 
-// RUN: llvm-profdata show --binary-ids  %t.profraw | FileCheck %s
-// CHECK: Instrumentation level: Front-end
-// CHECK: Binary IDs:
-// CHECK-NEXT: 0001020304050607000102030405060700010203
-// CHECK-NEXT: 0101010101010101020202020202020203030303
+RUN: llvm-profdata show --binary-ids  %t.profraw | FileCheck %s
+CHECK: Instrumentation level: Front-end
+CHECK: Binary IDs:
+CHECK-NEXT: 0001020304050607000102030405060700010203
+CHECK-NEXT: 0101010101010101020202020202020203030303
diff --git a/llvm/test/tools/llvm-profdata/large-binary-id-size.test b/llvm/test/tools/llvm-profdata/large-binary-id-size.test
index b62bdad4ddb21..4ffcc89c421ee 100644
--- a/llvm/test/tools/llvm-profdata/large-binary-id-size.test
+++ b/llvm/test/tools/llvm-profdata/large-binary-id-size.test
@@ -2,7 +2,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201rforpl\377' > %t.profraw
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\40\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -17,6 +17,9 @@ RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 
 // Check for a corrupted size being too large past the end of the file.
 RUN: printf '\7\7\7\7\7\7\7\7' >> %t.profraw
@@ -24,5 +27,5 @@ RUN: printf '\0\1\2\3\4\5\6\7' >> %t.profraw
 RUN: printf '\0\1\2\3\4\5\6\7' >> %t.profraw
 RUN: printf '\0\1\2\3\0\0\0\0' >> %t.profraw
 
-// RUN: not llvm-profdata show --binary-ids  %t.profraw 2>&1 | FileCheck %s
-// CHECK: malformed instrumentation profile data: not enough data to read binary id data
+RUN: not llvm-profdata show --binary-ids  %t.profraw 2>&1 | FileCheck %s
+CHECK: malformed instrumentation profile data: not enough data to read binary id data
diff --git a/llvm/test/tools/llvm-profdata/malformed-not-space-for-another-header.test b/llvm/test/tools/llvm-profdata/malformed-not-space-for-another-header.test
index 705e5efaf5875..4f9876e4d8fb2 100644
--- a/llvm/test/tools/llvm-profdata/malformed-not-space-for-another-header.test
+++ b/llvm/test/tools/llvm-profdata/malformed-not-space-for-another-header.test
@@ -18,7 +18,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201rforpl\377' > %t.profraw
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -26,6 +26,9 @@ RUN: printf '\1\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\10\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\4\0\1\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -46,10 +49,12 @@ RUN: printf '\254\275\030\333\114\302\370\134' >> %t.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\4\0\1\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\4\0\3\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 
 RUN: printf '\023\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\3\0foo\0\0\0' >> %t.profraw
diff --git a/llvm/test/tools/llvm-profdata/malformed-num-counters-zero.test b/llvm/test/tools/llvm-profdata/malformed-num-counters-zero.test
index 157c13b926a7e..e7c2fa7877c1a 100644
--- a/llvm/test/tools/llvm-profdata/malformed-num-counters-zero.test
+++ b/llvm/test/tools/llvm-profdata/malformed-num-counters-zero.test
@@ -18,7 +18,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201rforpl\377' > %t.profraw
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -26,6 +26,9 @@ RUN: printf '\1\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\10\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\4\0\1\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -56,6 +59,7 @@ RUN: cp %t.profraw %t-good.profraw
 // Make NumCounters = 0 so that we get "number of counters is zero" error message
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 
 RUN: printf '\023\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\3\0foo\0\0\0' >> %t.profraw
@@ -64,12 +68,14 @@ RUN: not llvm-profdata show %t.profraw 2>&1 | FileCheck %s --check-prefix=ZERO
 ZERO: malformed instrumentation profile data: number of counters is zero
 
 // Test a counter value greater than 2^56.
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bad.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t-bad.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bad.profraw
 // Counter value is 72057594037927937
 RUN: printf '\1\0\0\0\0\0\0\1' >> %t-bad.profraw
 RUN: printf '\3\0foo\0\0\0' >> %t-bad.profraw
 
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-good.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t-good.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-good.profraw
 // Counter value is 72057594037927937
diff --git a/llvm/test/tools/llvm-profdata/malformed-ptr-to-counter-array.test b/llvm/test/tools/llvm-profdata/malformed-ptr-to-counter-array.test
index 83cf76f68fb63..ee4b63986e43f 100644
--- a/llvm/test/tools/llvm-profdata/malformed-ptr-to-counter-array.test
+++ b/llvm/test/tools/llvm-profdata/malformed-ptr-to-counter-array.test
@@ -18,7 +18,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201rforpl\377' > %t.profraw
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -26,6 +26,9 @@ RUN: printf '\2\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\10\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\6\0\1\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
@@ -51,11 +54,12 @@ RUN: printf '\02\0\0\0\0\0\0\0' >> %t.profraw
 // the profile reader should error out.
 RUN: printf '\11\0\6\0\1\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
-
+RUN: printf '\0\0\6\0\2\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\02\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
 
 // Counter Section
 
diff --git a/llvm/test/tools/llvm-profdata/misaligned-binary-ids-size.test b/llvm/test/tools/llvm-profdata/misaligned-binary-ids-size.test
index 0f20a1b0b369c..b43b2408869c2 100644
--- a/llvm/test/tools/llvm-profdata/misaligned-binary-ids-size.test
+++ b/llvm/test/tools/llvm-profdata/misaligned-binary-ids-size.test
@@ -2,7 +2,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201rforpl\377' > %t.profraw
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t.profraw
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t.profraw
 // We should fail on this because the binary IDs is not a multiple of 8 bytes.
 RUN: printf '\77\0\0\0\0\0\0\0' >> %t.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t.profraw
diff --git a/llvm/test/tools/llvm-profdata/mismatched-raw-profile-header.test b/llvm/test/tools/llvm-profdata/mismatched-raw-profile-header.test
index dfa163f1f3439..8d71914a99995 100644
--- a/llvm/test/tools/llvm-profdata/mismatched-raw-profile-header.test
+++ b/llvm/test/tools/llvm-profdata/mismatched-raw-profile-header.test
@@ -11,6 +11,9 @@ RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\20' >> %t
 RUN: printf '\0\0\0\1\0\4\0\0' >> %t
 RUN: printf '\0\0\0\2\0\4\0\0' >> %t
diff --git a/llvm/test/tools/llvm-profdata/profile-version.test b/llvm/test/tools/llvm-profdata/profile-version.test
index e811699ac63ed..7f627cb89c7d0 100644
--- a/llvm/test/tools/llvm-profdata/profile-version.test
+++ b/llvm/test/tools/llvm-profdata/profile-version.test
@@ -2,7 +2,7 @@ Test the profile version.
 
 RUN: llvm-profdata merge -o %t.profdata %p/Inputs/basic.proftext
 RUN: llvm-profdata show --profile-version %t.profdata | FileCheck %s
-CHECK: Profile version: 13
+CHECK: Profile version: 14
 
 RUN: llvm-profdata merge -o %t.prev.profdata %p/Inputs/basic.proftext --write-prev-version
 RUN: llvm-profdata show --profile-version %t.prev.profdata | FileCheck %s --check-prefix=PREV
diff --git a/llvm/test/tools/llvm-profdata/raw-32-bits-be.test b/llvm/test/tools/llvm-profdata/raw-32-bits-be.test
index 6b3dc96b36270..4f24eb624a2a7 100644
--- a/llvm/test/tools/llvm-profdata/raw-32-bits-be.test
+++ b/llvm/test/tools/llvm-profdata/raw-32-bits-be.test
@@ -3,7 +3,7 @@
 UNSUPPORTED: system-windows
 // Header
 RUN: printf '\377lprofR\201' > %t
-RUN: printf '\0\0\0\0\0\0\0\12' >> %t
+RUN: printf '\0\0\0\0\0\0\0\13' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\2' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
@@ -11,6 +11,9 @@ RUN: printf '\0\0\0\0\0\0\0\3' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\4' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\20' >> %t
 RUN: printf '\0\0\0\0\1\0\0\0' >> %t
 RUN: printf '\0\0\0\0\3\0\0\0' >> %t
@@ -22,22 +25,28 @@ RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\134\370\302\114\333\030\275\254' >> %t
 RUN: printf '\0\0\0\0\0\0\0\1' >> %t
 RUN: printf '\1\0\0\0' >> %t
+RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\3\0\0\0' >> %t
 RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\0\0\0\1' >> %t
-RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0' >> %t
 RUN: printf '\0\0\0\3' >> %t
+RUN: printf '\0\0\0\0' >> %t
 
 RUN: printf '\344\023\165\112\031\035\265\067' >> %t
 RUN: printf '\0\0\0\0\0\0\0\2' >> %t
-RUN: printf '\0\377\377\330' >> %t
-RUN: printf '\2\377\377\323' >> %t
+RUN: printf '\0\377\377\320' >> %t
+RUN: printf '\0\0\0\0' >> %t
+RUN: printf '\2\377\377\313' >> %t
 RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\0\0\0\2' >> %t
-RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0' >> %t
 RUN: printf '\0\0\0\1' >> %t
+RUN: printf '\0\0\0\0' >> %t
 
 RUN: printf '\0\0\0\0\0\0\0\023' >> %t
 RUN: printf '\0\0\0\0\0\0\0\067' >> %t
diff --git a/llvm/test/tools/llvm-profdata/raw-32-bits-le.test b/llvm/test/tools/llvm-profdata/raw-32-bits-le.test
index 95625565f5c0c..96db225b16b78 100644
--- a/llvm/test/tools/llvm-profdata/raw-32-bits-le.test
+++ b/llvm/test/tools/llvm-profdata/raw-32-bits-le.test
@@ -2,7 +2,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201Rforpl\377' > %t
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\2\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
@@ -10,6 +10,9 @@ RUN: printf '\3\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\4\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\20\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\1\0\0\0\0' >> %t
 RUN: printf '\0\0\0\3\0\0\0\0' >> %t
@@ -21,22 +24,28 @@ RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\254\275\030\333\114\302\370\134' >> %t
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\1' >> %t
+RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\0\0\0\3' >> %t
 RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\1\0\0\0' >> %t
-RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0' >> %t
 RUN: printf '\3\0\0\0' >> %t
+RUN: printf '\0\0\0\0' >> %t
 
 RUN: printf '\067\265\035\031\112\165\023\344' >> %t
 RUN: printf '\02\0\0\0\0\0\0\0' >> %t
-RUN: printf '\330\377\377\0' >> %t
-RUN: printf '\323\377\377\2' >> %t
+RUN: printf '\320\377\377\0' >> %t
+RUN: printf '\0\0\0\0' >> %t
+RUN: printf '\313\377\377\2' >> %t
 RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0' >> %t
 RUN: printf '\2\0\0\0' >> %t
-RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0' >> %t
 RUN: printf '\1\0\0\0' >> %t
+RUN: printf '\0\0\0\0' >> %t
 
 RUN: printf '\023\0\0\0\0\0\0\0' >> %t
 RUN: printf '\067\0\0\0\0\0\0\0' >> %t
diff --git a/llvm/test/tools/llvm-profdata/raw-64-bits-be.test b/llvm/test/tools/llvm-profdata/raw-64-bits-be.test
index 5316ef5a3e559..8d29323ff472a 100644
--- a/llvm/test/tools/llvm-profdata/raw-64-bits-be.test
+++ b/llvm/test/tools/llvm-profdata/raw-64-bits-be.test
@@ -2,7 +2,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\377lprofr\201' > %t
-RUN: printf '\0\0\0\0\0\0\0\12' >> %t
+RUN: printf '\0\0\0\0\0\0\0\13' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\2' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
@@ -10,6 +10,9 @@ RUN: printf '\0\0\0\0\0\0\0\3' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\4' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\20' >> %t
 RUN: printf '\0\0\0\1\0\4\0\0' >> %t
 RUN: printf '\0\0\0\3\0\4\0\0' >> %t
@@ -21,22 +24,22 @@ RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\134\370\302\114\333\030\275\254' >> %t
 RUN: printf '\0\0\0\0\0\0\0\1' >> %t
 RUN: printf '\0\0\0\1\0\4\0\0' >> %t
-RUN: printf '\0\0\0\3\0\4\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\3\0\4\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
-RUN: printf '\0\0\0\1' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
-RUN: printf '\0\0\0\3' >> %t
+RUN: printf '\0\0\0\1\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\3' >> %t
 
 RUN: printf '\344\023\165\112\031\035\265\067' >> %t
-RUN: printf '\0\0\0\0\0\0\0\02' >> %t
-RUN: printf '\0\0\0\1\0\3\377\310' >> %t
-RUN: printf '\0\0\0\3\0\3\377\303' >> %t
+RUN: printf '\0\0\0\0\0\0\0\2' >> %t
+RUN: printf '\0\0\0\1\0\3\377\300' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\3\0\3\377\273' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
-RUN: printf '\0\0\0\02' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
-RUN: printf '\0\0\0\1' >> %t
+RUN: printf '\0\0\0\2\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\1' >> %t
 
 RUN: printf '\0\0\0\0\0\0\0\023' >> %t
 RUN: printf '\0\0\0\0\0\0\0\067' >> %t
diff --git a/llvm/test/tools/llvm-profdata/raw-64-bits-le.test b/llvm/test/tools/llvm-profdata/raw-64-bits-le.test
index 58f4da8cf0386..d90c18cd31b12 100644
--- a/llvm/test/tools/llvm-profdata/raw-64-bits-le.test
+++ b/llvm/test/tools/llvm-profdata/raw-64-bits-le.test
@@ -2,7 +2,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201rforpl\377' > %t
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\2\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
@@ -10,6 +10,9 @@ RUN: printf '\3\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\4\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\20\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\4\0\1\0\0\0' >> %t
 RUN: printf '\0\0\4\0\3\0\0\0' >> %t
@@ -21,22 +24,22 @@ RUN: printf '\0\0\0\0\0\0\0\0' >> %t
 RUN: printf '\254\275\030\333\114\302\370\134' >> %t
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t
 RUN: printf '\0\0\4\0\1\0\0\0' >> %t
-RUN: printf '\0\0\4\0\3\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\4\0\3\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
-RUN: printf '\1\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
-RUN: printf '\3\0\0\0' >> %t
+RUN: printf '\1\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\3\0\0\0' >> %t
 
 RUN: printf '\067\265\035\031\112\165\023\344' >> %t
 RUN: printf '\02\0\0\0\0\0\0\0' >> %t
-RUN: printf '\310\377\3\0\1\0\0\0' >> %t
-RUN: printf '\303\377\3\0\3\0\0\0' >> %t
+RUN: printf '\300\377\3\0\1\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
+RUN: printf '\273\377\3\0\3\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
-RUN: printf '\02\0\0\0' >> %t
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t
-RUN: printf '\1\0\0\0' >> %t
+RUN: printf '\02\0\0\0\0\0\0\0' >> %t
+RUN: printf '\0\0\0\0\1\0\0\0' >> %t
 
 RUN: printf '\023\0\0\0\0\0\0\0' >> %t
 RUN: printf '\067\0\0\0\0\0\0\0' >> %t
diff --git a/llvm/test/tools/llvm-profdata/raw-two-profiles.test b/llvm/test/tools/llvm-profdata/raw-two-profiles.test
index 47cc6fa4fd7fe..e14e70e747a9a 100644
--- a/llvm/test/tools/llvm-profdata/raw-two-profiles.test
+++ b/llvm/test/tools/llvm-profdata/raw-two-profiles.test
@@ -2,7 +2,7 @@
 // whenever \12 (LF) is in the input string.
 UNSUPPORTED: system-windows
 RUN: printf '\201rforpl\377' > %t-foo.profraw
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t-foo.profraw
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
@@ -10,6 +10,9 @@ RUN: printf '\1\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\10\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\4\0\1\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
@@ -24,6 +27,7 @@ RUN: printf '\0\0\4\0\1\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-foo.profraw
 
@@ -31,7 +35,7 @@ RUN: printf '\023\0\0\0\0\0\0\0' >> %t-foo.profraw
 RUN: printf '\3\0foo\0\0\0' >> %t-foo.profraw
 
 RUN: printf '\201rforpl\377' > %t-bar.profraw
-RUN: printf '\12\0\0\0\0\0\0\0' >> %t-bar.profraw
+RUN: printf '\13\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\1\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
@@ -39,6 +43,9 @@ RUN: printf '\2\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\10\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\6\0\1\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
@@ -53,6 +60,7 @@ RUN: printf '\0\0\6\0\1\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
+RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\02\0\0\0\0\0\0\0' >> %t-bar.profraw
 RUN: printf '\0\0\0\0\0\0\0\0' >> %t-bar.profraw
 
diff --git a/llvm/tools/llvm-profdata/llvm-profdata.cpp b/llvm/tools/llvm-profdata/llvm-profdata.cpp
index 6bf3179696e4a..67396f409a93a 100644
--- a/llvm/tools/llvm-profdata/llvm-profdata.cpp
+++ b/llvm/tools/llvm-profdata/llvm-profdata.cpp
@@ -31,6 +31,7 @@
 #include "llvm/Support/BalancedPartitioning.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Discriminator.h"
+#include "llvm/Support/Endian.h"
 #include "llvm/Support/Errc.h"
 #include "llvm/Support/FileSystem.h"
 #include "llvm/Support/Format.h"
@@ -825,6 +826,22 @@ loadInput(const WeightedFile &Input, SymbolRemapper *Remapper,
       I.Name = (*Remapper)(I.Name);
     const StringRef FuncName = I.Name;
     bool Reported = false;
+
+    if (!I.UniformCounts.empty()) {
+      size_t NumBlocks = I.Counts.size();
+      I.UniformityBits.resize((NumBlocks + 7) / 8, 0xFF);
+
+      for (size_t BlockIdx = 0; BlockIdx < NumBlocks; ++BlockIdx) {
+        uint64_t TotalCount = I.Counts[BlockIdx];
+        uint64_t UniformCount = I.UniformCounts[BlockIdx];
+
+        bool IsUniform =
+            (TotalCount == 0) || ((double)UniformCount / TotalCount >= 0.9);
+        if (!IsUniform)
+          I.UniformityBits[BlockIdx / 8] &= ~(1 << (BlockIdx % 8));
+      }
+    }
+
     WC->Writer.addRecord(std::move(I), Input.Weight, [&](Error E) {
       if (Reported) {
         consumeError(std::move(E));
@@ -2979,6 +2996,16 @@ static int showInstrProfile(ShowFormat SFormat, raw_fd_ostream &OS) {
           OS << (I == Start ? "" : ", ") << Func.Counts[I];
         }
         OS << "]\n";
+
+        // Show uniformity bits if present
+        if (!Func.UniformityBits.empty()) {
+          OS << "    Block uniformity: [";
+          for (size_t I = Start, E = Func.Counts.size(); I < E; ++I) {
+            bool IsUniform = Func.isBlockUniform(I);
+            OS << (I == Start ? "" : ", ") << (IsUniform ? "U" : "D");
+          }
+          OS << "]\n";
+        }
       }
 
       if (ShowIndirectCallTargets) {
diff --git a/offload/plugins-nextgen/common/include/GlobalHandler.h b/offload/plugins-nextgen/common/include/GlobalHandler.h
index fc8d6fe384754..8f333fcf1056e 100644
--- a/offload/plugins-nextgen/common/include/GlobalHandler.h
+++ b/offload/plugins-nextgen/common/include/GlobalHandler.h
@@ -72,11 +72,15 @@ struct __llvm_profile_gpu_sections {
 };
 
 extern "C" {
-extern int __attribute__((weak)) __llvm_write_custom_profile(
-    const char *Target, const __llvm_profile_data *DataBegin,
-    const __llvm_profile_data *DataEnd, const char *CountersBegin,
-    const char *CountersEnd, const char *NamesBegin, const char *NamesEnd,
-    const uint64_t *VersionOverride);
+extern int __attribute__((weak))
+__llvm_write_custom_profile(const char *Target,
+                            const __llvm_profile_data *DataBegin,
+                            const __llvm_profile_data *DataEnd,
+                            const char *CountersBegin, const char *CountersEnd,
+                            const char *UniformCountersBegin,
+                            const char *UniformCountersEnd,
+                            const char *NamesBegin, const char *NamesEnd,
+                            const uint64_t *VersionOverride);
 }
 /// PGO profiling data extracted from a GPU device via __llvm_profile_sections.
 struct GPUProfGlobals {
diff --git a/offload/plugins-nextgen/common/src/GlobalHandler.cpp b/offload/plugins-nextgen/common/src/GlobalHandler.cpp
index 9216834b1e15e..9b7f9bfec3163 100644
--- a/offload/plugins-nextgen/common/src/GlobalHandler.cpp
+++ b/offload/plugins-nextgen/common/src/GlobalHandler.cpp
@@ -295,7 +295,9 @@ Error GPUProfGlobals::write() const {
       reinterpret_cast<const __llvm_profile_data *>(DataBegin),
       reinterpret_cast<const __llvm_profile_data *>(DataBegin +
                                                     DataSection.size()),
-      CountersBegin, CountersBegin + CountersSection.size(), NamesBegin,
+      CountersBegin, CountersBegin + CountersSection.size(),
+      /*UniformCountersBegin=*/nullptr,
+      /*UniformCountersEnd=*/nullptr, NamesBegin,
       NamesBegin + NamesSection.size(), &Version);
   if (Result != 0)
     return Plugin::error(ErrorCode::HOST_IO,



More information about the cfe-commits mailing list