[llvm-branch-commits] [llvm] [AMDGPU] Add FUNC_WAVE32 for object linking info flag (PR #207633)

Shilei Tian via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Sat Nov 21 13:19:36 PST 2026


https://github.com/shiltian updated https://github.com/llvm/llvm-project/pull/207633

>From f577862c32c15d2d2d39a4039bd6bcf4d493c16f Mon Sep 17 00:00:00 2001
From: Shilei Tian <i at tianshilei.me>
Date: Sat, 4 Jul 2026 18:45:38 -0400
Subject: [PATCH] [AMDGPU] Add FUNC_ALTERNATE_WAVESIZE for object linking info
 flag

Add an INFO_FLAGS bit that is set when a function does not use the
default wave size of its target. The default wave size is wave32 if the
target supports wave32, and wave64 otherwise.

So the bit is 0 for a function that uses the default wave size, on
every target. Today only +wavefrontsize64 on a target that supports both
wave sizes sets it. The linker knows the target, so it can get the wave
size of each function from the bit.
---
 llvm/docs/AMDGPUUsage.md                      |  3 +++
 .../llvm/Support/AMDGPUObjLinkingInfo.h       |  6 ++++-
 llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp   |  4 ++++
 .../AMDGPU/AsmParser/AMDGPUAsmParser.cpp      |  2 ++
 .../MCTargetDesc/AMDGPUTargetStreamer.cpp     |  4 ++++
 .../MCTargetDesc/AMDGPUTargetStreamer.h       |  1 +
 .../CodeGen/AMDGPU/lds-link-time-codegen.ll   | 22 +++++++++++++++----
 llvm/test/MC/AMDGPU/amdgpu-info-roundtrip.s   | 10 ++++-----
 8 files changed, 42 insertions(+), 10 deletions(-)

diff --git a/llvm/docs/AMDGPUUsage.md b/llvm/docs/AMDGPUUsage.md
index 21013418f842f..7c16b5b9973e0 100644
--- a/llvm/docs/AMDGPUUsage.md
+++ b/llvm/docs/AMDGPUUsage.md
@@ -4526,6 +4526,9 @@ if needed.
   * - 0x8
     - `FUNC_HALF_SIMD_MODE`
     - Function uses half SIMD execution mode; otherwise full SIMD mode
+  * - 0x10
+    - `FUNC_ALTERNATE_WAVESIZE`
+    - Function uses the alternate wave size of the target; otherwise the default (wave32 if supported, else wave64)
   :::
 
   Symbol references (`INFO_FUNC`, `INFO_USE`, `INFO_CALL`) generate
diff --git a/llvm/include/llvm/Support/AMDGPUObjLinkingInfo.h b/llvm/include/llvm/Support/AMDGPUObjLinkingInfo.h
index 9f2db5033b61e..7baaa77d96fe8 100644
--- a/llvm/include/llvm/Support/AMDGPUObjLinkingInfo.h
+++ b/llvm/include/llvm/Support/AMDGPUObjLinkingInfo.h
@@ -70,7 +70,11 @@ enum class FuncInfoFlags : uint32_t {
   FUNC_HAS_DYN_STACK = 1U << 2,
   /// Function uses half SIMD mode. If unset, the function uses full SIMD mode.
   FUNC_HALF_SIMD_MODE = 1U << 3,
-  LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/FUNC_HALF_SIMD_MODE),
+  /// Function uses the alternate wave size of the target. If unset, the
+  /// function uses the default wave size of the target: wave32 if the target
+  /// supports wave32, and wave64 otherwise.
+  FUNC_ALTERNATE_WAVESIZE = 1U << 4,
+  LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/FUNC_ALTERNATE_WAVESIZE),
 };
 
 } // namespace AMDGPU
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
index dc58920f24d51..7d6145d037715 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
@@ -987,6 +987,9 @@ bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
         *ResourceUsage;
     uint32_t Occupancy = static_cast<uint32_t>(
         MF.getInfo<SIMachineFunctionInfo>()->getMinWavesPerEU());
+    // The default wave size of a target is wave32 if the target supports
+    // wave32, and wave64 otherwise.
+    unsigned DefaultWaveSize = STM.supportsWave32() ? 32 : 64;
     FunctionInfos.push_back(
         {/*NumSGPR=*/static_cast<uint32_t>(RU.NumExplicitSGPR),
          /*NumArchVGPR=*/static_cast<uint32_t>(RU.NumVGPR),
@@ -994,6 +997,7 @@ bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
          /*PrivateSegmentSize=*/static_cast<uint32_t>(RU.PrivateSegmentSize),
          /*Occupancy=*/Occupancy,
          /*UsesHalfSIMDMode=*/!AMDGPU::isFullSIMDMode(STM),
+         /*UsesAlternateWaveSize=*/STM.getWavefrontSize() != DefaultWaveSize,
          /*UsesVCC=*/RU.UsesVCC,
          /*UsesFlatScratch=*/RU.UsesFlatScratch,
          /*HasDynStack=*/RU.HasDynamicallySizedStack,
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 0a80578dbccbb..b10bb70d4d107 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -6977,6 +6977,8 @@ bool AMDGPUAsmParser::ParseDirectiveAMDGPUInfo() {
       FI.HasDynStack = !!(Flags & AMDGPU::FuncInfoFlags::FUNC_HAS_DYN_STACK);
       FI.UsesHalfSIMDMode =
           !!(Flags & AMDGPU::FuncInfoFlags::FUNC_HALF_SIMD_MODE);
+      FI.UsesAlternateWaveSize =
+          !!(Flags & AMDGPU::FuncInfoFlags::FUNC_ALTERNATE_WAVESIZE);
       HasScalarAttrs = true;
     } else if (Dir == "num_sgpr") {
       int64_t Val;
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp
index 86a849f036e00..d0439fb2e5ad0 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp
@@ -673,6 +673,8 @@ void AMDGPUTargetAsmStreamer::emitAMDGPUInfo(
         Flags |= AMDGPU::FuncInfoFlags::FUNC_HAS_DYN_STACK;
       if (Info->UsesHalfSIMDMode)
         Flags |= AMDGPU::FuncInfoFlags::FUNC_HALF_SIMD_MODE;
+      if (Info->UsesAlternateWaveSize)
+        Flags |= AMDGPU::FuncInfoFlags::FUNC_ALTERNATE_WAVESIZE;
       OS << "\t\t.amdgpu_flags " << llvm::to_underlying(Flags) << '\n';
       OS << "\t\t.amdgpu_num_sgpr " << Info->NumSGPR << '\n';
       OS << "\t\t.amdgpu_num_vgpr " << Info->NumArchVGPR << '\n';
@@ -1156,6 +1158,8 @@ void AMDGPUTargetELFStreamer::emitAMDGPUInfo(
         Flags |= AMDGPU::FuncInfoFlags::FUNC_HAS_DYN_STACK;
       if (Info->UsesHalfSIMDMode)
         Flags |= AMDGPU::FuncInfoFlags::FUNC_HALF_SIMD_MODE;
+      if (Info->UsesAlternateWaveSize)
+        Flags |= AMDGPU::FuncInfoFlags::FUNC_ALTERNATE_WAVESIZE;
       EmitU32Entry(AMDGPU::InfoKind::INFO_FLAGS, llvm::to_underlying(Flags));
       EmitU32Entry(AMDGPU::InfoKind::INFO_NUM_SGPR, Info->NumSGPR);
       EmitU32Entry(AMDGPU::InfoKind::INFO_NUM_VGPR, Info->NumArchVGPR);
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h
index 43f277259dc2f..ea12da71898f0 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h
@@ -37,6 +37,7 @@ struct FuncInfo {
   uint32_t PrivateSegmentSize = 0;
   uint32_t Occupancy = 0;
   bool UsesHalfSIMDMode = false;
+  bool UsesAlternateWaveSize = false;
   bool UsesVCC = false;
   bool UsesFlatScratch = false;
   bool HasDynStack = false;
diff --git a/llvm/test/CodeGen/AMDGPU/lds-link-time-codegen.ll b/llvm/test/CodeGen/AMDGPU/lds-link-time-codegen.ll
index f4bc469804d80..7bfb2beef67f1 100644
--- a/llvm/test/CodeGen/AMDGPU/lds-link-time-codegen.ll
+++ b/llvm/test/CodeGen/AMDGPU/lds-link-time-codegen.ll
@@ -1,9 +1,9 @@
 ; RUN: llc -mtriple=amdgpu9.00-amd-amdhsa -amdgpu-enable-object-linking < %s | FileCheck -check-prefixes=ASM %s --implicit-check-not=.amdgpu_num_agpr
 ; RUN: llc -mtriple=amdgpu9.00-amd-amdhsa -amdgpu-enable-object-linking -filetype=obj < %s | llvm-readobj -r --syms --sections - | FileCheck -check-prefixes=ELF %s
 ; RUN: llc -mtriple=amdgpu11.00-amd-amdhsa -amdgpu-enable-object-linking -filetype=asm < %s | FileCheck -check-prefixes=FULL %s
-; RUN: llc -mtriple=amdgpu11.00-amd-amdhsa -mattr=+wavefrontsize64 -amdgpu-enable-object-linking -filetype=asm < %s | FileCheck -check-prefixes=FULL %s
+; RUN: llc -mtriple=amdgpu11.00-amd-amdhsa -mattr=+wavefrontsize64 -amdgpu-enable-object-linking -filetype=asm < %s | FileCheck -check-prefixes=FULL-WAVE64 %s
 ; RUN: llc -mtriple=amdgpu11.00-amd-amdhsa -mattr=+cumode -amdgpu-enable-object-linking -filetype=asm < %s | FileCheck -check-prefixes=HALF %s
-; RUN: llc -mtriple=amdgpu11.00-amd-amdhsa -mattr=+cumode,+wavefrontsize64 -amdgpu-enable-object-linking -filetype=asm < %s | FileCheck -check-prefixes=HALF %s
+; RUN: llc -mtriple=amdgpu11.00-amd-amdhsa -mattr=+cumode,+wavefrontsize64 -amdgpu-enable-object-linking -filetype=asm < %s | FileCheck -check-prefixes=HALF-WAVE64 %s
 ; RUN: llc -mtriple=amdgpu12.50-amd-amdhsa -amdgpu-enable-object-linking -filetype=asm < %s | FileCheck -check-prefixes=FULL %s
 ; RUN: llc -mtriple=amdgpu12.50-amd-amdhsa -mattr=+cumode -amdgpu-enable-object-linking -filetype=asm < %s | FileCheck -check-prefixes=FULL %s
 
@@ -47,8 +47,10 @@
 ; COM: FUNC_HALF_SIMD_MODE (0x8): set when the function uses half SIMD mode
 ; COM: (two of the four SIMD32s). gfx11 default is full SIMD; +cumode selects
 ; COM: half SIMD and sets the flag. gfx1250 is always full SIMD, even with
-; COM: +cumode. Checked for both +wavefrontsize32 (default) and
-; COM: +wavefrontsize64 on gfx11.
+; COM: +cumode.
+; COM: FUNC_ALTERNATE_WAVESIZE (0x10): set when the function does not use the
+; COM: default wave size of the target. The default is wave32 on gfx11 and
+; COM: gfx1250, so only +wavefrontsize64 on gfx11 sets the flag.
 ; FULL:      .amdgpu_info device_func
 ; FULL-NEXT:   .amdgpu_flags 0
 ; FULL:      .amdgpu_info test_kernel
@@ -59,6 +61,18 @@
 ; HALF:      .amdgpu_info test_kernel
 ; HALF-NEXT:   .amdgpu_flags 8
 
+; COM: ALTERNATE_WAVESIZE only = 0x10 = 16.
+; FULL-WAVE64:      .amdgpu_info device_func
+; FULL-WAVE64-NEXT:   .amdgpu_flags 16
+; FULL-WAVE64:      .amdgpu_info test_kernel
+; FULL-WAVE64-NEXT:   .amdgpu_flags 16
+
+; COM: HALF_SIMD_MODE | ALTERNATE_WAVESIZE = 0x8 | 0x10 = 24.
+; HALF-WAVE64:      .amdgpu_info device_func
+; HALF-WAVE64-NEXT:   .amdgpu_flags 24
+; HALF-WAVE64:      .amdgpu_info test_kernel
+; HALF-WAVE64-NEXT:   .amdgpu_flags 24
+
 ; SHN_AMDGPU_LDS directives.
 ; ASM-DAG: .amdgpu_lds lds_large, 256, 16
 ; ASM-DAG: .amdgpu_lds lds_small, 128, 4
diff --git a/llvm/test/MC/AMDGPU/amdgpu-info-roundtrip.s b/llvm/test/MC/AMDGPU/amdgpu-info-roundtrip.s
index a86fd03e7006b..e6303ef5d99db 100644
--- a/llvm/test/MC/AMDGPU/amdgpu-info-roundtrip.s
+++ b/llvm/test/MC/AMDGPU/amdgpu-info-roundtrip.s
@@ -31,11 +31,11 @@ addr_taken_func:
 
 	.globl	extern_func
 
-// COM: Kernel: flags=15 (VCC|FLAT_SCRATCH|HAS_DYN_STACK|HALF_SIMD_MODE), resources,
-// COM: call edge, use edge, indirect call, and type ID. Non-zero AGPR to verify
-// COM: conditional emission.
+// COM: Kernel: flags=31 (VCC|FLAT_SCRATCH|HAS_DYN_STACK|HALF_SIMD_MODE|
+// COM: ALTERNATE_WAVESIZE), resources, call edge, use edge, indirect call, and
+// COM: type ID. Non-zero AGPR to verify conditional emission.
 	.amdgpu_info my_kernel
-		.amdgpu_flags 15
+		.amdgpu_flags 31
 		.amdgpu_num_sgpr 33
 		.amdgpu_num_vgpr 32
 		.amdgpu_num_agpr 4
@@ -69,7 +69,7 @@ addr_taken_func:
 	.end_amdgpu_info
 
 // ASM: .amdgpu_info my_kernel
-// ASM: .amdgpu_flags 15
+// ASM: .amdgpu_flags 31
 // ASM: .amdgpu_num_sgpr 33
 // ASM: .amdgpu_num_vgpr 32
 // ASM: .amdgpu_num_agpr 4



More information about the llvm-branch-commits mailing list