[llvm-branch-commits] [llvm] [AMDGPU] Add FUNC_FULL_SIMD_MODE for object linking info flag (PR #207632)

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/207632

>From aa0358cb37ae5e286260e3f05cbf55c5176c397c Mon Sep 17 00:00:00 2001
From: Shilei Tian <i at tianshilei.me>
Date: Sat, 4 Jul 2026 18:12:47 -0400
Subject: [PATCH] [AMDGPU] Add FUNC_HALF_SIMD_MODE for object linking info flag

Add an INFO_FLAGS bit that is set when a function uses half SIMD mode,
where the waves of a workgroup can use only two of the four SIMDs. Today
this is CU mode on a target that supports WGP mode. If the bit is not
set, the function uses full SIMD mode.

Full SIMD mode is the default on every target: a gfx6-gfx9 CU has four
SIMDs, gfx10 and later targets use WGP mode by default, and gfx1250 is
always full SIMD, even in CU mode. So the bit is 0 for a default compile
on every target.
---
 llvm/docs/AMDGPUUsage.md                      |  3 +++
 .../llvm/Support/AMDGPUObjLinkingInfo.h       |  4 +++-
 llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp   |  1 +
 .../AMDGPU/AsmParser/AMDGPUAsmParser.cpp      |  2 ++
 .../MCTargetDesc/AMDGPUTargetStreamer.cpp     |  4 ++++
 .../MCTargetDesc/AMDGPUTargetStreamer.h       |  1 +
 .../CodeGen/AMDGPU/lds-link-time-codegen.ll   | 21 +++++++++++++++++++
 llvm/test/MC/AMDGPU/amdgpu-info-roundtrip.s   | 10 ++++-----
 8 files changed, 40 insertions(+), 6 deletions(-)

diff --git a/llvm/docs/AMDGPUUsage.md b/llvm/docs/AMDGPUUsage.md
index 00686c6ffc80f5..21013418f842f4 100644
--- a/llvm/docs/AMDGPUUsage.md
+++ b/llvm/docs/AMDGPUUsage.md
@@ -4523,6 +4523,9 @@ if needed.
   * - 0x4
     - `FUNC_HAS_DYN_STACK`
     - Function has dynamic stack allocation
+  * - 0x8
+    - `FUNC_HALF_SIMD_MODE`
+    - Function uses half SIMD execution mode; otherwise full SIMD mode
   :::
 
   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 19c4e2193bfe74..9f2db5033b61e5 100644
--- a/llvm/include/llvm/Support/AMDGPUObjLinkingInfo.h
+++ b/llvm/include/llvm/Support/AMDGPUObjLinkingInfo.h
@@ -68,7 +68,9 @@ enum class FuncInfoFlags : uint32_t {
   FUNC_USES_VCC = 1U << 0,
   FUNC_USES_FLAT_SCRATCH = 1U << 1,
   FUNC_HAS_DYN_STACK = 1U << 2,
-  LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/FUNC_HAS_DYN_STACK),
+  /// 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),
 };
 
 } // namespace AMDGPU
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
index 6a9c1c4caf7e2f..dc58920f24d51a 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
@@ -993,6 +993,7 @@ bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
          /*NumAccVGPR=*/static_cast<uint32_t>(RU.NumAGPR),
          /*PrivateSegmentSize=*/static_cast<uint32_t>(RU.PrivateSegmentSize),
          /*Occupancy=*/Occupancy,
+         /*UsesHalfSIMDMode=*/!AMDGPU::isFullSIMDMode(STM),
          /*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 72820355d55b1a..0a80578dbccbbf 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -6975,6 +6975,8 @@ bool AMDGPUAsmParser::ParseDirectiveAMDGPUInfo() {
       FI.UsesFlatScratch =
           !!(Flags & AMDGPU::FuncInfoFlags::FUNC_USES_FLAT_SCRATCH);
       FI.HasDynStack = !!(Flags & AMDGPU::FuncInfoFlags::FUNC_HAS_DYN_STACK);
+      FI.UsesHalfSIMDMode =
+          !!(Flags & AMDGPU::FuncInfoFlags::FUNC_HALF_SIMD_MODE);
       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 d4641b339786f4..86a849f036e00e 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp
@@ -671,6 +671,8 @@ void AMDGPUTargetAsmStreamer::emitAMDGPUInfo(
         Flags |= AMDGPU::FuncInfoFlags::FUNC_USES_FLAT_SCRATCH;
       if (Info->HasDynStack)
         Flags |= AMDGPU::FuncInfoFlags::FUNC_HAS_DYN_STACK;
+      if (Info->UsesHalfSIMDMode)
+        Flags |= AMDGPU::FuncInfoFlags::FUNC_HALF_SIMD_MODE;
       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';
@@ -1152,6 +1154,8 @@ void AMDGPUTargetELFStreamer::emitAMDGPUInfo(
         Flags |= AMDGPU::FuncInfoFlags::FUNC_USES_FLAT_SCRATCH;
       if (Info->HasDynStack)
         Flags |= AMDGPU::FuncInfoFlags::FUNC_HAS_DYN_STACK;
+      if (Info->UsesHalfSIMDMode)
+        Flags |= AMDGPU::FuncInfoFlags::FUNC_HALF_SIMD_MODE;
       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 34219028dca03e..43f277259dc2f9 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h
@@ -36,6 +36,7 @@ struct FuncInfo {
   uint32_t NumAccVGPR = 0;
   uint32_t PrivateSegmentSize = 0;
   uint32_t Occupancy = 0;
+  bool UsesHalfSIMDMode = 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 e7c0e7a562d752..f4bc469804d808 100644
--- a/llvm/test/CodeGen/AMDGPU/lds-link-time-codegen.ll
+++ b/llvm/test/CodeGen/AMDGPU/lds-link-time-codegen.ll
@@ -1,5 +1,11 @@
 ; 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=+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=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
 
 ; Test that with object linking enabled, external LDS declarations produce
 ; @abs32 at lo relocations, SHN_AMDGPU_LDS symbols, .amdgpu_lds directives,
@@ -38,6 +44,21 @@
 ; ASM-DAG:   .amdgpu_call device_func
 ; ASM-DAG: .end_amdgpu_info
 
+; 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.
+; FULL:      .amdgpu_info device_func
+; FULL-NEXT:   .amdgpu_flags 0
+; FULL:      .amdgpu_info test_kernel
+; FULL-NEXT:   .amdgpu_flags 0
+
+; HALF:      .amdgpu_info device_func
+; HALF-NEXT:   .amdgpu_flags 8
+; HALF:      .amdgpu_info test_kernel
+; HALF-NEXT:   .amdgpu_flags 8
+
 ; 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 47680c8a3470fd..a86fd03e7006be 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=7 (KERNEL|VCC|FLAT_SCRATCH), resources, call edge, use
-// COM: edge, indirect call, and type ID. Non-zero AGPR to verify conditional
-// COM: emission.
+// 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.
 	.amdgpu_info my_kernel
-		.amdgpu_flags 7
+		.amdgpu_flags 15
 		.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 7
+// ASM: .amdgpu_flags 15
 // ASM: .amdgpu_num_sgpr 33
 // ASM: .amdgpu_num_vgpr 32
 // ASM: .amdgpu_num_agpr 4



More information about the llvm-branch-commits mailing list