[llvm-branch-commits] [llvm] AMDGPU: Figure out required AGPR count for inline asm (PR #150910)
Matt Arsenault via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Tue Jul 29 08:27:09 PDT 2025
https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/150910
>From 2cef45d9731db40cc9297c5922f2b69c4d09ee62 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Mon, 28 Jul 2025 16:11:28 +0900
Subject: [PATCH] AMDGPU: Figure out required AGPR count for inline asm
For now just try to compute the minimum number of AGPRs required
to allocate the asm. Leave the attributor changes to turn this
into an integer value for later.
---
llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp | 59 +++++-
.../AMDGPU/amdgpu-attributor-no-agpr.ll | 199 ++++++++++++++++++
2 files changed, 251 insertions(+), 7 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
index 59cc1df292f46..1f5634a96a231 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
@@ -1200,16 +1200,61 @@ AAAMDWavesPerEU &AAAMDWavesPerEU::createForPosition(const IRPosition &IRP,
llvm_unreachable("AAAMDWavesPerEU is only valid for function position");
}
-static bool inlineAsmUsesAGPRs(const InlineAsm *IA) {
- for (const auto &CI : IA->ParseConstraints()) {
+/// Compute the minimum number of AGPRs required to allocate the inline asm.
+static unsigned inlineAsmGetNumRequiredAGPRs(const InlineAsm *IA,
+ const CallBase &Call) {
+ unsigned ArgNo = 0;
+ unsigned ResNo = 0;
+ unsigned AGPRDefCount = 0;
+ unsigned AGPRUseCount = 0;
+ unsigned MaxPhysReg = 0;
+ const DataLayout &DL = Call.getFunction()->getParent()->getDataLayout();
+
+ for (const InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) {
+ Type *Ty = nullptr;
+ switch (CI.Type) {
+ case InlineAsm::isOutput: {
+ Ty = Call.getType();
+ if (auto *STy = dyn_cast<StructType>(Ty))
+ Ty = STy->getElementType(ResNo);
+ ++ResNo;
+ break;
+ }
+ case InlineAsm::isInput: {
+ Ty = Call.getArgOperand(ArgNo++)->getType();
+ break;
+ }
+ case InlineAsm::isLabel:
+ continue;
+ case InlineAsm::isClobber:
+ // Parse the physical register reference.
+ break;
+ }
+
for (StringRef Code : CI.Codes) {
- Code.consume_front("{");
- if (Code.starts_with("a"))
- return true;
+ if (Code.starts_with("a")) {
+ // Virtual register, compute number of registers based on the type.
+ //
+ // We ought to be going through TargetLowering to get the number of
+ // registers, but we should avoid the dependence on CodeGen here.
+ unsigned RegCount = divideCeil(DL.getTypeSizeInBits(Ty), 32);
+ if (CI.Type == InlineAsm::isOutput) {
+ AGPRDefCount += RegCount;
+ if (CI.isEarlyClobber)
+ AGPRUseCount += RegCount;
+ } else
+ AGPRUseCount += RegCount;
+ } else {
+ // Physical register reference
+ auto [Kind, RegIdx, NumRegs] = AMDGPU::parseAsmConstraintPhysReg(Code);
+ if (Kind == 'a')
+ MaxPhysReg = std::max(MaxPhysReg, std::min(RegIdx + NumRegs, 256u));
+ }
}
}
- return false;
+ unsigned MaxVirtReg = std::max(AGPRUseCount, AGPRDefCount);
+ return std::min(MaxVirtReg + MaxPhysReg, 256u);
}
// TODO: Migrate to range merge of amdgpu-agpr-alloc.
@@ -1251,7 +1296,7 @@ struct AAAMDGPUNoAGPR
const Function *Callee = dyn_cast<Function>(CalleeOp);
if (!Callee) {
if (const InlineAsm *IA = dyn_cast<InlineAsm>(CalleeOp))
- return !inlineAsmUsesAGPRs(IA);
+ return inlineAsmGetNumRequiredAGPRs(IA, CB) == 0;
return false;
}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-no-agpr.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-no-agpr.ll
index 181dab8d4ca79..e502995cdb8ea 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-no-agpr.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-no-agpr.ll
@@ -251,6 +251,205 @@ define amdgpu_kernel void @indirect_calls_none_agpr(i1 %cond) {
ret void
}
+define amdgpu_kernel void @kernel_uses_asm_virtreg_def_struct_0() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_def_struct_0(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call { i32, i32 } asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call {i32, i32} asm sideeffect "; def $0", "=a,=a"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_virtreg_use_struct_1() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_use_struct_1(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call { i32, <2 x i32> } asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call {i32, <2 x i32>} asm sideeffect "; def $0", "=a,=a"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_virtreg_use_struct_2() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_use_struct_2(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call { i32, <2 x i32> } asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call {i32, <2 x i32>} asm sideeffect "; def $0", "=a,=v"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_virtreg_ptr_ty() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_ptr_ty(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: call void asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ call void asm sideeffect "; use $0", "a"(ptr poison)
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_virtreg_def_ptr_ty() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_def_ptr_ty(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call ptr asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call ptr asm sideeffect "; def $0", "=a"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_virtreg_def_vector_ptr_ty() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_def_vector_ptr_ty(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call <2 x ptr> asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call <2 x ptr> asm sideeffect "; def $0", "=a"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_physreg_def_struct_0() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_physreg_def_struct_0(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call { i32, i32 } asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call {i32, i32} asm sideeffect "; def $0", "={a0},={a[4:5]}"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_clobber() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_clobber(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: call void asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ call void asm sideeffect "; clobber $0", "~{a4}"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_clobber_tuple() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_clobber_tuple(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: call void asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ call void asm sideeffect "; clobber $0", "~{a[10:13]}"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_clobber_oob() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_clobber_oob(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: call void asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ call void asm sideeffect "; clobber $0", "~{a256}"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_clobber_max() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_clobber_max(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: call void asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ call void asm sideeffect "; clobber $0", "~{a255}"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_physreg_oob() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_physreg_oob(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: call void asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ call void asm sideeffect "; use $0", "{a256}"(i32 poison)
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_virtreg_def_max_ty() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_def_max_ty(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call <32 x i32> asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call <32 x i32> asm sideeffect "; def $0", "=a"()
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_virtreg_use_max_ty() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_use_max_ty(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: call void asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ call void asm sideeffect "; use $0", "a"(<32 x i32> poison)
+ ret void
+}
+
+define amdgpu_kernel void @kernel_uses_asm_virtreg_use_def_max_ty() {
+; CHECK-LABEL: define amdgpu_kernel void @kernel_uses_asm_virtreg_use_def_max_ty(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call <32 x i32> asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call <32 x i32> asm sideeffect "; use $0", "=a,a"(<32 x i32> poison)
+ ret void
+}
+
+define amdgpu_kernel void @vreg_use_exceeds_register_file() {
+; CHECK-LABEL: define amdgpu_kernel void @vreg_use_exceeds_register_file(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: call void asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ call void asm sideeffect "; use $0", "a"(<257 x i32> poison)
+ ret void
+}
+
+define amdgpu_kernel void @vreg_def_exceeds_register_file() {
+; CHECK-LABEL: define amdgpu_kernel void @vreg_def_exceeds_register_file(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call <257 x i32> asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call <257 x i32> asm sideeffect "; def $0", "=a"()
+ ret void
+}
+
+define amdgpu_kernel void @multiple() {
+; CHECK-LABEL: define amdgpu_kernel void @multiple(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call { <16 x i32>, <8 x i32>, <8 x i32> } asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call {<16 x i32>, <8 x i32>, <8 x i32>} asm sideeffect "; def $0", "=a,=a,=a,a,a,a"(<4 x i32> splat (i32 0), <8 x i32> splat (i32 1), i64 999)
+ ret void
+}
+
+define amdgpu_kernel void @earlyclobber_0() {
+; CHECK-LABEL: define amdgpu_kernel void @earlyclobber_0(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call <8 x i32> asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call <8 x i32> asm sideeffect "; def $0", "=&a,a"(i32 0)
+ ret void
+}
+
+define amdgpu_kernel void @earlyclobber_1() {
+; CHECK-LABEL: define amdgpu_kernel void @earlyclobber_1(
+; CHECK-SAME: ) #[[ATTR0]] {
+; CHECK-NEXT: [[DEF:%.*]] = call { <8 x i32>, <16 x i32> } asm sideeffect "
+; CHECK-NEXT: ret void
+;
+ %def = call { <8 x i32>, <16 x i32 > } asm sideeffect "; def $0, $1", "=&a,=&a,a,a"(i32 0, <16 x i32> splat (i32 1))
+ ret void
+}
attributes #0 = { "amdgpu-agpr-alloc"="0" }
;.
More information about the llvm-branch-commits
mailing list