[llvm] [AMDGPU] Force the third source operand of the MAI instructions to VGPR if no AGPRs are used. (PR #69720)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Oct 20 06:03:26 PDT 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: None (alex-t)
<details>
<summary>Changes</summary>
eaf85b9c28 "[AMDGPU] Select VGPR versions of MFMA if possible" prevents the compiler from reserving AGPRs if a kernel has no inline asm explicitly using AGPRs, no calls, and runs at least 2 waves with not more than 256 VGPRs. This, in turn, makes it impossible to allocate AGPR if necessary. As a result, regalloc fails in case we have an MAI instruction that has at least one AGPR operand.
This change checks if we have AGPRs and forces operands to VGPR if we do not have them.
---
Full diff: https://github.com/llvm/llvm-project/pull/69720.diff
2 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+21-13)
- (added) llvm/test/CodeGen/AMDGPU/smfmac_no_agprs.ll (+53)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index a7d8e11461733fc..99569c8071bebee 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -14136,7 +14136,9 @@ void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
SDNode *Node) const {
const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
- MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
+ MachineFunction *MF = MI.getParent()->getParent();
+ MachineRegisterInfo &MRI = MF->getRegInfo();
+ SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();
if (TII->isVOP3(MI.getOpcode())) {
// Make sure constant bus requirements are respected.
@@ -14147,9 +14149,14 @@ void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
// use between vgpr and agpr as agpr tuples tend to be big.
if (!MI.getDesc().operands().empty()) {
unsigned Opc = MI.getOpcode();
+ bool NoAGPRs = !Info->mayNeedAGPRs();
+ SmallVector<int> Opnds;
+ Opnds.push_back(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0));
+ Opnds.push_back(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1));
+ if (NoAGPRs)
+ Opnds.push_back(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2));
const SIRegisterInfo *TRI = Subtarget->getRegisterInfo();
- for (auto I : { AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0),
- AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1) }) {
+ for (auto I : Opnds) {
if (I == -1)
break;
MachineOperand &Op = MI.getOperand(I);
@@ -14169,18 +14176,19 @@ void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
MRI.setRegClass(Op.getReg(), NewRC);
}
- // Resolve the rest of AV operands to AGPRs.
- if (auto *Src2 = TII->getNamedOperand(MI, AMDGPU::OpName::src2)) {
- if (Src2->isReg() && Src2->getReg().isVirtual()) {
- auto *RC = TRI->getRegClassForReg(MRI, Src2->getReg());
- if (TRI->isVectorSuperClass(RC)) {
- auto *NewRC = TRI->getEquivalentAGPRClass(RC);
- MRI.setRegClass(Src2->getReg(), NewRC);
- if (Src2->isTied())
- MRI.setRegClass(MI.getOperand(0).getReg(), NewRC);
+ if (!NoAGPRs)
+ // Resolve the rest of AV operands to AGPRs.
+ if (auto *Src2 = TII->getNamedOperand(MI, AMDGPU::OpName::src2)) {
+ if (Src2->isReg() && Src2->getReg().isVirtual()) {
+ auto *RC = TRI->getRegClassForReg(MRI, Src2->getReg());
+ if (TRI->isVectorSuperClass(RC)) {
+ auto *NewRC = TRI->getEquivalentAGPRClass(RC);
+ MRI.setRegClass(Src2->getReg(), NewRC);
+ if (Src2->isTied())
+ MRI.setRegClass(MI.getOperand(0).getReg(), NewRC);
+ }
}
}
- }
}
return;
diff --git a/llvm/test/CodeGen/AMDGPU/smfmac_no_agprs.ll b/llvm/test/CodeGen/AMDGPU/smfmac_no_agprs.ll
new file mode 100644
index 000000000000000..a8bd0afe3ce16fd
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/smfmac_no_agprs.ll
@@ -0,0 +1,53 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx940 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=MI300 %s
+
+
+define protected amdgpu_kernel void @test(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; MI300-LABEL: test:
+; MI300: ; %bb.0: ; %entry
+; MI300-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x0
+; MI300-NEXT: v_mov_b32_e32 v0, 0
+; MI300-NEXT: v_mov_b32_e32 v2, v0
+; MI300-NEXT: v_mov_b32_e32 v3, v0
+; MI300-NEXT: v_mov_b32_e32 v1, v0
+; MI300-NEXT: s_waitcnt lgkmcnt(0)
+; MI300-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x0
+; MI300-NEXT: v_mov_b64_e32 v[10:11], v[2:3]
+; MI300-NEXT: v_mov_b64_e32 v[8:9], v[0:1]
+; MI300-NEXT: s_waitcnt lgkmcnt(0)
+; MI300-NEXT: v_mov_b32_e32 v12, s4
+; MI300-NEXT: v_mov_b32_e32 v13, s5
+; MI300-NEXT: v_mov_b32_e32 v4, s6
+; MI300-NEXT: v_mov_b32_e32 v5, s7
+; MI300-NEXT: v_mov_b32_e32 v6, s7
+; MI300-NEXT: v_mov_b32_e32 v7, s7
+; MI300-NEXT: s_nop 1
+; MI300-NEXT: v_smfmac_i32_16x16x64_i8 v[8:11], v[12:13], v[4:7], v13
+; MI300-NEXT: s_nop 6
+; MI300-NEXT: global_store_dword v0, v11, s[2:3] offset:12 sc0 sc1
+; MI300-NEXT: s_endpgm
+entry:
+ %arrayidx = getelementptr inbounds i32, ptr addrspace(1) %in, i64 0
+ %arrayidx1 = getelementptr inbounds i32, ptr addrspace(1) %in, i64 1
+ %arrayidx2 = getelementptr inbounds i32, ptr addrspace(1) %in, i64 2
+ %arrayidx3 = getelementptr inbounds i32, ptr addrspace(1) %in, i64 3
+ %0 = load i32, ptr addrspace(1) %arrayidx
+ %1 = load i32, ptr addrspace(1) %arrayidx1
+ %2 = load i32, ptr addrspace(1) %arrayidx2
+ %3 = load i32, ptr addrspace(1) %arrayidx3
+ %src1.0 = insertelement <2 x i32> undef, i32 %0, i64 0
+ %src1 = insertelement <2 x i32> %src1.0, i32 %1, i64 1
+ %src2.0 = insertelement <4 x i32> undef, i32 %2, i64 0
+ %src2.1 = insertelement <4 x i32> %src2.0, i32 %3, i64 1
+ %src2.2 = insertelement <4 x i32> %src2.1, i32 %3, i64 2
+ %src2 = insertelement <4 x i32> %src2.2, i32 %3, i64 3
+ %4 = tail call <4 x i32> @llvm.amdgcn.smfmac.i32.16x16x64.i8(<2 x i32> %src1, <4 x i32> %src2, <4 x i32> zeroinitializer, i32 %1, i32 0, i32 0)
+ %vecext = extractelement <4 x i32> %4, i64 0
+ %vecext.1 = extractelement <4 x i32> %4, i64 1
+ %vecext.2 = extractelement <4 x i32> %4, i64 2
+ %vecext.3 = extractelement <4 x i32> %4, i64 3
+ %arrayidx4 = getelementptr inbounds i32, ptr addrspace(1) %out, i64 3
+ store i32 %vecext.3, ptr addrspace(1) %arrayidx4
+ ret void
+}
+declare <4 x i32> @llvm.amdgcn.smfmac.i32.16x16x64.i8(<2 x i32>, <4 x i32>, <4 x i32>, i32, i32 immarg, i32 immarg)
``````````
</details>
https://github.com/llvm/llvm-project/pull/69720
More information about the llvm-commits
mailing list