[llvm-branch-commits] [llvm] SelectionDAG: Use the physical register class for tied inline asm inputs (PR #238705)
Matt Arsenault via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Tue Nov 3 23:27:58 PST 2026
https://github.com/arsenm created https://github.com/llvm/llvm-project/pull/238705
When an inline asm input is tied to a physical register output, the
input's virtual register was created in the class for its value type.
That class is not required to contain the physical register, in which
case the input is materialized in the wrong bank and then copied into
the output register. On AMDGPU, an immediate tied to an AGPR output was
materialized in an SGPR, then copied to the AGPR, through a VGPR on
gfx908.
Use the minimal class of the physical register when the class for the
value type does not contain it. Untyped operands already did this
unconditionally; this extends it to typed operands. Keep the value
type's class if the physical register's minimal class is not
allocatable.
Co-authored-by: Claude Opus 5.5 <noreply at anthropic.com>
>From b61a0eeb23acd8c0abf30328f9b3aec634c6e57d Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Fri, 23 Oct 2026 07:58:14 -0700
Subject: [PATCH] SelectionDAG: Use the physical register class for tied inline
asm inputs
When an inline asm input is tied to a physical register output, the
input's virtual register was created in the class for its value type.
That class is not required to contain the physical register, in which
case the input is materialized in the wrong bank and then copied into
the output register. On AMDGPU, an immediate tied to an AGPR output was
materialized in an SGPR, then copied to the AGPR, through a VGPR on
gfx908.
Use the minimal class of the physical register when the class for the
value type does not contain it. Untyped operands already did this
unconditionally; this extends it to typed operands. Keep the value
type's class if the physical register's minimal class is not
allocatable.
Co-authored-by: Claude Opus 5.5 <noreply at anthropic.com>
---
.../SelectionDAG/SelectionDAGBuilder.cpp | 19 ++++++++++++++----
.../AMDGPU/inline-asm-tied-physreg-class.ll | 20 ++++++++-----------
2 files changed, 23 insertions(+), 16 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 2e35f199c7fba..21ce32906ef46 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -10700,10 +10700,21 @@ static bool prepareDAGLevelOperands(ConstraintDecisionInfo &Info,
auto *R = cast<RegisterSDNode>(Info.AsmNodeOperands[CurOp + 1]);
Register TiedReg = R->getReg();
MVT RegVT = R->getSimpleValueType(0);
- const TargetRegisterClass *RC =
- TiedReg.isVirtual() ? MRI.getRegClass(TiedReg)
- : RegVT != MVT::Untyped ? TLI.getRegClassFor(RegVT)
- : TRI.getMinimalPhysRegClass(TiedReg);
+ const TargetRegisterClass *RC;
+ if (TiedReg.isVirtual()) {
+ RC = MRI.getRegClass(TiedReg);
+ } else if (RegVT == MVT::Untyped) {
+ RC = TRI.getMinimalPhysRegClass(TiedReg);
+ } else {
+ RC = TLI.getRegClassFor(RegVT);
+ if (!RC->contains(TiedReg)) {
+ const TargetRegisterClass *PhysRC =
+ TRI.getMinimalPhysRegClass(TiedReg);
+ if (PhysRC->isAllocatable())
+ RC = PhysRC;
+ }
+ }
+
for (unsigned I = 0, E = Flag.getNumOperandRegisters(); I != E; ++I)
Regs.push_back(MRI.createVirtualRegister(RC));
diff --git a/llvm/test/CodeGen/AMDGPU/inline-asm-tied-physreg-class.ll b/llvm/test/CodeGen/AMDGPU/inline-asm-tied-physreg-class.ll
index 1461e33d15491..18c8411b56442 100644
--- a/llvm/test/CodeGen/AMDGPU/inline-asm-tied-physreg-class.ll
+++ b/llvm/test/CodeGen/AMDGPU/inline-asm-tied-physreg-class.ll
@@ -9,10 +9,7 @@ define void @tied_agpr_imm() {
; GFX908-LABEL: tied_agpr_imm:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: s_mov_b32 s4, 7
-; GFX908-NEXT: v_mov_b32_e32 v1, s4
-; GFX908-NEXT: s_nop 1
-; GFX908-NEXT: v_accvgpr_write_b32 a1, v1
+; GFX908-NEXT: v_accvgpr_write_b32 a1, 7
; GFX908-NEXT: ;;#ASMSTART
; GFX908-NEXT: ; def a1
; GFX908-NEXT: ;;#ASMEND
@@ -24,8 +21,7 @@ define void @tied_agpr_imm() {
; GFX90A-LABEL: tied_agpr_imm:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: s_mov_b32 s4, 7
-; GFX90A-NEXT: v_accvgpr_write_b32 a1, s4
+; GFX90A-NEXT: v_accvgpr_write_b32 a1, 7
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def a1
; GFX90A-NEXT: ;;#ASMEND
@@ -102,8 +98,8 @@ define void @tied_agpr64(i64 %x) {
; GFX908-LABEL: tied_agpr64:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_accvgpr_write_b32 a3, v1
; GFX908-NEXT: v_accvgpr_write_b32 a2, v0
+; GFX908-NEXT: v_accvgpr_write_b32 a3, v1
; GFX908-NEXT: ;;#ASMSTART
; GFX908-NEXT: ; def a[2:3]
; GFX908-NEXT: ;;#ASMEND
@@ -115,8 +111,8 @@ define void @tied_agpr64(i64 %x) {
; GFX90A-LABEL: tied_agpr64:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_accvgpr_write_b32 a3, v1
; GFX90A-NEXT: v_accvgpr_write_b32 a2, v0
+; GFX90A-NEXT: v_accvgpr_write_b32 a3, v1
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def a[2:3]
; GFX90A-NEXT: ;;#ASMEND
@@ -133,11 +129,11 @@ define void @tied_agpr64_sgpr(i64 inreg %x) {
; GFX908-LABEL: tied_agpr64_sgpr:
; GFX908: ; %bb.0:
; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: v_mov_b32_e32 v2, s16
-; GFX908-NEXT: v_mov_b32_e32 v0, s17
+; GFX908-NEXT: v_mov_b32_e32 v0, s16
+; GFX908-NEXT: v_mov_b32_e32 v1, s17
; GFX908-NEXT: s_nop 0
-; GFX908-NEXT: v_accvgpr_write_b32 a2, v2
-; GFX908-NEXT: v_accvgpr_write_b32 a3, v0
+; GFX908-NEXT: v_accvgpr_write_b32 a2, v0
+; GFX908-NEXT: v_accvgpr_write_b32 a3, v1
; GFX908-NEXT: ;;#ASMSTART
; GFX908-NEXT: ; def a[2:3]
; GFX908-NEXT: ;;#ASMEND
More information about the llvm-branch-commits
mailing list