[llvm-branch-commits] [llvm] [AMDGPU][SIInstrInfo] Materialise t16 PHI subreg operands (PR #222155)
Zach Goldthorpe via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Oct 30 07:55:42 PDT 2026
https://github.com/zGoldthorpe updated https://github.com/llvm/llvm-project/pull/222155
>From ef0344de2372e2728a9062733f82c7d7dc5b589b Mon Sep 17 00:00:00 2001
From: Zach Goldthorpe <Zach.Goldthorpe at amd.com>
Date: Tue, 8 Sep 2026 16:10:59 -0500
Subject: [PATCH 1/6] [AMDGPU][SIInstrInfo] Materialise t16 PHI subreg operands
---
llvm/lib/Target/AMDGPU/SIInstrInfo.cpp | 25 ++++++++++++++++++-------
1 file changed, 18 insertions(+), 7 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 43ee5330e770c..9e929630ba6e5 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -8267,9 +8267,7 @@ void SIInstrInfo::legalizeOperandsVALUt16(MachineInstr &MI, unsigned OpIdx,
unsigned Opcode = MI.getOpcode();
MachineBasicBlock *MBB = MI.getParent();
// Legalize operands and check for size mismatch
- if (OpIdx >= MI.getNumExplicitOperands() ||
- OpIdx >= get(Opcode).getNumOperands() ||
- get(Opcode).operands()[OpIdx].RegClass == -1)
+ if (OpIdx >= MI.getNumExplicitOperands())
return;
MachineOperand &Op = MI.getOperand(OpIdx);
@@ -8280,8 +8278,16 @@ void SIInstrInfo::legalizeOperandsVALUt16(MachineInstr &MI, unsigned OpIdx,
if (!RI.isVGPRClass(CurrRC))
return;
- int16_t RCID = getOpRegClassID(get(Opcode).operands()[OpIdx]);
- const TargetRegisterClass *ExpectedRC = RI.getRegClass(RCID);
+ const TargetRegisterClass *ExpectedRC;
+ if (MI.isPHI()) {
+ // A PHI is generic, so it carries no operand register classes.
+ ExpectedRC = MRI.getRegClass(MI.getOperand(0).getReg());
+ } else {
+ if (OpIdx >= get(Opcode).getNumOperands() ||
+ get(Opcode).operands()[OpIdx].RegClass == -1)
+ return;
+ ExpectedRC = RI.getRegClass(getOpRegClassID(get(Opcode).operands()[OpIdx]));
+ }
if (RI.getMatchingSuperRegClass(CurrRC, ExpectedRC, AMDGPU::lo16)) {
// Default to the lo16 only if the subregister is not specified.
if (Op.getSubReg() == AMDGPU::NoSubRegister)
@@ -8293,10 +8299,15 @@ void SIInstrInfo::legalizeOperandsVALUt16(MachineInstr &MI, unsigned OpIdx,
RI.getSubRegisterClass(CurrRC, Op.getSubReg());
if (RI.getMatchingSuperRegClass(ExpectedRC, CurrSRC, AMDGPU::lo16)) {
const DebugLoc &DL = MI.getDebugLoc();
+ MachineBasicBlock::iterator I = MI;
+ if (MI.isPHI()) {
+ MBB = MI.getOperand(OpIdx + 1).getMBB();
+ I = MBB->getFirstTerminator();
+ }
Register NewDstReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
Register Undef = MRI.createVirtualRegister(&AMDGPU::VGPR_16RegClass);
- BuildMI(*MBB, MI, DL, get(AMDGPU::IMPLICIT_DEF), Undef);
- BuildMI(*MBB, MI, DL, get(AMDGPU::REG_SEQUENCE), NewDstReg)
+ BuildMI(*MBB, I, DL, get(AMDGPU::IMPLICIT_DEF), Undef);
+ BuildMI(*MBB, I, DL, get(AMDGPU::REG_SEQUENCE), NewDstReg)
.addReg(Op.getReg(), {}, Op.getSubReg())
.addImm(AMDGPU::lo16)
.addReg(Undef)
>From 6f24c7022bb225fd0f799f18d2b95f7f40ea3d1e Mon Sep 17 00:00:00 2001
From: Zach Goldthorpe <Zach.Goldthorpe at amd.com>
Date: Wed, 9 Sep 2026 11:21:22 -0500
Subject: [PATCH 2/6] Refactor `ExpectedRC` determination
---
llvm/lib/Target/AMDGPU/SIInstrInfo.cpp | 30 +++++++++++++++++---------
1 file changed, 20 insertions(+), 10 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 9e929630ba6e5..f8f3768f49d8f 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -8278,16 +8278,26 @@ void SIInstrInfo::legalizeOperandsVALUt16(MachineInstr &MI, unsigned OpIdx,
if (!RI.isVGPRClass(CurrRC))
return;
- const TargetRegisterClass *ExpectedRC;
- if (MI.isPHI()) {
- // A PHI is generic, so it carries no operand register classes.
- ExpectedRC = MRI.getRegClass(MI.getOperand(0).getReg());
- } else {
- if (OpIdx >= get(Opcode).getNumOperands() ||
- get(Opcode).operands()[OpIdx].RegClass == -1)
- return;
- ExpectedRC = RI.getRegClass(getOpRegClassID(get(Opcode).operands()[OpIdx]));
- }
+ const TargetRegisterClass *ExpectedRC = [&]() -> const TargetRegisterClass * {
+ if (MI.isPHI()) {
+ // The expected register class of the PHI node operands is determined by
+ // its target operand.
+ return getOpRegClass(MI, 0);
+ }
+
+ if (OpIdx >= get(Opcode).getNumOperands())
+ return nullptr;
+
+ const MCOperandInfo &Operand = get(Opcode).operands()[OpIdx];
+ if (Operand.RegClass == -1)
+ return nullptr;
+
+ return RI.getRegClass(getOpRegClassID(Operand));
+ }();
+
+ if (!ExpectedRC)
+ return;
+
if (RI.getMatchingSuperRegClass(CurrRC, ExpectedRC, AMDGPU::lo16)) {
// Default to the lo16 only if the subregister is not specified.
if (Op.getSubReg() == AMDGPU::NoSubRegister)
>From 4367ae5fba23a6c065b1f1ad48ffedfbdcabdd6c Mon Sep 17 00:00:00 2001
From: Zach Goldthorpe <Zach.Goldthorpe at amd.com>
Date: Thu, 24 Sep 2026 15:46:27 -0500
Subject: [PATCH 3/6] Update lit test
(change was lost during rebase)
---
llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir | 9 +++++----
1 file changed, 5 insertions(+), 4 deletions(-)
diff --git a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir
index 0e6fd543b972c..db7287128a212 100644
--- a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir
+++ b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir
@@ -346,17 +346,18 @@ body: |
; GCN-NEXT: {{ $}}
; GCN-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
; GCN-NEXT: [[V_TRUNC_F16_t16_e64_:%[0-9]+]]:vgpr_16 = V_TRUNC_F16_t16_e64 0, [[DEF]].lo16, 0, 0, 0, implicit $mode, implicit $exec
- ; GCN-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY [[V_TRUNC_F16_t16_e64_]], implicit $exec
+ ; GCN-NEXT: [[DEF1:%[0-9]+]]:vgpr_16 = IMPLICIT_DEF
+ ; GCN-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vgpr_32 = REG_SEQUENCE [[V_TRUNC_F16_t16_e64_]], %subreg.lo16, [[DEF1]], %subreg.hi16
; GCN-NEXT: S_CBRANCH_SCC1 %bb.2, implicit $scc
; GCN-NEXT: {{ $}}
; GCN-NEXT: bb.1:
; GCN-NEXT: successors: %bb.2(0x80000000)
; GCN-NEXT: {{ $}}
- ; GCN-NEXT: [[DEF1:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
- ; GCN-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[DEF1]]
+ ; GCN-NEXT: [[DEF2:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
+ ; GCN-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY [[DEF2]]
; GCN-NEXT: {{ $}}
; GCN-NEXT: bb.2:
- ; GCN-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI [[COPY]], %bb.0, [[COPY1]], %bb.1
+ ; GCN-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI [[REG_SEQUENCE]], %bb.0, [[COPY]], %bb.1
; GCN-NEXT: S_ENDPGM 0, implicit [[PHI]]
bb.0:
liveins: $scc
>From 27ab67f2c440dc8e4db151f9ea506f66de1a0a68 Mon Sep 17 00:00:00 2001
From: Zach Goldthorpe <Zach.Goldthorpe at amd.com>
Date: Fri, 25 Sep 2026 10:16:26 -0500
Subject: [PATCH 4/6] Update lit tests
---
.../test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir | 10 ++++++----
1 file changed, 6 insertions(+), 4 deletions(-)
diff --git a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir
index db7287128a212..0c6885ac2a3b5 100644
--- a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir
+++ b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-f16-true16.mir
@@ -274,10 +274,11 @@ body: |
; GCN-NEXT: successors: %bb.2(0x80000000)
; GCN-NEXT: {{ $}}
; GCN-NEXT: [[V_TRUNC_F16_t16_e64_:%[0-9]+]]:vgpr_16 = V_TRUNC_F16_t16_e64 0, [[DEF]].lo16, 0, 0, 0, implicit $mode, implicit $exec
- ; GCN-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[V_TRUNC_F16_t16_e64_]], implicit $exec
+ ; GCN-NEXT: [[DEF2:%[0-9]+]]:vgpr_16 = IMPLICIT_DEF
+ ; GCN-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vgpr_32 = REG_SEQUENCE [[V_TRUNC_F16_t16_e64_]], %subreg.lo16, [[DEF2]], %subreg.hi16
; GCN-NEXT: {{ $}}
; GCN-NEXT: bb.2:
- ; GCN-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI [[COPY]], %bb.0, [[COPY1]], %bb.1
+ ; GCN-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI [[COPY]], %bb.0, [[REG_SEQUENCE]], %bb.1
; GCN-NEXT: S_ENDPGM 0, implicit [[PHI]]
bb.0:
liveins: $scc
@@ -314,10 +315,11 @@ body: |
; GCN-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[DEF]]
; GCN-NEXT: [[V_CVT_F32_F16_t16_e64_:%[0-9]+]]:vgpr_32 = V_CVT_F32_F16_t16_e64 0, [[COPY1]].hi16, 0, 0, 0, implicit $mode, implicit $exec
; GCN-NEXT: [[V_CVT_F16_F32_t16_e64_:%[0-9]+]]:vgpr_16 = V_CVT_F16_F32_t16_e64 0, [[V_CVT_F32_F16_t16_e64_]], 0, 0, 0, implicit $mode, implicit $exec
- ; GCN-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[V_CVT_F16_F32_t16_e64_]], implicit $exec
+ ; GCN-NEXT: [[DEF2:%[0-9]+]]:vgpr_16 = IMPLICIT_DEF
+ ; GCN-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vgpr_32 = REG_SEQUENCE [[V_CVT_F16_F32_t16_e64_]], %subreg.lo16, [[DEF2]], %subreg.hi16
; GCN-NEXT: {{ $}}
; GCN-NEXT: bb.2:
- ; GCN-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI [[COPY]], %bb.0, [[COPY2]], %bb.1
+ ; GCN-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI [[COPY]], %bb.0, [[REG_SEQUENCE]], %bb.1
; GCN-NEXT: S_ENDPGM 0, implicit [[PHI]]
bb.0:
liveins: $scc
>From 0eb85656c138ad8f38346c30d1960af9f4b092fb Mon Sep 17 00:00:00 2001
From: Zach Goldthorpe <Zach.Goldthorpe at amd.com>
Date: Mon, 28 Sep 2026 12:15:59 -0500
Subject: [PATCH 5/6] Resolve merge conflict in lit test
---
llvm/test/CodeGen/AMDGPU/v_swap_b16.ll | 12 ++++--------
1 file changed, 4 insertions(+), 8 deletions(-)
diff --git a/llvm/test/CodeGen/AMDGPU/v_swap_b16.ll b/llvm/test/CodeGen/AMDGPU/v_swap_b16.ll
index e4938c3449b2e..8c5de745402a8 100644
--- a/llvm/test/CodeGen/AMDGPU/v_swap_b16.ll
+++ b/llvm/test/CodeGen/AMDGPU/v_swap_b16.ll
@@ -8,14 +8,12 @@ define half @swap(half %a, half %b, i32 %i) {
; GFX11-TRUE16-LABEL: swap:
; GFX11-TRUE16: ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT: .LBB0_1: ; %loop
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v2, -1, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_swap_b16 v0.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l
+; GFX11-TRUE16-NEXT: v_swap_b16 v1.l, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0
@@ -51,14 +49,12 @@ define half @swap(half %a, half %b, i32 %i) {
; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0
; GFX12-TRUE16-NEXT: .LBB0_1: ; %loop
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: v_add_nc_u32_e32 v2, -1, v2
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX12-TRUE16-NEXT: v_swap_b16 v0.l, v0.h
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l
+; GFX12-TRUE16-NEXT: v_swap_b16 v1.l, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
; GFX12-TRUE16-NEXT: s_wait_alu depctr_sa_sdst(0)
>From 3f4497576ea7f8f4298d3ecd74b2301cc7342702 Mon Sep 17 00:00:00 2001
From: Zach Goldthorpe <Zach.Goldthorpe at amd.com>
Date: Thu, 29 Oct 2026 16:30:16 -0500
Subject: [PATCH 6/6] Update lit test after merge
---
llvm/test/CodeGen/AMDGPU/v_swap_b16.ll | 20 ++++++++++----------
1 file changed, 10 insertions(+), 10 deletions(-)
diff --git a/llvm/test/CodeGen/AMDGPU/v_swap_b16.ll b/llvm/test/CodeGen/AMDGPU/v_swap_b16.ll
index 8c5de745402a8..2ba7bc1f8b6c7 100644
--- a/llvm/test/CodeGen/AMDGPU/v_swap_b16.ll
+++ b/llvm/test/CodeGen/AMDGPU/v_swap_b16.ll
@@ -111,17 +111,16 @@ define half @swap_B(half %a, half %b, half %c, i32 %i) {
; GFX11-TRUE16-LABEL: swap_B:
; GFX11-TRUE16: ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT: .LBB1_1: ; %loop
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v3, -1, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l
; GFX11-TRUE16-NEXT: ;;#ASMSTART
-; GFX11-TRUE16-NEXT: ; use v0.h
+; GFX11-TRUE16-NEXT: ; use v0.l
; GFX11-TRUE16-NEXT: ;;#ASMEND
-; GFX11-TRUE16-NEXT: v_swap_b16 v0.l, v0.h
+; GFX11-TRUE16-NEXT: v_swap_b16 v1.l, v0.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
; GFX11-TRUE16-NEXT: ;;#ASMSTART
; GFX11-TRUE16-NEXT: ; use v1.l
@@ -136,6 +135,7 @@ define half @swap_B(half %a, half %b, half %c, i32 %i) {
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB1_1
; GFX11-TRUE16-NEXT: ; %bb.2: ; %ret
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.l
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: swap_B:
@@ -173,17 +173,16 @@ define half @swap_B(half %a, half %b, half %c, i32 %i) {
; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0
; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0
; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0
; GFX12-TRUE16-NEXT: .LBB1_1: ; %loop
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: v_add_nc_u32_e32 v3, -1, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l
; GFX12-TRUE16-NEXT: ;;#ASMSTART
-; GFX12-TRUE16-NEXT: ; use v0.h
+; GFX12-TRUE16-NEXT: ; use v0.l
; GFX12-TRUE16-NEXT: ;;#ASMEND
-; GFX12-TRUE16-NEXT: v_swap_b16 v0.l, v0.h
+; GFX12-TRUE16-NEXT: v_swap_b16 v1.l, v0.l
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
; GFX12-TRUE16-NEXT: ;;#ASMSTART
; GFX12-TRUE16-NEXT: ; use v1.l
@@ -199,6 +198,7 @@ define half @swap_B(half %a, half %b, half %c, i32 %i) {
; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB1_1
; GFX12-TRUE16-NEXT: ; %bb.2: ; %ret
; GFX12-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.l
; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX12-FAKE16-LABEL: swap_B:
More information about the llvm-branch-commits
mailing list