Address review comments: Scalarize v2s16 for uniform operation

Author: Abhinav Garg

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeHelper.cpp

@@ -616,6 +616,24 @@ void RegBankLegalizeHelper::lowerSplitTo32(MachineInstr &MI) {
   MI.eraseFromParent();
 }
 
+void RegBankLegalizeHelper::lowerSplitTo16(MachineInstr &MI) {
+  Register Dst = MI.getOperand(0).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  assert(DstTy == V2S16);
+  auto [Val0_Lo_32, Val0_Hi_32] = unpackAExt(MI.getOperand(1).getReg());
+  auto [Val1_Lo_32, Val1_Hi_32] = unpackAExt(MI.getOperand(2).getReg());
+  unsigned Opc = MI.getOpcode();
+  auto Flags = MI.getFlags();
+  auto Val0_Lo = B.buildTrunc(SgprRB_S16, Val0_Lo_32);
+  auto Val0_Hi = B.buildTrunc(SgprRB_S16, Val0_Hi_32);
+  auto Val1_Lo = B.buildTrunc(SgprRB_S16, Val1_Lo_32);
+  auto Val1_Hi = B.buildTrunc(SgprRB_S16, Val1_Hi_32);
+  auto Lo = B.buildInstr(Opc, {SgprRB_S16}, {Val0_Lo, Val1_Lo}, Flags);
+  auto Hi = B.buildInstr(Opc, {SgprRB_S16}, {Val0_Hi, Val1_Hi}, Flags);
+  B.buildMergeLikeInstr(Dst, {Lo.getReg(0), Hi.getReg(0)});
+  MI.eraseFromParent();
+}
+
 void RegBankLegalizeHelper::lowerSplitTo32Select(MachineInstr &MI) {
   Register Dst = MI.getOperand(0).getReg();
   LLT DstTy = MRI.getType(Dst);
@@ -688,6 +706,8 @@ void RegBankLegalizeHelper::lower(MachineInstr &MI,
     return lowerUnpackBitShift(MI);
   case UnpackMinMax:
     return lowerUnpackMinMax(MI);
+  case ScalarizeToS16:
+    return lowerSplitTo16(MI);
   case Ext32To64: {
     const RegisterBank *RB = MRI.getRegBank(MI.getOperand(0).getReg());
     MachineInstrBuilder Hi;

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeHelper.h

@@ -72,6 +72,7 @@ class RegBankLegalizeHelper {
   static constexpr LLT P6 = LLT::pointer(6, 32);
 
   MachineRegisterInfo::VRegAttrs SgprRB_S32 = {SgprRB, S32};
+  MachineRegisterInfo::VRegAttrs SgprRB_S16 = {SgprRB, S16};
   MachineRegisterInfo::VRegAttrs VgprRB_S32 = {VgprRB, S32};
   MachineRegisterInfo::VRegAttrs VccRB_S1 = {VccRB, S1};
 
@@ -121,6 +122,7 @@ class RegBankLegalizeHelper {
   void lowerV_BFE(MachineInstr &MI);
   void lowerS_BFE(MachineInstr &MI);
   void lowerSplitTo32(MachineInstr &MI);
+  void lowerSplitTo16(MachineInstr &MI);
   void lowerSplitTo32Select(MachineInstr &MI);
   void lowerSplitTo32SExtInReg(MachineInstr &MI);
   void lowerUnpackMinMax(MachineInstr &MI);

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp

@@ -914,7 +914,9 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
       .Div(S32, {{Vgpr32}, {Vgpr32, Vgpr32}})
       .Uni(S64, {{UniInVgprS64}, {Vgpr64, Vgpr64}})
       .Div(S64, {{Vgpr64}, {Vgpr64, Vgpr64}})
-      .Uni(V2S16, {{UniInVgprV2S16}, {VgprV2S16, VgprV2S16}})
+      .Uni(V2S16, {{UniInVgprV2S16}, {VgprV2S16, VgprV2S16}}, !hasSALUFloat)
+      .Uni(V2S16, {{SgprV2S16}, {SgprV2S16, SgprV2S16}, ScalarizeToS16},
+           hasSALUFloat)
       .Div(V2S16, {{VgprV2S16}, {VgprV2S16, VgprV2S16}})
       .Any({{UniV2S32}, {{UniInVgprV2S32}, {VgprV2S32, VgprV2S32}}})
       .Any({{DivV2S32}, {{VgprV2S32}, {VgprV2S32, VgprV2S32}}});

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.h

@@ -221,6 +221,7 @@ enum LoweringMethodID {
   V_BFE,
   VgprToVccCopy,
   SplitTo32,
+  ScalarizeToS16,
   SplitTo32Select,
   SplitTo32SExtInReg,
   Ext32To64,

llvm/test/CodeGen/AMDGPU/GlobalISel/fadd.ll

@@ -1,90 +1,67 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
-; RUN: llc -mtriple=amdgcn-amd-amdpal -mattr=-real-true16 -mcpu=gfx1100 -o - %s | FileCheck -check-prefixes=GCN,GFX11,GFX11-SDAG,GFX11-SDAG-FAKE16 %s
-; RUN: llc -mtriple=amdgcn-amd-amdpal -mattr=+real-true16 -mcpu=gfx1100 -o - %s | FileCheck -check-prefixes=GCN,GFX11,GFX11-SDAG,GFX11-SDAG-TRUE16 %s
-; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=-real-true16 -mcpu=gfx1100 -o - %s | FileCheck -check-prefixes=GCN,GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s
-; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=+real-true16 -mcpu=gfx1100 -o - %s | FileCheck -check-prefixes=GCN,GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s
-; RUN: llc -mtriple=amdgcn-amd-amdpal -mattr=-real-true16 -mcpu=gfx1200 -o - %s | FileCheck -check-prefixes=GCN,GFX12,GFX12-SDAG,GFX12-SDAG-FAKE16 %s
-; RUN: llc -mtriple=amdgcn-amd-amdpal -mattr=+real-true16 -mcpu=gfx1200 -o - %s | FileCheck -check-prefixes=GCN,GFX12,GFX12-SDAG,GFX12-SDAG-TRUE16 %s
-; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=-real-true16 -mcpu=gfx1200 -o - %s | FileCheck -check-prefixes=GCN,GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s
-; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=+real-true16 -mcpu=gfx1200 -o - %s | FileCheck -check-prefixes=GCN,GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=-real-true16 -mcpu=gfx1100 -o - %s | FileCheck -check-prefixes=GCN,GFX11,GFX11-FAKE16 %s
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=+real-true16 -mcpu=gfx1100 -o - %s | FileCheck -check-prefixes=GCN,GFX11,GFX11-TRUE16 %s
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=-real-true16 -mcpu=gfx1200 -o - %s | FileCheck -check-prefixes=GCN,GFX12,GFX12-FAKE16 %s
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=+real-true16 -mcpu=gfx1200 -o - %s | FileCheck -check-prefixes=GCN,GFX12,GFX12-TRUE16 %s
 
 define amdgpu_ps half @fadd_s16_uniform(half inreg %a, half inreg %b) {
-; GFX11-SDAG-FAKE16-LABEL: fadd_s16_uniform:
-; GFX11-SDAG-FAKE16:       ; %bb.0:
-; GFX11-SDAG-FAKE16-NEXT:    v_add_f16_e64 v0, s0, s1
-; GFX11-SDAG-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-SDAG-TRUE16-LABEL: fadd_s16_uniform:
-; GFX11-SDAG-TRUE16:       ; %bb.0:
-; GFX11-SDAG-TRUE16-NEXT:    v_add_f16_e64 v0.l, s0, s1
-; GFX11-SDAG-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-GISEL-FAKE16-LABEL: fadd_s16_uniform:
-; GFX11-GISEL-FAKE16:       ; %bb.0:
-; GFX11-GISEL-FAKE16-NEXT:    v_add_f16_e64 v0, s0, s1
-; GFX11-GISEL-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-GISEL-FAKE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11-GISEL-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-GISEL-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-GISEL-TRUE16-LABEL: fadd_s16_uniform:
-; GFX11-GISEL-TRUE16:       ; %bb.0:
-; GFX11-GISEL-TRUE16-NEXT:    v_add_f16_e64 v0.l, s0, s1
-; GFX11-GISEL-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-GISEL-TRUE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11-GISEL-TRUE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-GISEL-TRUE16-NEXT:    ; return to shader part epilog
+; GFX11-FAKE16-LABEL: fadd_s16_uniform:
+; GFX11-FAKE16:       ; %bb.0:
+; GFX11-FAKE16-NEXT:    v_add_f16_e64 v0, s0, s1
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-FAKE16-NEXT:    ; return to shader part epilog
+;
+; GFX11-TRUE16-LABEL: fadd_s16_uniform:
+; GFX11-TRUE16:       ; %bb.0:
+; GFX11-TRUE16-NEXT:    v_add_f16_e64 v0.l, s0, s1
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-TRUE16-NEXT:    ; return to shader part epilog
 ;
 ; GFX12-LABEL: fadd_s16_uniform:
 ; GFX12:       ; %bb.0:
 ; GFX12-NEXT:    s_add_f16 s0, s0, s1
 ; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_3)
 ; GFX12-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX12-NEXT:    ; return to shader part epilog
+; -LABEL: fadd_s16_uniform:
+; :       ; %bb.0:
+; -NEXT:    v_add_f16_e64 v0.l, s0, s1
+; -NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; -NEXT:    v_readfirstlane_b32 s0, v0
+; -NEXT:    v_mov_b32_e32 v0, s0
+; -NEXT:    ; return to shader part epilog
   %fadd = fadd half %a, %b
   ret half %fadd
 }
 
 define amdgpu_ps half @fadd_s16_div(half %a, half %b) {
-; GFX11-SDAG-FAKE16-LABEL: fadd_s16_div:
-; GFX11-SDAG-FAKE16:       ; %bb.0:
-; GFX11-SDAG-FAKE16-NEXT:    v_add_f16_e32 v0, v0, v1
-; GFX11-SDAG-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-SDAG-TRUE16-LABEL: fadd_s16_div:
-; GFX11-SDAG-TRUE16:       ; %bb.0:
-; GFX11-SDAG-TRUE16-NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
-; GFX11-SDAG-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-GISEL-FAKE16-LABEL: fadd_s16_div:
-; GFX11-GISEL-FAKE16:       ; %bb.0:
-; GFX11-GISEL-FAKE16-NEXT:    v_add_f16_e32 v0, v0, v1
-; GFX11-GISEL-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-GISEL-TRUE16-LABEL: fadd_s16_div:
-; GFX11-GISEL-TRUE16:       ; %bb.0:
-; GFX11-GISEL-TRUE16-NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
-; GFX11-GISEL-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-SDAG-FAKE16-LABEL: fadd_s16_div:
-; GFX12-SDAG-FAKE16:       ; %bb.0:
-; GFX12-SDAG-FAKE16-NEXT:    v_add_f16_e32 v0, v0, v1
-; GFX12-SDAG-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-SDAG-TRUE16-LABEL: fadd_s16_div:
-; GFX12-SDAG-TRUE16:       ; %bb.0:
-; GFX12-SDAG-TRUE16-NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
-; GFX12-SDAG-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-GISEL-FAKE16-LABEL: fadd_s16_div:
-; GFX12-GISEL-FAKE16:       ; %bb.0:
-; GFX12-GISEL-FAKE16-NEXT:    v_add_f16_e32 v0, v0, v1
-; GFX12-GISEL-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-GISEL-TRUE16-LABEL: fadd_s16_div:
-; GFX12-GISEL-TRUE16:       ; %bb.0:
-; GFX12-GISEL-TRUE16-NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
-; GFX12-GISEL-TRUE16-NEXT:    ; return to shader part epilog
+; GFX11-FAKE16-LABEL: fadd_s16_div:
+; GFX11-FAKE16:       ; %bb.0:
+; GFX11-FAKE16-NEXT:    v_add_f16_e32 v0, v0, v1
+; GFX11-FAKE16-NEXT:    ; return to shader part epilog
+;
+; GFX11-TRUE16-LABEL: fadd_s16_div:
+; GFX11-TRUE16:       ; %bb.0:
+; GFX11-TRUE16-NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT:    ; return to shader part epilog
+;
+; GFX12-FAKE16-LABEL: fadd_s16_div:
+; GFX12-FAKE16:       ; %bb.0:
+; GFX12-FAKE16-NEXT:    v_add_f16_e32 v0, v0, v1
+; GFX12-FAKE16-NEXT:    ; return to shader part epilog
+;
+; GFX12-TRUE16-LABEL: fadd_s16_div:
+; GFX12-TRUE16:       ; %bb.0:
+; GFX12-TRUE16-NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
+; GFX12-TRUE16-NEXT:    ; return to shader part epilog
+; -LABEL: fadd_s16_div:
+; :       ; %bb.0:
+; -NEXT:    v_add_f16_e32 v0.l, v0.l, v1.l
+; -NEXT:    ; return to shader part epilog
   %fadd = fadd half %a, %b
   ret half %fadd
 }
@@ -155,92 +132,58 @@ define amdgpu_ps double @fadd_s64_div(double %a, double %b) {
   ret double %fadd
 }
 
-define <2 x half> @fadd_v2s16_uniform(<2 x half> inreg %a, <2 x half> inreg %b) {
+define amdgpu_ps <2 x half> @fadd_v2s16_uniform(<2 x half> inreg %a, <2 x half> inreg %b) {
 ; GFX11-LABEL: fadd_v2s16_uniform:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    v_pk_add_f16 v0, s0, s1
-; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:    ; return to shader part epilog
 ;
 ; GFX12-LABEL: fadd_v2s16_uniform:
 ; GFX12:       ; %bb.0:
-; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT:    s_wait_expcnt 0x0
-; GFX12-NEXT:    s_wait_samplecnt 0x0
-; GFX12-NEXT:    s_wait_bvhcnt 0x0
-; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    v_pk_add_f16 v0, s0, s1
-; GFX12-NEXT:    s_setpc_b64 s[30:31]
+; GFX12-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX12-NEXT:    s_lshr_b32 s3, s1, 16
+; GFX12-NEXT:    s_add_f16 s0, s0, s1
+; GFX12-NEXT:    s_add_f16 s1, s2, s3
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_3) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
+; GFX12-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-NEXT:    ; return to shader part epilog
   %fadd = fadd <2 x half> %a, %b
   ret <2 x half> %fadd
 }
 
-define <2 x half> @fadd_v2s16_div(<2 x half> %a, <2 x half> %b) {
-; GFX11-LABEL: fadd_v2s16_div:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    v_pk_add_f16 v0, v0, v1
-; GFX11-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: fadd_v2s16_div:
-; GFX12:       ; %bb.0:
-; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT:    s_wait_expcnt 0x0
-; GFX12-NEXT:    s_wait_samplecnt 0x0
-; GFX12-NEXT:    s_wait_bvhcnt 0x0
-; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    v_pk_add_f16 v0, v0, v1
-; GFX12-NEXT:    s_setpc_b64 s[30:31]
+define amdgpu_ps <2 x half> @fadd_v2s16_div(<2 x half> %a, <2 x half> %b) {
+; GCN-LABEL: fadd_v2s16_div:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    v_pk_add_f16 v0, v0, v1
+; GCN-NEXT:    ; return to shader part epilog
   %fadd = fadd <2 x half> %a, %b
   ret <2 x half> %fadd
 }
 
-define <2 x float> @fadd_v2s32_uniform(<2 x float> inreg %a, <2 x float> inreg %b) {
+define amdgpu_ps <2 x float> @fadd_v2s32_uniform(<2 x float> inreg %a, <2 x float> inreg %b) {
 ; GFX11-LABEL: fadd_v2s32_uniform:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    v_add_f32_e64 v0, s0, s2
 ; GFX11-NEXT:    v_add_f32_e64 v1, s1, s3
-; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:    ; return to shader part epilog
 ;
 ; GFX12-LABEL: fadd_v2s32_uniform:
 ; GFX12:       ; %bb.0:
-; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT:    s_wait_expcnt 0x0
-; GFX12-NEXT:    s_wait_samplecnt 0x0
-; GFX12-NEXT:    s_wait_bvhcnt 0x0
-; GFX12-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-NEXT:    s_add_f32 s0, s0, s2
 ; GFX12-NEXT:    s_add_f32 s1, s1, s3
-; GFX12-NEXT:    s_wait_alu 0xfffe
-; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_2)
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_3)
 ; GFX12-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX12-NEXT:    s_setpc_b64 s[30:31]
+; GFX12-NEXT:    ; return to shader part epilog
   %fadd = fadd <2 x float> %a, %b
   ret <2 x float> %fadd
 }
 
-define <2 x float> @fadd_v2s32_div(<2 x float> %a, <2 x float> %b) {
-; GFX11-LABEL: fadd_v2s32_div:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    v_dual_add_f32 v0, v0, v2 :: v_dual_add_f32 v1, v1, v3
-; GFX11-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX12-LABEL: fadd_v2s32_div:
-; GFX12:       ; %bb.0:
-; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
-; GFX12-NEXT:    s_wait_expcnt 0x0
-; GFX12-NEXT:    s_wait_samplecnt 0x0
-; GFX12-NEXT:    s_wait_bvhcnt 0x0
-; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    v_dual_add_f32 v0, v0, v2 :: v_dual_add_f32 v1, v1, v3
-; GFX12-NEXT:    s_setpc_b64 s[30:31]
+define amdgpu_ps <2 x float> @fadd_v2s32_div(<2 x float> %a, <2 x float> %b) {
+; GCN-LABEL: fadd_v2s32_div:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    v_dual_add_f32 v0, v0, v2 :: v_dual_add_f32 v1, v1, v3
+; GCN-NEXT:    ; return to shader part epilog
   %fadd = fadd <2 x float> %a, %b
   ret <2 x float> %fadd
 }
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; GFX11-GISEL: {{.*}}
-; GFX11-SDAG: {{.*}}
-; GFX12-GISEL: {{.*}}
-; GFX12-SDAG: {{.*}}