[AMDGPU] Support V_PK_MIN3/MAX3_NUM_F16 on gfx1250

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -167,6 +167,12 @@ def FeatureMinimum3Maximum3F16 : SubtargetFeature<"minimum3-maximum3-f16",
   "Has v_minimum3_f16 and v_maximum3_f16 instructions"
 >;
 
+def FeatureMin3Max3PKF16 : SubtargetFeature<"min3-max3-pkf16",
+  "HasMin3Max3PKF16",
+  "true",
+  "Has v_pk_min3_num_f16 and v_pk_max3_num_f16 instructions"
+>;
+
 def FeatureMinimum3Maximum3PKF16 : SubtargetFeature<"minimum3-maximum3-pkf16",
   "HasMinimum3Maximum3PKF16",
   "true",
@@ -2001,6 +2007,7 @@ def FeatureISAVersion12_50 : FeatureSet<
    FeatureBF16ConversionInsts,
    FeatureBF16PackedInsts,
    FeatureCvtPkF16F32Inst,
+   FeatureMin3Max3PKF16,
    FeatureMinimum3Maximum3PKF16,
    FeaturePrngInst,
    FeaturePermlane16Swap,
@@ -2361,6 +2368,10 @@ def HasMinimum3Maximum3F16 :
   Predicate<"Subtarget->hasMinimum3Maximum3F16()">,
   AssemblerPredicate<(all_of FeatureMinimum3Maximum3F16)>;
 
+def HasMin3Max3PKF16 :
+  Predicate<"Subtarget->hasMin3Max3PKF16()">,
+  AssemblerPredicate<(all_of FeatureMin3Max3PKF16)>;
+
 def HasMinimum3Maximum3PKF16 :
   Predicate<"Subtarget->hasMinimum3Maximum3PKF16()">,
   AssemblerPredicate<(all_of FeatureMinimum3Maximum3PKF16)>;

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -265,6 +265,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   bool HasIEEEMinimumMaximumInsts = false;
   bool HasMinimum3Maximum3F32 = false;
   bool HasMinimum3Maximum3F16 = false;
+  bool HasMin3Max3PKF16 = false;
   bool HasMinimum3Maximum3PKF16 = false;
   bool HasLshlAddU64Inst = false;
   bool HasAddSubU64Insts = false;
@@ -1388,6 +1389,8 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
     return HasMinimum3Maximum3F16;
   }
 
+  bool hasMin3Max3PKF16() const { return HasMin3Max3PKF16; }
+
   bool hasTanhInsts() const { return HasTanhInsts; }
 
   bool hasAddPC64Inst() const { return GFX1250Insts; }

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -14068,7 +14068,8 @@ static bool supportsMin3Max3(const GCNSubtarget &Subtarget, unsigned Opc,
   case ISD::FMAXIMUMNUM:
   case AMDGPUISD::FMIN_LEGACY:
   case AMDGPUISD::FMAX_LEGACY:
-    return (VT == MVT::f32) || (VT == MVT::f16 && Subtarget.hasMin3Max3_16());
+    return (VT == MVT::f32) || (VT == MVT::f16 && Subtarget.hasMin3Max3_16()) ||
+           (VT == MVT::v2f16 && Subtarget.hasMin3Max3PKF16());
   case ISD::FMINIMUM:
   case ISD::FMAXIMUM:
     return (VT == MVT::f32 && Subtarget.hasMinimum3Maximum3F32()) ||

llvm/lib/Target/AMDGPU/VOP3PInstructions.td

@@ -144,10 +144,17 @@ def : VOP3PSatPat<usubsat, V_PK_SUB_U16>;
 def : VOP3PSatPat<ssubsat, V_PK_SUB_I16>;
 } // End SubtargetPredicate = HasVOP3PInsts
 
-let SubtargetPredicate = HasMinimum3Maximum3PKF16, FPDPRounding = 1 in {
+let isCommutable = 1, FPDPRounding = 1 in {
+let SubtargetPredicate = HasMin3Max3PKF16 in {
+defm V_PK_MIN3_NUM_F16 : VOP3PInst<"v_pk_min3_num_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, AMDGPUfmin3>;
+defm V_PK_MAX3_NUM_F16 : VOP3PInst<"v_pk_max3_num_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, AMDGPUfmax3>;
+}
+
+let SubtargetPredicate = HasMinimum3Maximum3PKF16 in {
 defm V_PK_MINIMUM3_F16 : VOP3PInst<"v_pk_minimum3_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, AMDGPUfminimum3>;
 defm V_PK_MAXIMUM3_F16 : VOP3PInst<"v_pk_maximum3_f16", VOP3P_Profile<VOP_V2F16_V2F16_V2F16_V2F16>, AMDGPUfmaximum3>;
 }
+} // End isCommutable = 1, FPDPRounding = 1
 
 // TODO: Make sure we're doing the right thing with denormals. Note
 // that FMA and MAD will differ.
@@ -2237,6 +2244,8 @@ defm V_PK_MIN_NUM_BF16 : VOP3P_Real_gfx1250<0x2b>;
 defm V_PK_MAX_NUM_BF16 : VOP3P_Real_gfx1250<0x2c>;
 defm V_PK_MINIMUM3_F16 : VOP3P_Real_gfx1250<0x36>;
 defm V_PK_MAXIMUM3_F16 : VOP3P_Real_gfx1250<0x37>;
+defm V_PK_MIN3_NUM_F16 : VOP3P_Real_gfx1250<0x38>;
+defm V_PK_MAX3_NUM_F16 : VOP3P_Real_gfx1250<0x39>;
 
 defm V_PK_MINIMUM_F16 : VOP3P_Real_gfx12<0x1d>;
 defm V_PK_MAXIMUM_F16 : VOP3P_Real_gfx12<0x1e>;

llvm/test/CodeGen/AMDGPU/fmax3.ll

@@ -6,6 +6,8 @@
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GFX11,GFX11-FAKE16 %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GFX12,GFX12-TRUE16 %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GFX12,GFX12-FAKE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1250 -mattr=+real-true16 -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GFX1250,GFX1250-TRUE16 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1250 -mattr=-real-true16 -mattr=-flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GFX1250,GFX1250-FAKE16 %s
 
 define amdgpu_kernel void @test_fmax3_olt_0_f32(ptr addrspace(1) %out, ptr addrspace(1) %aptr, ptr addrspace(1) %bptr, ptr addrspace(1) %cptr) #0 {
 ; SI-LABEL: test_fmax3_olt_0_f32:
@@ -157,6 +159,36 @@ define amdgpu_kernel void @test_fmax3_olt_0_f32(ptr addrspace(1) %out, ptr addrs
 ; GFX12-NEXT:    v_max3_num_f32 v0, v0, v1, v2
 ; GFX12-NEXT:    buffer_store_b32 v0, off, s[8:11], null
 ; GFX12-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: test_fmax3_olt_0_f32:
+; GFX1250:       ; %bb.0:
+; GFX1250-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
+; GFX1250-NEXT:    s_mov_b32 s10, -1
+; GFX1250-NEXT:    s_mov_b32 s11, 0x31016000
+; GFX1250-NEXT:    s_mov_b32 s14, s10
+; GFX1250-NEXT:    s_mov_b32 s15, s11
+; GFX1250-NEXT:    s_mov_b32 s18, s10
+; GFX1250-NEXT:    s_mov_b32 s19, s11
+; GFX1250-NEXT:    s_mov_b32 s22, s10
+; GFX1250-NEXT:    s_mov_b32 s23, s11
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    s_mov_b32 s12, s2
+; GFX1250-NEXT:    s_mov_b32 s13, s3
+; GFX1250-NEXT:    s_mov_b32 s16, s4
+; GFX1250-NEXT:    s_mov_b32 s17, s5
+; GFX1250-NEXT:    s_mov_b32 s20, s6
+; GFX1250-NEXT:    s_mov_b32 s21, s7
+; GFX1250-NEXT:    buffer_load_b32 v0, off, s[12:15], null scope:SCOPE_SYS
+; GFX1250-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-NEXT:    buffer_load_b32 v1, off, s[16:19], null scope:SCOPE_SYS
+; GFX1250-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-NEXT:    buffer_load_b32 v2, off, s[20:23], null scope:SCOPE_SYS
+; GFX1250-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-NEXT:    s_mov_b32 s8, s0
+; GFX1250-NEXT:    s_mov_b32 s9, s1
+; GFX1250-NEXT:    v_max3_num_f32 v0, v0, v1, v2
+; GFX1250-NEXT:    buffer_store_b32 v0, off, s[8:11], null
+; GFX1250-NEXT:    s_endpgm
   %a = load volatile  float, ptr addrspace(1) %aptr, align 4
   %b = load volatile float, ptr addrspace(1) %bptr, align 4
   %c = load volatile float, ptr addrspace(1) %cptr, align 4
@@ -317,6 +349,36 @@ define amdgpu_kernel void @test_fmax3_olt_1_f32(ptr addrspace(1) %out, ptr addrs
 ; GFX12-NEXT:    v_max3_num_f32 v0, v2, v0, v1
 ; GFX12-NEXT:    buffer_store_b32 v0, off, s[8:11], null
 ; GFX12-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: test_fmax3_olt_1_f32:
+; GFX1250:       ; %bb.0:
+; GFX1250-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
+; GFX1250-NEXT:    s_mov_b32 s10, -1
+; GFX1250-NEXT:    s_mov_b32 s11, 0x31016000
+; GFX1250-NEXT:    s_mov_b32 s14, s10
+; GFX1250-NEXT:    s_mov_b32 s15, s11
+; GFX1250-NEXT:    s_mov_b32 s18, s10
+; GFX1250-NEXT:    s_mov_b32 s19, s11
+; GFX1250-NEXT:    s_mov_b32 s22, s10
+; GFX1250-NEXT:    s_mov_b32 s23, s11
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    s_mov_b32 s12, s2
+; GFX1250-NEXT:    s_mov_b32 s13, s3
+; GFX1250-NEXT:    s_mov_b32 s16, s4
+; GFX1250-NEXT:    s_mov_b32 s17, s5
+; GFX1250-NEXT:    s_mov_b32 s20, s6
+; GFX1250-NEXT:    s_mov_b32 s21, s7
+; GFX1250-NEXT:    buffer_load_b32 v0, off, s[12:15], null scope:SCOPE_SYS
+; GFX1250-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-NEXT:    buffer_load_b32 v1, off, s[16:19], null scope:SCOPE_SYS
+; GFX1250-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-NEXT:    buffer_load_b32 v2, off, s[20:23], null scope:SCOPE_SYS
+; GFX1250-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-NEXT:    s_mov_b32 s8, s0
+; GFX1250-NEXT:    s_mov_b32 s9, s1
+; GFX1250-NEXT:    v_max3_num_f32 v0, v2, v0, v1
+; GFX1250-NEXT:    buffer_store_b32 v0, off, s[8:11], null
+; GFX1250-NEXT:    s_endpgm
   %a = load volatile float, ptr addrspace(1) %aptr, align 4
   %b = load volatile float, ptr addrspace(1) %bptr, align 4
   %c = load volatile float, ptr addrspace(1) %cptr, align 4
@@ -544,6 +606,66 @@ define amdgpu_kernel void @test_fmax3_olt_0_f16(ptr addrspace(1) %out, ptr addrs
 ; GFX12-FAKE16-NEXT:    v_max3_num_f16 v0, v0, v1, v2
 ; GFX12-FAKE16-NEXT:    buffer_store_b16 v0, off, s[8:11], null
 ; GFX12-FAKE16-NEXT:    s_endpgm
+;
+; GFX1250-TRUE16-LABEL: test_fmax3_olt_0_f16:
+; GFX1250-TRUE16:       ; %bb.0:
+; GFX1250-TRUE16-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s10, -1
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s14, s10
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s15, s11
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s18, s10
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s19, s11
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s22, s10
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s23, s11
+; GFX1250-TRUE16-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s12, s2
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s13, s3
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s16, s4
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s17, s5
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s20, s6
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s21, s7
+; GFX1250-TRUE16-NEXT:    buffer_load_u16 v0, off, s[12:15], null scope:SCOPE_SYS
+; GFX1250-TRUE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-TRUE16-NEXT:    buffer_load_u16 v1, off, s[16:19], null scope:SCOPE_SYS
+; GFX1250-TRUE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-TRUE16-NEXT:    buffer_load_u16 v2, off, s[20:23], null scope:SCOPE_SYS
+; GFX1250-TRUE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s8, s0
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s9, s1
+; GFX1250-TRUE16-NEXT:    v_max3_num_f16 v0.l, v0.l, v1.l, v2.l
+; GFX1250-TRUE16-NEXT:    buffer_store_b16 v0, off, s[8:11], null
+; GFX1250-TRUE16-NEXT:    s_endpgm
+;
+; GFX1250-FAKE16-LABEL: test_fmax3_olt_0_f16:
+; GFX1250-FAKE16:       ; %bb.0:
+; GFX1250-FAKE16-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s10, -1
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s11, 0x31016000
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s14, s10
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s15, s11
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s18, s10
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s19, s11
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s22, s10
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s23, s11
+; GFX1250-FAKE16-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s12, s2
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s13, s3
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s16, s4
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s17, s5
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s20, s6
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s21, s7
+; GFX1250-FAKE16-NEXT:    buffer_load_u16 v0, off, s[12:15], null scope:SCOPE_SYS
+; GFX1250-FAKE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-FAKE16-NEXT:    buffer_load_u16 v1, off, s[16:19], null scope:SCOPE_SYS
+; GFX1250-FAKE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-FAKE16-NEXT:    buffer_load_u16 v2, off, s[20:23], null scope:SCOPE_SYS
+; GFX1250-FAKE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s8, s0
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s9, s1
+; GFX1250-FAKE16-NEXT:    v_max3_num_f16 v0, v0, v1, v2
+; GFX1250-FAKE16-NEXT:    buffer_store_b16 v0, off, s[8:11], null
+; GFX1250-FAKE16-NEXT:    s_endpgm
   %a = load volatile half, ptr addrspace(1) %aptr, align 2
   %b = load volatile half, ptr addrspace(1) %bptr, align 2
   %c = load volatile half, ptr addrspace(1) %cptr, align 2
@@ -772,6 +894,66 @@ define amdgpu_kernel void @test_fmax3_olt_1_f16(ptr addrspace(1) %out, ptr addrs
 ; GFX12-FAKE16-NEXT:    v_max3_num_f16 v0, v2, v0, v1
 ; GFX12-FAKE16-NEXT:    buffer_store_b16 v0, off, s[8:11], null
 ; GFX12-FAKE16-NEXT:    s_endpgm
+;
+; GFX1250-TRUE16-LABEL: test_fmax3_olt_1_f16:
+; GFX1250-TRUE16:       ; %bb.0:
+; GFX1250-TRUE16-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s10, -1
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s14, s10
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s15, s11
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s18, s10
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s19, s11
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s22, s10
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s23, s11
+; GFX1250-TRUE16-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s12, s2
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s13, s3
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s16, s4
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s17, s5
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s20, s6
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s21, s7
+; GFX1250-TRUE16-NEXT:    buffer_load_u16 v1, off, s[12:15], null scope:SCOPE_SYS
+; GFX1250-TRUE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-TRUE16-NEXT:    buffer_load_u16 v2, off, s[16:19], null scope:SCOPE_SYS
+; GFX1250-TRUE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-TRUE16-NEXT:    buffer_load_u16 v0, off, s[20:23], null scope:SCOPE_SYS
+; GFX1250-TRUE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s8, s0
+; GFX1250-TRUE16-NEXT:    s_mov_b32 s9, s1
+; GFX1250-TRUE16-NEXT:    v_max3_num_f16 v0.l, v0.l, v1.l, v2.l
+; GFX1250-TRUE16-NEXT:    buffer_store_b16 v0, off, s[8:11], null
+; GFX1250-TRUE16-NEXT:    s_endpgm
+;
+; GFX1250-FAKE16-LABEL: test_fmax3_olt_1_f16:
+; GFX1250-FAKE16:       ; %bb.0:
+; GFX1250-FAKE16-NEXT:    s_load_b256 s[0:7], s[4:5], 0x24
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s10, -1
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s11, 0x31016000
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s14, s10
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s15, s11
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s18, s10
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s19, s11
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s22, s10
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s23, s11
+; GFX1250-FAKE16-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s12, s2
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s13, s3
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s16, s4
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s17, s5
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s20, s6
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s21, s7
+; GFX1250-FAKE16-NEXT:    buffer_load_u16 v0, off, s[12:15], null scope:SCOPE_SYS
+; GFX1250-FAKE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-FAKE16-NEXT:    buffer_load_u16 v1, off, s[16:19], null scope:SCOPE_SYS
+; GFX1250-FAKE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-FAKE16-NEXT:    buffer_load_u16 v2, off, s[20:23], null scope:SCOPE_SYS
+; GFX1250-FAKE16-NEXT:    s_wait_loadcnt 0x0
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s8, s0
+; GFX1250-FAKE16-NEXT:    s_mov_b32 s9, s1
+; GFX1250-FAKE16-NEXT:    v_max3_num_f16 v0, v2, v0, v1
+; GFX1250-FAKE16-NEXT:    buffer_store_b16 v0, off, s[8:11], null
+; GFX1250-FAKE16-NEXT:    s_endpgm
   %a = load volatile half, ptr addrspace(1) %aptr, align 2
   %b = load volatile half, ptr addrspace(1) %bptr, align 2
   %c = load volatile half, ptr addrspace(1) %cptr, align 2
@@ -850,6 +1032,15 @@ define <2 x half> @no_fmax3_v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c, <
 ; GFX12-NEXT:    v_pk_max_num_f16 v0, v2, v0
 ; GFX12-NEXT:    v_pk_max_num_f16 v0, v0, v3
 ; GFX12-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: no_fmax3_v2f16:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_pk_max_num_f16 v0, v0, v1
+; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT:    v_pk_max3_num_f16 v0, v2, v0, v3
+; GFX1250-NEXT:    s_set_pc_i64 s[30:31]
 entry:
   %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %a, <2 x half> %b)
   %max1 = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %c, <2 x half> %max)