[llvm] [AMDGPU] Add subtarget features for MAD NC and 64-bit MIN/MAX instructions (PR #196326)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Mariusz Sikora (mariusz-sikora-at-amd)

<details>
<summary>Changes</summary>



---
Full diff: https://github.com/llvm/llvm-project/pull/196326.diff


2 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/AMDGPU.td (+10) 
- (modified) llvm/lib/Target/AMDGPU/VOP3Instructions.td (+3-3) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index 25fc64d178858..dde14ef6574a8 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -1317,10 +1317,18 @@ defm MadU32Inst : AMDGPUSubtargetFeature<"mad-u32-inst",
   "Has v_mad_u32 instruction"
 >;
 
+defm MadNC64_32Insts : AMDGPUSubtargetFeature<"mad-nc-64-32-insts",
+  "Has  v_mad_nc_{u64_u32|i64_i32} instructions"
+>;
+
 defm AddMinMaxInsts : AMDGPUSubtargetFeature<"add-min-max-insts",
   "Has v_add_{min|max}_{i|u}32 instructions"
 >;
 
+defm MinMax64Insts : AMDGPUSubtargetFeature<"min-max-64-insts",
+  "Has v_{min|max}_{i|u}64 instructions"
+>;
+
 defm PkAddMinMaxInsts : AMDGPUSubtargetFeature<"pk-add-min-max-insts",
   "Has v_pk_add_{min|max}_{i|u}16 instructions"
 >;
@@ -2097,7 +2105,9 @@ def FeatureISAVersion12_50_Common : FeatureSet<
    FeatureLshlAddU64Inst,
    FeatureAddSubU64Insts,
    FeatureMadU32Inst,
+   FeatureMadNC64_32Insts,
    FeatureAddMinMaxInsts,
+   FeatureMinMax64Insts,
    FeaturePkAddMinMaxInsts,
    FeatureLdsBarrierArriveAtomic,
    FeatureSetPrioIncWgInst,
diff --git a/llvm/lib/Target/AMDGPU/VOP3Instructions.td b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
index 41e76063386ec..4970d24125d45 100644
--- a/llvm/lib/Target/AMDGPU/VOP3Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
@@ -191,7 +191,7 @@ let SubtargetPredicate = HasLerpInst in
 let SchedRW = [WriteIntMul] in {
   let SubtargetPredicate = HasMadU32Inst in
     defm V_MAD_U32 : VOP3Inst <"v_mad_u32", V_MAD_U32_PROF>;
-  let SubtargetPredicate = isGFX1250Plus in {
+  let SubtargetPredicate = HasMadNC64_32Insts in {
     defm V_MAD_NC_U64_U32 : VOP3Inst<"v_mad_nc_u64_u32", VOP_I32_I32_I64_DPP>;
     defm V_MAD_NC_I64_I32 : VOP3Inst<"v_mad_nc_i64_i32", VOP_I32_I32_I64_DPP>;
   }
@@ -232,12 +232,12 @@ defm V_MAXIMUM_F64 : VOP3Inst <"v_maximum_f64", VOP3_Profile<VOP_F64_F64_F64>, f
 } // End SchedRW = [WriteDoubleAdd], IsDPMACCInstruction = 1
 } // End SubtargetPredicate = HasIEEEMinimumMaximumInsts, ReadsModeReg = 0, AddedComplexity = 1
 
-let SubtargetPredicate = isGFX1250Plus, SchedRW = [WriteDoubleAdd] in {
+let SubtargetPredicate = HasMinMax64Insts, SchedRW = [WriteDoubleAdd] in {
 defm V_MAX_I64 : VOP3Inst <"v_max_i64", VOP_I64_I64_I64_DPP, smax>;
 defm V_MAX_U64 : VOP3Inst <"v_max_u64", VOP_I64_I64_I64_DPP, umax>;
 defm V_MIN_I64 : VOP3Inst <"v_min_i64", VOP_I64_I64_I64_DPP, smin>;
 defm V_MIN_U64 : VOP3Inst <"v_min_u64", VOP_I64_I64_I64_DPP, umin>;
-} // End SubtargetPredicate = isGFX1250Plus, SchedRW = [WriteDoubleAdd]
+} // End SubtargetPredicate = HasMinMax64Insts, SchedRW = [WriteDoubleAdd]
 
 } // End isReMaterializable = 1
 

``````````

</details>


https://github.com/llvm/llvm-project/pull/196326