AMDGPU: Restrict src0 to VGPRs only for certain cvt scale opcodes. (llvm#127464)

The Src0 operand width higher that 32-bits of cvt_scale opcodes
operating on FP6/BF6/FP4 need to be restricted to take only VGPRs.

Author: Pravin Jagtap

llvm/lib/Target/AMDGPU/SIInstrInfo.td

@@ -1803,19 +1803,16 @@ class getVOP3SrcForVT<ValueType VT, bit IsTrue16 = 0> {
         1                  : VSrc_b32);
 }
 
-// Returns the vreg register class to use for sources of VOP3 instructions for the
-// given VT.
-class getVOP3VRegSrcForVT<ValueType VT, bit IsTrue16 = 0, bit IsFake16 = 0> {
-   RegisterOperand ret =
-   !cond(!eq(VT.Size, 128) : RegisterOperand<VReg_128>,
-         !eq(VT.Size, 96)  : RegisterOperand<VReg_96>,
-         !eq(VT.Size, 64)  : RegisterOperand<VReg_64>,
-         !eq(VT.Size, 48)  : RegisterOperand<VReg_64>,
-         !eq(VT.Size, 16)  : !if(IsTrue16,
-                                 !if(IsFake16, RegisterOperand<VGPR_32>,
-                                               RegisterOperand<VGPR_16>),
-                                 RegisterOperand<VGPR_32>),
-         1                 : RegisterOperand<VGPR_32>);
+// VGPR only VOP3 src with 9 bit encoding
+class getVOP3VRegSrcForVT<ValueType VT> {
+  RegisterOperand ret = !cond(!eq(VT.Size, 1024) : VRegSrc_1024,
+                              !eq(VT.Size, 512)  : VRegSrc_512,
+                              !eq(VT.Size, 256)  : VRegSrc_256,
+                              !eq(VT.Size, 192)  : VRegSrc_192,
+                              !eq(VT.Size, 128)  : VRegSrc_128,
+                              !eq(VT.Size, 96)   : VRegSrc_96,
+                              !eq(VT.Size, 64)   : VRegSrc_64,
+                              1 : VRegSrc_32);
 }
 
 // Src2 of VOP3 DPP instructions cannot be a literal
@@ -2859,6 +2856,7 @@ def VOP_V2I16_F32_F32_F32 : VOPProfile<[v2i16, f32, f32, f32]>;
 def VOP_V2I16_V2F16_F32 : VOPProfile<[v2i16, v2f16, f32, untyped]>;
 def VOP_V2I16_V2BF16_F32 : VOPProfile<[v2i16, v2bf16, f32, untyped]>;
 def VOP_I32_F32_F32_F32 : VOPProfile<[i32, f32, f32, f32]>;
+def VOP_I32_V2F32_I32_F32 : VOPProfile<[i32, v2f32, i32, f32]>;
 def VOP_I32_V2F16_F32_F32 : VOPProfile<[i32, v2f16, f32, f32]>;
 def VOP_I32_V2BF16_F32_F32: VOPProfile<[i32, v2bf16, f32, f32]>;
 def VOP_BF16_F32_I32 : VOPProfile<[bf16, f32, i32, untyped]>;

llvm/lib/Target/AMDGPU/VOP2Instructions.td

@@ -418,12 +418,27 @@ def VOP_MADMK_F16_fake16 : VOP_MADMK <f16> {
 }
 def VOP_MADMK_F32 : VOP_MADMK <f32>;
 
+// Returns the vreg register class to use for sources of VOP3 instructions for the
+// given VT.
+class getVOP3VRegForVT<ValueType VT, bit IsTrue16 = 0, bit IsFake16 = 0> {
+   RegisterOperand ret =
+   !cond(!eq(VT.Size, 128) : RegisterOperand<VReg_128>,
+         !eq(VT.Size, 96)  : RegisterOperand<VReg_96>,
+         !eq(VT.Size, 64)  : RegisterOperand<VReg_64>,
+         !eq(VT.Size, 48)  : RegisterOperand<VReg_64>,
+         !eq(VT.Size, 16)  : !if(IsTrue16,
+                                 !if(IsFake16, RegisterOperand<VGPR_32>,
+                                               RegisterOperand<VGPR_16>),
+                                 RegisterOperand<VGPR_32>),
+         1                 : RegisterOperand<VGPR_32>);
+}
+
 // FIXME: Remove src2_modifiers. It isn't used, so is wasting memory
 // and processing time but it makes it easier to convert to mad.
 class VOP_MAC <ValueType vt0, ValueType vt1=vt0> : VOPProfile <[vt0, vt1, vt1, vt0]> {
   let Ins32 = (ins Src0RC32:$src0, Src1RC32:$src1, getVregSrcForVT<Src2VT>.ret:$src2);
   // Src2 must accept the same operand types as vdst, namely VGPRs only
-  let Src2RC64 = getVOP3VRegSrcForVT<Src2VT, IsTrue16, !not(IsRealTrue16)>.ret;
+  let Src2RC64 = getVOP3VRegForVT<Src2VT, IsTrue16, !not(IsRealTrue16)>.ret;
   let Ins64 = getIns64<Src0RC64, Src1RC64, Src2RC64, 3,
                        0, HasModifiers, HasModifiers, HasOMod,
                        Src0Mod, Src1Mod, Src2Mod>.ret;

llvm/lib/Target/AMDGPU/VOP3Instructions.td

@@ -1052,7 +1052,11 @@ class VOP3_CVT_SCALE_SR_PK_F4_F16BF16_TiedInput_Profile<ValueType Src0Ty> :
   let HasFP4DstByteSel = 1;
 }
 
-def VOP3_CVT_SCALE_SR_PK_F4_F32_TiedInput_Profile : VOP3_Profile<VOPProfile<[i32, v2f32, i32, f32]>, VOP3_OPSEL> {
+class VOP3_CVT_SCALE_SR_PK_F4_F32_TiedInput_Profile<VOPProfile P>
+    : VOP3_Profile<P, VOP3_OPSEL> {
+
+  let Src0RC64 = !if(!gt(P.Src0VT.Size, 32), getVOP3VRegSrcForVT<P.Src0VT>.ret,
+                     getVOP3SrcForVT<P.Src0VT>.ret);
   let InsVOP3OpSel = (ins PackedF32InputMods: $src0_modifiers, Src0RC64:$src0,
                           Int32InputMods:     $src1_modifiers, Src1RC64:$src1,
                           FP32InputMods:      $src2_modifiers, Src2RC64:$src2,
@@ -1100,6 +1104,11 @@ class VOP3_CVT_SCALEF32_PK_F864_Profile<VOPProfile P> : VOP3_Profile<P> {
   let HasExt32BitDPP = 0;
   let HasExtVOP3DPP = 0;
   let HasExt64BitDPP = 0;
+
+  // All convert opcodes operating on FP6/BF6/FP4 data must use VGPR sources for
+  // any operand slots > 32 bit.
+  let Src0RC64 = !if(!gt(P.Src0VT.Size, 32), getVOP3VRegSrcForVT<P.Src0VT>.ret,
+                     getVOP3SrcForVT<P.Src0VT>.ret);
 }
 
 let SubtargetPredicate = HasFP8ConversionScaleInsts, mayRaiseFPException = 0 in {
@@ -1141,7 +1150,10 @@ let SubtargetPredicate = HasFP4ConversionScaleInsts, mayRaiseFPException = 0 in
     let Constraints = "@earlyclobber $vdst" in {
       defm V_CVT_SCALEF32_SR_PK_FP4_F16:  VOP3Inst<"v_cvt_scalef32_sr_pk_fp4_f16", VOP3_CVT_SCALE_SR_PK_F4_F16BF16_TiedInput_Profile<v2f16>>;
       defm V_CVT_SCALEF32_SR_PK_FP4_BF16: VOP3Inst<"v_cvt_scalef32_sr_pk_fp4_bf16", VOP3_CVT_SCALE_SR_PK_F4_F16BF16_TiedInput_Profile<v2bf16>>;
-      defm V_CVT_SCALEF32_SR_PK_FP4_F32:  VOP3Inst<"v_cvt_scalef32_sr_pk_fp4_f32", VOP3_CVT_SCALE_SR_PK_F4_F32_TiedInput_Profile>;
+      defm V_CVT_SCALEF32_SR_PK_FP4_F32
+          : VOP3Inst<"v_cvt_scalef32_sr_pk_fp4_f32",
+                     VOP3_CVT_SCALE_SR_PK_F4_F32_TiedInput_Profile<
+                         VOP_I32_V2F32_I32_F32>>;
     }
   }
   defm V_CVT_SCALEF32_PK_F16_FP4 : VOP3Inst<"v_cvt_scalef32_pk_f16_fp4", VOP3_CVT_SCALE_PK_F16BF16F32_FP4FP8BF8_Profile<v2f16>>;