[AMDGPU] Fix wrong MSB encoding for V_FMAMK instructions

These instructions use `src0`, `imm`, `src1` as operand.

Fixes SWDEV-566579.

Author: shiltian

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

@@ -3439,17 +3439,31 @@ getVGPRLoweringOperandTables(const MCInstrDesc &Desc) {
       AMDGPU::OpName::src0Y, AMDGPU::OpName::vsrc1Y, AMDGPU::OpName::vsrc2Y,
       AMDGPU::OpName::vdstY};
 
+  // VOP2 MADMK instructions use src0, imm, src1 scheme.
+  static const AMDGPU::OpName VOP2MADMKOps[4] = {
+      AMDGPU::OpName::src0, AMDGPU::OpName::imm, AMDGPU::OpName::src1,
+      AMDGPU::OpName::vdst};
+
   unsigned TSFlags = Desc.TSFlags;
 
   if (TSFlags &
       (SIInstrFlags::VOP1 | SIInstrFlags::VOP2 | SIInstrFlags::VOP3 |
        SIInstrFlags::VOP3P | SIInstrFlags::VOPC | SIInstrFlags::DPP)) {
+    switch (Desc.getOpcode()) {
     // LD_SCALE operands ignore MSB.
-    if (Desc.getOpcode() == AMDGPU::V_WMMA_LD_SCALE_PAIRED_B32 ||
-        Desc.getOpcode() == AMDGPU::V_WMMA_LD_SCALE_PAIRED_B32_gfx1250 ||
-        Desc.getOpcode() == AMDGPU::V_WMMA_LD_SCALE16_PAIRED_B64 ||
-        Desc.getOpcode() == AMDGPU::V_WMMA_LD_SCALE16_PAIRED_B64_gfx1250)
+    case AMDGPU::V_WMMA_LD_SCALE_PAIRED_B32:
+    case AMDGPU::V_WMMA_LD_SCALE_PAIRED_B32_gfx1250:
+    case AMDGPU::V_WMMA_LD_SCALE16_PAIRED_B64:
+    case AMDGPU::V_WMMA_LD_SCALE16_PAIRED_B64_gfx1250:
       return {};
+    case AMDGPU::V_FMAMK_F32:
+    case AMDGPU::V_FMAMK_F32_gfx12:
+    case AMDGPU::V_FMAMK_F64:
+    case AMDGPU::V_FMAMK_F64_gfx1250:
+      return {VOP2MADMKOps, nullptr};
+    default:
+      break;
+    }
     return {VOPOps, nullptr};
   }
 

llvm/test/CodeGen/AMDGPU/vgpr-lowering-gfx1250.mir

@@ -332,23 +332,39 @@ body:             |
     ; GCN-NEXT: v_fmaak_f32 v0 /*v256*/, v1, v2 /*v258*/, 0x1
     $vgpr256 = V_FMAAK_F32 undef $vgpr1, undef $vgpr258, 1, implicit $exec, implicit $mode
 
-    ; GCN-NEXT: s_set_vgpr_msb 0x4445
+    ; GCN-NEXT: s_set_vgpr_msb 0x4451
     ; GCN-NEXT: v_fmamk_f32 v0 /*v256*/, v1 /*v257*/, 0x1, v2 /*v258*/
     $vgpr256 = V_FMAMK_F32 undef $vgpr257, 1, undef $vgpr258, implicit $exec, implicit $mode
 
-    ; GCN-NEXT: s_set_vgpr_msb 0x4505
+    ; GCN-NEXT: s_set_vgpr_msb 0x5111
     ; GCN-NEXT: v_fmamk_f32 v0, v1 /*v257*/, 0x1, v2 /*v258*/
     $vgpr0 = V_FMAMK_F32 undef $vgpr257, 1, undef $vgpr258, implicit $exec, implicit $mode
 
-    ; GCN-NEXT: s_set_vgpr_msb 0x541
+    ; GCN-NEXT: s_set_vgpr_msb 0x1141
     ; GCN-NEXT: v_fmamk_f32 v0 /*v256*/, v1 /*v257*/, 0x1, v2
     $vgpr256 = V_FMAMK_F32 undef $vgpr257, 1, undef $vgpr2, implicit $exec, implicit $mode
 
-    ; GCN-NEXT: s_set_vgpr_msb 0x4144
+    ; GCN-NEXT: s_set_vgpr_msb 0x4150
     ; GCN-NEXT: v_fmamk_f32 v0 /*v256*/, v1, 0x1, v2 /*v258*/
     $vgpr256 = V_FMAMK_F32 undef $vgpr1, 1, undef $vgpr258, implicit $exec, implicit $mode
 
-    ; ASM: NumVgprs: 259
+    ; GCN-NEXT: s_set_vgpr_msb 0x5051
+    ; GCN-NEXT: v_fmamk_f64 v[4:5] /*v[260:261]*/, v[100:101] /*v[356:357]*/, 0x1, v[2:3] /*v[258:259]*/
+    $vgpr260_vgpr261 = V_FMAMK_F64 undef $vgpr356_vgpr357, 1, undef $vgpr258_vgpr259, implicit $exec, implicit $mode
+
+    ; GCN-NEXT: s_set_vgpr_msb 0x5101
+    ; GCN-NEXT: v_fmamk_f64 v[0:1], v[100:101] /*v[356:357]*/, 0x1, v[2:3]
+    $vgpr0_vgpr1 = V_FMAMK_F64 undef $vgpr356_vgpr357, 1, undef $vgpr2_vgpr3, implicit $exec, implicit $mode
+
+    ; GCN-NEXT: s_set_vgpr_msb 0x110
+    ; GCN-NEXT: v_fmamk_f64 v[0:1], v[2:3], 0x1, v[100:101] /*v[356:357]*/
+    $vgpr0_vgpr1 = V_FMAMK_F64 undef $vgpr2_vgpr3, 1, undef $vgpr356_vgpr357, implicit $exec, implicit $mode
+
+    ; GCN-NEXT: s_set_vgpr_msb 0x1040
+    ; GCN-NEXT: v_fmamk_f64 v[0:1] /*v[256:257]*/, v[2:3], 0x1, v[4:5]
+    $vgpr256_vgpr257 = V_FMAMK_F64 undef $vgpr2_vgpr3, 1, undef $vgpr4_vgpr5, implicit $exec, implicit $mode
+
+    ; ASM: NumVgprs: 358
 
 ...