[AMDGPU] Allow unaligned VGPR for ds_read_b96_tr_b6 (llvm#125169) (#400)

All load transpose instructions follow gfx950 standard of even aligned
VGPR except ds_read_b96_tr_b6, which allows unaligned VGPR.

Co-authored-by: Sirish Pande
[[email protected]](mailto:[email protected])

Author: Jagtap, Pravin

llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

@@ -4803,7 +4803,10 @@ bool AMDGPUAsmParser::validateAGPRLdSt(const MCInst &Inst) const {
 
 bool AMDGPUAsmParser::validateVGPRAlign(const MCInst &Inst) const {
   auto FB = getFeatureBits();
-  if (!FB[AMDGPU::FeatureGFX90AInsts])
+  unsigned Opc = Inst.getOpcode();
+  // DS_READ_B96_TR_B6 is the only DS instruction in GFX950, that allows
+  // unaligned VGPR. All others only allow even aligned VGPRs.
+  if (!(FB[AMDGPU::FeatureGFX90AInsts]) || Opc == AMDGPU::DS_READ_B96_TR_B6_vi)
     return true;
 
   const MCRegisterInfo *MRI = getMRI();

llvm/test/MC/AMDGPU/gfx950-unsupported.s

@@ -239,10 +239,6 @@ ds_read_b64_tr_b16 v[2:3], v2 offset:-64
 //===----------------------------------------------------------------------===//
 // ds_read_b96_tr_b6
 //===----------------------------------------------------------------------===//
-ds_read_b96_tr_b6 v[1:3], v0
-// ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid register class: vgpr tuples must be 64 bit aligned
-// W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
-
 ds_read_b96_tr_b6 v1, v0
 // ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid operand for instruction
 // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU

llvm/test/MC/AMDGPU/gfx950_asm_read_tr.s

@@ -32,3 +32,11 @@ ds_read_b96_tr_b6 v[0:2], v0
 ds_read_b96_tr_b6 v[2:4], v2 offset:64
 // GFX940-ERR: [[@LINE-1]]:{{[0-9]+}}: error: instruction not supported on this GPU
 // GFX950: encoding: [0x40,0x00,0xc2,0xd9,0x02,0x00,0x00,0x02]
+
+ds_read_b96_tr_b6 v[1:3], v0
+// GFX940-ERR: [[@LINE-1]]:{{[0-9]+}}: error: instruction not supported on this GPU
+// GFX950: encoding: [0x00,0x00,0xc2,0xd9,0x00,0x00,0x00,0x01]
+
+ds_read_b96_tr_b6 v[1:3], v2 offset:64
+// GFX940-ERR: [[@LINE-1]]:{{[0-9]+}}: error: instruction not supported on this GPU
+// GFX950: encoding: [0x40,0x00,0xc2,0xd9,0x02,0x00,0x00,0x01]

llvm/test/MC/Disassembler/AMDGPU/gfx950_dasm_ds_read_tr.txt

@@ -35,3 +35,9 @@
 
 # GFX950: ds_read_b96_tr_b6 v[2:4], v2 offset:64  ; encoding: [0x40,0x00,0xc2,0xd9,0x02,0x00,0x00,0x02]
 0x40,0x00,0xc2,0xd9,0x02,0x00,0x00,0x02
+
+# GFX950: ds_read_b96_tr_b6 v[1:3], v0            ; encoding: [0x00,0x00,0xc2,0xd9,0x00,0x00,0x00,0x01]
+0x00,0x00,0xc2,0xd9,0x00,0x00,0x00,0x01
+
+# GFX950: ds_read_b96_tr_b6 v[1:3], v2 offset:64  ; encoding: [0x40,0x00,0xc2,0xd9,0x02,0x00,0x00,0x01]
+0x40,0x00,0xc2,0xd9,0x02,0x00,0x00,0x01