[RISCV] Add @earlyclobber to SiFive custom matrix multiply instruction.

[topperc]

All of these have a constraint that vd and vs1 cannot overlap. Some of them have an additional widening constraint for vs2.

We should use earlyclobber to protect this.

This is unlikely to be an issue in practice due to the instrinsic being ternary so vd is also a source. The intrinsic has a different type for this source than the other sources. You would have to do something crazy to get the register allocator to overlap the registers.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Craig Topper (topperc)

Changes

All of these have a constraint that vd and vs1 cannot overlap. Some of them have an additional widening constraint for vs2.

We should use earlyclobber to protect this.

This is unlikely to be an issue in practice due to the instrinsic being ternary so vd is also a source. The intrinsic has a different type for this source than the other sources. You would have to do something crazy to get the register allocator to overlap the registers.

---

Full diff: https://github.com/llvm/llvm-project/pull/124060.diff

1 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td (+9-9)

diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td b/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td
index 20adda91f6bde1..b55c6c55689361 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td
@@ -374,28 +374,28 @@ multiclass VPseudoVC_XVW<LMULInfo m, DAGOperand RS1Class,
   }
 }
 
-multiclass VPseudoSiFiveVMACC<string mx, VReg vd_type, VReg vs2_type,
-                              string Constraint = ""> {
+multiclass VPseudoSiFiveVMACC<string mx, VReg vd_type, VReg vs2_type> {
   def "Pseudo" # NAME # "_" # mx
-      : VPseudoTernaryNoMaskWithPolicy<vd_type, V_M1.vrclass, vs2_type, Constraint>;
+      : VPseudoTernaryNoMaskWithPolicy<vd_type, V_M1.vrclass, vs2_type,
+                                       "@earlyclobber $rd">;
 }
 
-multiclass VPseudoSiFiveVQMACCDOD<string Constraint = ""> {
+multiclass VPseudoSiFiveVQMACCDOD {
   foreach m = MxListVF8 in
     let VLMul = m.value in
-    defm NAME : VPseudoSiFiveVMACC<m.MX, m.vrclass, m.vrclass, Constraint>;
+    defm NAME : VPseudoSiFiveVMACC<m.MX, m.vrclass, m.vrclass>;
 }
 
-multiclass VPseudoSiFiveVQMACCQOQ<string Constraint = ""> {
+multiclass VPseudoSiFiveVQMACCQOQ {
   foreach m = [V_MF2, V_M1, V_M2, V_M4] in
     let VLMul = m.value in
-    defm NAME : VPseudoSiFiveVMACC<m.MX, m.wvrclass, m.vrclass, Constraint>;
+    defm NAME : VPseudoSiFiveVMACC<m.MX, m.wvrclass, m.vrclass>;
 }
 
-multiclass VPseudoSiFiveVFWMACC<string Constraint = ""> {
+multiclass VPseudoSiFiveVFWMACC {
   foreach m = MxListVF2 in
     let VLMul = m.value in
-    defm NAME : VPseudoSiFiveVMACC<m.MX, m.wvrclass, m.vrclass, Constraint>;
+    defm NAME : VPseudoSiFiveVMACC<m.MX, m.wvrclass, m.vrclass>;
 }
 
 multiclass VPseudoSiFiveVFNRCLIP<string Constraint = "@earlyclobber $rd"> {

[wangpc-pp]

I don't know this can be tested, but LGTM since this is a vendor extension.