[llvm-branch-commits] [RISCV] Make fixed-point instructions commutable (PR #90372)

via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Sat Apr 27 21:03:01 PDT 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

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

Author: Pengcheng Wang (wangpc-pp)

<details>
<summary>Changes</summary>

This PR includes:
* vsadd.vv/vsaddu.vv
* vaadd.vv/vaaddu.vv
* vsmul.vv


---
Full diff: https://github.com/llvm/llvm-project/pull/90372.diff


3 Files Affected:

- (modified) llvm/lib/Target/RISCV/RISCVInstrInfo.cpp (+5) 
- (modified) llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td (+16-13) 
- (modified) llvm/test/CodeGen/RISCV/rvv/commutable.ll (+6-8) 


``````````diff
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
index 3efd09aeae879d..8cb9a40a98bcd8 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
@@ -3132,6 +3132,11 @@ bool RISCVInstrInfo::findCommutedOpIndices(const MachineInstr &MI,
   case CASE_RVV_OPCODE_WIDEN(VWMACC_VV):
   case CASE_RVV_OPCODE_WIDEN(VWMACCU_VV):
   case CASE_RVV_OPCODE_UNMASK(VADC_VVM):
+  case CASE_RVV_OPCODE(VSADD_VV):
+  case CASE_RVV_OPCODE(VSADDU_VV):
+  case CASE_RVV_OPCODE(VAADD_VV):
+  case CASE_RVV_OPCODE(VAADDU_VV):
+  case CASE_RVV_OPCODE(VSMUL_VV):
     // Operands 2 and 3 are commutable.
     return fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, 2, 3);
   case CASE_VFMA_SPLATS(FMADD):
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
index e9715b40adc079..fc60a9cc7cd30e 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td
@@ -2146,8 +2146,9 @@ multiclass VPseudoBinaryRoundingMode<VReg RetClass,
                                      string Constraint = "",
                                      int sew = 0,
                                      int UsesVXRM = 1,
-                                     int TargetConstraintType = 1> {
-  let VLMul = MInfo.value, SEW=sew in {
+                                     int TargetConstraintType = 1,
+                                     bit Commutable = 0> {
+  let VLMul = MInfo.value, SEW=sew, isCommutable = Commutable in {
     defvar suffix = !if(sew, "_" # MInfo.MX # "_E" # sew, "_" # MInfo.MX);
     def suffix : VPseudoBinaryNoMaskRoundingMode<RetClass, Op1Class, Op2Class,
                                                  Constraint, UsesVXRM,
@@ -2232,8 +2233,9 @@ multiclass VPseudoBinaryV_VV<LMULInfo m, string Constraint = "", int sew = 0, bi
   defm _VV : VPseudoBinary<m.vrclass, m.vrclass, m.vrclass, m, Constraint, sew, Commutable=Commutable>;
 }
 
-multiclass VPseudoBinaryV_VV_RM<LMULInfo m, string Constraint = ""> {
-  defm _VV : VPseudoBinaryRoundingMode<m.vrclass, m.vrclass, m.vrclass, m, Constraint>;
+multiclass VPseudoBinaryV_VV_RM<LMULInfo m, string Constraint = "", bit Commutable = 0> {
+  defm _VV : VPseudoBinaryRoundingMode<m.vrclass, m.vrclass, m.vrclass, m, Constraint,
+                                       Commutable=Commutable>;
 }
 
 // Similar to VPseudoBinaryV_VV, but uses MxListF.
@@ -2715,10 +2717,11 @@ multiclass VPseudoVGTR_VV_VX_VI<Operand ImmType = simm5, string Constraint = "">
   }
 }
 
-multiclass VPseudoVSALU_VV_VX_VI<Operand ImmType = simm5, string Constraint = ""> {
+multiclass VPseudoVSALU_VV_VX_VI<Operand ImmType = simm5, string Constraint = "",
+                                 bit Commutable = 0> {
   foreach m = MxList in {
     defvar mx = m.MX;
-    defm "" : VPseudoBinaryV_VV<m, Constraint>,
+    defm "" : VPseudoBinaryV_VV<m, Constraint, Commutable=Commutable>,
               SchedBinary<"WriteVSALUV", "ReadVSALUV", "ReadVSALUX", mx,
                           forceMergeOpRead=true>;
     defm "" : VPseudoBinaryV_VX<m, Constraint>,
@@ -2788,7 +2791,7 @@ multiclass VPseudoVSALU_VV_VX {
 multiclass VPseudoVSMUL_VV_VX_RM {
   foreach m = MxList in {
     defvar mx = m.MX;
-    defm "" : VPseudoBinaryV_VV_RM<m>,
+    defm "" : VPseudoBinaryV_VV_RM<m, Commutable=1>,
               SchedBinary<"WriteVSMulV", "ReadVSMulV", "ReadVSMulV", mx,
                           forceMergeOpRead=true>;
     defm "" : VPseudoBinaryV_VX_RM<m>,
@@ -2797,10 +2800,10 @@ multiclass VPseudoVSMUL_VV_VX_RM {
   }
 }
 
-multiclass VPseudoVAALU_VV_VX_RM {
+multiclass VPseudoVAALU_VV_VX_RM<bit Commutable = 0> {
   foreach m = MxList in {
     defvar mx = m.MX;
-    defm "" : VPseudoBinaryV_VV_RM<m>,
+    defm "" : VPseudoBinaryV_VV_RM<m, Commutable=Commutable>,
               SchedBinary<"WriteVAALUV", "ReadVAALUV", "ReadVAALUV", mx,
                           forceMergeOpRead=true>;
     defm "" : VPseudoBinaryV_VX_RM<m>,
@@ -6448,8 +6451,8 @@ defm PseudoVMV_V : VPseudoUnaryVMV_V_X_I;
 // 12.1. Vector Single-Width Saturating Add and Subtract
 //===----------------------------------------------------------------------===//
 let Defs = [VXSAT], hasSideEffects = 1 in {
-  defm PseudoVSADDU : VPseudoVSALU_VV_VX_VI;
-  defm PseudoVSADD  : VPseudoVSALU_VV_VX_VI;
+  defm PseudoVSADDU : VPseudoVSALU_VV_VX_VI<Commutable=1>;
+  defm PseudoVSADD  : VPseudoVSALU_VV_VX_VI<Commutable=1>;
   defm PseudoVSSUBU : VPseudoVSALU_VV_VX;
   defm PseudoVSSUB  : VPseudoVSALU_VV_VX;
 }
@@ -6457,8 +6460,8 @@ let Defs = [VXSAT], hasSideEffects = 1 in {
 //===----------------------------------------------------------------------===//
 // 12.2. Vector Single-Width Averaging Add and Subtract
 //===----------------------------------------------------------------------===//
-defm PseudoVAADDU : VPseudoVAALU_VV_VX_RM;
-defm PseudoVAADD  : VPseudoVAALU_VV_VX_RM;
+defm PseudoVAADDU : VPseudoVAALU_VV_VX_RM<Commutable=1>;
+defm PseudoVAADD  : VPseudoVAALU_VV_VX_RM<Commutable=1>;
 defm PseudoVASUBU : VPseudoVAALU_VV_VX_RM;
 defm PseudoVASUB  : VPseudoVAALU_VV_VX_RM;
 
diff --git a/llvm/test/CodeGen/RISCV/rvv/commutable.ll b/llvm/test/CodeGen/RISCV/rvv/commutable.ll
index e383c1b477c45d..06a6327d3892b6 100644
--- a/llvm/test/CodeGen/RISCV/rvv/commutable.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/commutable.ll
@@ -724,10 +724,9 @@ define <vscale x 1 x i64> @commutable_vaadd_vv(<vscale x 1 x i64> %0, <vscale x
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    vsetvli zero, a0, e64, m1, ta, ma
 ; CHECK-NEXT:    csrwi vxrm, 0
-; CHECK-NEXT:    vaadd.vv v10, v8, v9
-; CHECK-NEXT:    vaadd.vv v8, v9, v8
+; CHECK-NEXT:    vaadd.vv v8, v8, v9
 ; CHECK-NEXT:    vsetvli a0, zero, e64, m1, ta, ma
-; CHECK-NEXT:    vadd.vv v8, v10, v8
+; CHECK-NEXT:    vadd.vv v8, v8, v8
 ; CHECK-NEXT:    ret
 entry:
   %a = call <vscale x 1 x i64> @llvm.riscv.vaadd.nxv1i64.nxv1i64(<vscale x 1 x i64> undef, <vscale x 1 x i64> %0, <vscale x 1 x i64> %1, iXLen 0, iXLen %2)
@@ -743,7 +742,7 @@ define <vscale x 1 x i64> @commutable_vaadd_vv_masked(<vscale x 1 x i64> %0, <vs
 ; CHECK-NEXT:    vsetvli zero, a0, e64, m1, ta, ma
 ; CHECK-NEXT:    csrwi vxrm, 0
 ; CHECK-NEXT:    vaadd.vv v10, v8, v9, v0.t
-; CHECK-NEXT:    vaadd.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vaadd.vv v8, v8, v9, v0.t
 ; CHECK-NEXT:    vsetvli a0, zero, e64, m1, ta, ma
 ; CHECK-NEXT:    vadd.vv v8, v10, v8
 ; CHECK-NEXT:    ret
@@ -760,10 +759,9 @@ define <vscale x 1 x i64> @commutable_vaaddu_vv(<vscale x 1 x i64> %0, <vscale x
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    vsetvli zero, a0, e64, m1, ta, ma
 ; CHECK-NEXT:    csrwi vxrm, 0
-; CHECK-NEXT:    vaaddu.vv v10, v8, v9
-; CHECK-NEXT:    vaaddu.vv v8, v9, v8
+; CHECK-NEXT:    vaaddu.vv v8, v8, v9
 ; CHECK-NEXT:    vsetvli a0, zero, e64, m1, ta, ma
-; CHECK-NEXT:    vadd.vv v8, v10, v8
+; CHECK-NEXT:    vadd.vv v8, v8, v8
 ; CHECK-NEXT:    ret
 entry:
   %a = call <vscale x 1 x i64> @llvm.riscv.vaaddu.nxv1i64.nxv1i64(<vscale x 1 x i64> undef, <vscale x 1 x i64> %0, <vscale x 1 x i64> %1, iXLen 0, iXLen %2)
@@ -779,7 +777,7 @@ define <vscale x 1 x i64> @commutable_vaaddu_vv_masked(<vscale x 1 x i64> %0, <v
 ; CHECK-NEXT:    vsetvli zero, a0, e64, m1, ta, ma
 ; CHECK-NEXT:    csrwi vxrm, 0
 ; CHECK-NEXT:    vaaddu.vv v10, v8, v9, v0.t
-; CHECK-NEXT:    vaaddu.vv v8, v9, v8, v0.t
+; CHECK-NEXT:    vaaddu.vv v8, v8, v9, v0.t
 ; CHECK-NEXT:    vsetvli a0, zero, e64, m1, ta, ma
 ; CHECK-NEXT:    vadd.vv v8, v10, v8
 ; CHECK-NEXT:    ret

``````````

</details>


https://github.com/llvm/llvm-project/pull/90372


More information about the llvm-branch-commits mailing list