[PATCH] D113805: [RISCV] Improve codegen for i32 udiv/urem by constant on RV64.

[Craig Topper via Phabricator via llvm-commits]

craig.topper updated this revision to Diff 386959.
craig.topper added a comment.

Remove TODO


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113805/new/

https://reviews.llvm.org/D113805

Files:
  llvm/lib/Target/RISCV/RISCVInstrInfoM.td
  llvm/test/CodeGen/RISCV/div.ll
  llvm/test/CodeGen/RISCV/urem-lkk.ll


Index: llvm/test/CodeGen/RISCV/urem-lkk.ll
===================================================================
--- llvm/test/CodeGen/RISCV/urem-lkk.ll
+++ llvm/test/CodeGen/RISCV/urem-lkk.ll
@@ -103,11 +103,10 @@
 ; RV64IM-LABEL: fold_urem_positive_even:
 ; RV64IM:       # %bb.0:
 ; RV64IM-NEXT:    slli a1, a0, 32
-; RV64IM-NEXT:    srli a1, a1, 32
-; RV64IM-NEXT:    lui a2, 253241
-; RV64IM-NEXT:    slli a2, a2, 2
-; RV64IM-NEXT:    addi a2, a2, -61
-; RV64IM-NEXT:    mul a1, a1, a2
+; RV64IM-NEXT:    lui a2, 1012964
+; RV64IM-NEXT:    addiw a2, a2, -61
+; RV64IM-NEXT:    slli a2, a2, 32
+; RV64IM-NEXT:    mulhu a1, a1, a2
 ; RV64IM-NEXT:    srli a1, a1, 42
 ; RV64IM-NEXT:    addi a2, zero, 1060
 ; RV64IM-NEXT:    mulw a1, a1, a2
Index: llvm/test/CodeGen/RISCV/div.ll
===================================================================
--- llvm/test/CodeGen/RISCV/div.ll
+++ llvm/test/CodeGen/RISCV/div.ll
@@ -78,12 +78,10 @@
 ; RV64IM-LABEL: udiv_constant:
 ; RV64IM:       # %bb.0:
 ; RV64IM-NEXT:    slli a0, a0, 32
-; RV64IM-NEXT:    srli a0, a0, 32
-; RV64IM-NEXT:    lui a1, 205
+; RV64IM-NEXT:    lui a1, 838861
 ; RV64IM-NEXT:    addiw a1, a1, -819
-; RV64IM-NEXT:    slli a1, a1, 12
-; RV64IM-NEXT:    addi a1, a1, -819
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a1, 32
+; RV64IM-NEXT:    mulhu a0, a0, a1
 ; RV64IM-NEXT:    srli a0, a0, 34
 ; RV64IM-NEXT:    ret
   %1 = udiv i32 %a, 5
Index: llvm/lib/Target/RISCV/RISCVInstrInfoM.td
===================================================================
--- llvm/lib/Target/RISCV/RISCVInstrInfoM.td
+++ llvm/lib/Target/RISCV/RISCVInstrInfoM.td
@@ -96,13 +96,24 @@
           (REMW GPR:$rs1, GPR:$rs2)>;
 } // Predicates = [HasStdExtM, IsRV64]
 
+// Pattern to detect constants with no more than 32 active bits that can't
+// be materialized with lui+addiw.
+def uimm32_not_simm32 : PatLeaf<(XLenVT GPR:$a), [{
+  auto *C = dyn_cast<ConstantSDNode>(N);
+  return C && C->hasOneUse() && isUInt<32>(C->getZExtValue()) &&
+         !isInt<32>(C->getSExtValue());
+}]>;
+
 let Predicates = [HasStdExtM, IsRV64, NotHasStdExtZba] in {
 // Special case for calculating the full 64-bit product of a 32x32 unsigned
 // multiply where the inputs aren't known to be zero extended. We can shift the
 // inputs left by 32 and use a MULHU. This saves two SRLIs needed to finish
 // zeroing the upper 32 bits.
-// TODO: If one of the operands is zero extended and the other isn't, we might
-// still be better off shifting both left by 32.
 def : Pat<(i64 (mul (and GPR:$rs1, 0xffffffff), (and GPR:$rs2, 0xffffffff))),
           (MULHU (SLLI GPR:$rs1, 32), (SLLI GPR:$rs2, 32))>;
+// The RHS could also be a constant that is hard to materialize. By shifting
+// left we can allow constant materialization to use LUI+ADDIW via
+// hasAllWUsers.
+def : Pat<(i64 (mul (and GPR:$rs1, 0xffffffff), uimm32_not_simm32:$rs2)),
+          (MULHU (SLLI GPR:$rs1, 32), (SLLI GPR:$rs2, 32))>;
 } // Predicates = [HasStdExtM, IsRV64, NotHasStdExtZba]


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D113805.386959.patch
Type: text/x-patch
Size: 3011 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20211112/83fed0f1/attachment.bin>