[llvm] [RISCV] Implement Intrinsics and CodeGen Support for XCValu Extension… (PR #78138)

[Craig Topper via llvm-commits]

================
@@ -704,3 +749,112 @@ let Predicates = [HasVendorXCVbitmanip, IsRV32] in {
             (CV_BITREV GPR:$rs1, cv_tuimm2:$radix, cv_tuimm5:$pts)>;
   def : Pat<(bitreverse (XLenVT GPR:$rs)), (CV_BITREV GPR:$rs, 0, 0)>;
 }
+
+class PatCoreVAluGpr <string intr, string asm> :
+  PatGpr<!cast<Intrinsic>("int_riscv_cv_alu_" # intr),
+            !cast<RVInst>("CV_" # asm)>;
+class PatCoreVAluGprGpr <string intr, string asm> :
+  PatGprGpr<!cast<Intrinsic>("int_riscv_cv_alu_" # intr),
+               !cast<RVInst>("CV_" # asm)>;
+
+multiclass PatCoreVAluGprImm <Intrinsic intr> {
+  def "CV_" # NAME # "_PSEUDO" :
+    Pseudo<(outs GPR:$rd), (ins GPR:$rs, cv_uimm32:$imm), []>;
+  def : PatGprGpr<intr, !cast<RVInst>("CV_" # NAME # "R")>;
+  def : PatGprImm<intr, !cast<RVInst>("CV_" # NAME # "_PSEUDO"), cv_uimm32>;
+}
+
+multiclass PatCoreVAluGprGprImm <Intrinsic intr> {
+  def "CV_" # NAME # "_PSEUDO" :
+    Pseudo<(outs GPR:$rd), (ins GPR:$rs1, GPR:$rs2, uimm5:$imm), []>;
+  def : Pat<(intr GPR:$rs1, GPR:$rs2, GPR:$rs3),
+            (!cast<RVInst>("CV_" # NAME # "R") GPR:$rs1, GPR:$rs2, GPR:$rs3)>;
+  def : Pat<(intr GPR:$rs1, GPR:$rs2, uimm5:$imm),
+            (!cast<RVInst>("CV_" # NAME # "_PSEUDO") GPR:$rs1, GPR:$rs2,
+            uimm5:$imm)>;
+}
+
+let Predicates = [HasVendorXCValu, IsRV32], AddedComplexity = 1 in {
+  def : PatGpr<abs, CV_ABS>;
+  def : PatGprGpr<setle, CV_SLET>;
+  def : PatGprGpr<setule, CV_SLETU>;
+  def : PatGprGpr<smin, CV_MIN>;
+  def : PatGprGpr<umin, CV_MINU>;
+  def : PatGprGpr<smax, CV_MAX>;
+  def : PatGprGpr<umax, CV_MAXU>;
+
+  def : Pat<(sext_inreg (XLenVT GPR:$rs1), i16), (CV_EXTHS GPR:$rs1)>;
+  def : Pat<(sext_inreg (XLenVT GPR:$rs1), i8), (CV_EXTBS GPR:$rs1)>;
+
+  def : Pat<(and (XLenVT GPR:$rs1), 0xffff), (CV_EXTHZ GPR:$rs1)>;
+  def : Pat<(and (XLenVT GPR:$rs1), 0xff),   (CV_EXTBZ GPR:$rs1)>;
+
+  def : Pat<(clip powerOf2Minus1:$upperBound, (XLenVT GPR:$rs1)),
+            (CV_CLIP GPR:$rs1, (trailing1sPlus1 imm:$upperBound))>;
+  def : Pat<(between (not GPR:$rs2), GPR:$rs2, (XLenVT GPR:$rs1)),
+            (CV_CLIPR GPR:$rs1, GPR:$rs2)>;
+  def : Pat<(betweenu powerOf2Minus1:$upperBound, (XLenVT GPR:$rs1)),
+            (CV_CLIPU GPR:$rs1, (trailing1sPlus1 imm:$upperBound))>;
+  def : Pat<(betweenu GPR:$rs2, (XLenVT GPR:$rs1)),
+            (CV_CLIPUR GPR:$rs1, GPR:$rs2)>;
+
+  def : Pat<(sra (add (XLenVT GPR:$rs1), (XLenVT GPR:$rs2)), uimm5:$imm5),
+            (CV_ADDN GPR:$rs1, GPR:$rs2, uimm5:$imm5)>;
+  def : Pat<(srl (add (XLenVT GPR:$rs1), (XLenVT GPR:$rs2)), uimm5:$imm5),
----------------
topperc wrote:

It's not clear to me from the spec that this is the right pattern. I would expect the addition and shift be evaluated with XLen+1 bits so the carry out from the addition is not lost before the shift.

The pattern you have here produces 0 for ((0x80000000+0x80000000)>>1) but I would expect the the cv.addun instruction would produce 0x80000000 when the same inputs and a shift of 1.

https://github.com/llvm/llvm-project/pull/78138