[RISCV] Extend pattern matches involving shNadd to support disjoint or (#87001)

I tried to add representative tests while not duplicating complete
coverage.  If there's other tests you'd like to see, let me know.

2 files changed, 114 insertions, 25 deletions

diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td b/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td
index 549bc039fabb..434b071e628a 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td
@@ -676,38 +676,38 @@ let Predicates = [HasStdExtZba] in {
 
 foreach i = {1,2,3} in {
   defvar shxadd = !cast<Instruction>("SH"#i#"ADD");
-  def : Pat<(XLenVT (add_non_imm12 (shl GPR:$rs1, (XLenVT i)), GPR:$rs2)),
+  def : Pat<(XLenVT (add_like_non_imm12 (shl GPR:$rs1, (XLenVT i)), GPR:$rs2)),
             (shxadd GPR:$rs1, GPR:$rs2)>;
 
   defvar pat = !cast<ComplexPattern>("sh"#i#"add_op");
   // More complex cases use a ComplexPattern.
-  def : Pat<(XLenVT (add_non_imm12 pat:$rs1, GPR:$rs2)),
+  def : Pat<(XLenVT (add_like_non_imm12 pat:$rs1, GPR:$rs2)),
             (shxadd pat:$rs1, GPR:$rs2)>;
 }
 
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 6)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 6)), GPR:$rs2),
           (SH1ADD (XLenVT (SH1ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 10)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 10)), GPR:$rs2),
           (SH1ADD (XLenVT (SH2ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 18)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 18)), GPR:$rs2),
           (SH1ADD (XLenVT (SH3ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 12)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 12)), GPR:$rs2),
           (SH2ADD (XLenVT (SH1ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 20)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 20)), GPR:$rs2),
           (SH2ADD (XLenVT (SH2ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 36)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 36)), GPR:$rs2),
           (SH2ADD (XLenVT (SH3ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 24)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 24)), GPR:$rs2),
           (SH3ADD (XLenVT (SH1ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 40)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 40)), GPR:$rs2),
           (SH3ADD (XLenVT (SH2ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
-def : Pat<(add (mul_oneuse GPR:$rs1, (XLenVT 72)), GPR:$rs2),
+def : Pat<(add_like (mul_oneuse GPR:$rs1, (XLenVT 72)), GPR:$rs2),
           (SH3ADD (XLenVT (SH3ADD GPR:$rs1, GPR:$rs1)), GPR:$rs2)>;
 
-def : Pat<(add (XLenVT GPR:$r), CSImm12MulBy4:$i),
+def : Pat<(add_like (XLenVT GPR:$r), CSImm12MulBy4:$i),
           (SH2ADD (XLenVT (ADDI (XLenVT X0), (SimmShiftRightBy2XForm CSImm12MulBy4:$i))),
                   GPR:$r)>;
-def : Pat<(add (XLenVT GPR:$r), CSImm12MulBy8:$i),
+def : Pat<(add_like (XLenVT GPR:$r), CSImm12MulBy8:$i),
           (SH3ADD (XLenVT (ADDI (XLenVT X0), (SimmShiftRightBy3XForm CSImm12MulBy8:$i))),
                   GPR:$r)>;
 
@@ -762,37 +762,37 @@ def : Pat<(i64 (and GPR:$rs, 0xFFFFFFFF)), (ADD_UW GPR:$rs, (XLenVT X0))>;
 
 foreach i = {1,2,3} in {
   defvar shxadd_uw = !cast<Instruction>("SH"#i#"ADD_UW");
-  def : Pat<(i64 (add_non_imm12 (shl (and GPR:$rs1, 0xFFFFFFFF), (i64 i)), (XLenVT GPR:$rs2))),
+  def : Pat<(i64 (add_like_non_imm12 (shl (and GPR:$rs1, 0xFFFFFFFF), (i64 i)), (XLenVT GPR:$rs2))),
             (shxadd_uw GPR:$rs1, GPR:$rs2)>;
 }
 
-def : Pat<(i64 (add_non_imm12 (and (shl GPR:$rs1, (i64 1)), 0x1FFFFFFFF), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and (shl GPR:$rs1, (i64 1)), 0x1FFFFFFFF), (XLenVT GPR:$rs2))),
           (SH1ADD_UW GPR:$rs1, GPR:$rs2)>;
-def : Pat<(i64 (add_non_imm12 (and (shl GPR:$rs1, (i64 2)), 0x3FFFFFFFF), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and (shl GPR:$rs1, (i64 2)), 0x3FFFFFFFF), (XLenVT GPR:$rs2))),
           (SH2ADD_UW GPR:$rs1, GPR:$rs2)>;
-def : Pat<(i64 (add_non_imm12 (and (shl GPR:$rs1, (i64 3)), 0x7FFFFFFFF), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and (shl GPR:$rs1, (i64 3)), 0x7FFFFFFFF), (XLenVT GPR:$rs2))),
           (SH3ADD_UW GPR:$rs1, GPR:$rs2)>;
 
 // More complex cases use a ComplexPattern.
 foreach i = {1,2,3} in {
   defvar pat = !cast<ComplexPattern>("sh"#i#"add_uw_op");
-  def : Pat<(i64 (add_non_imm12 pat:$rs1, (XLenVT GPR:$rs2))),
+  def : Pat<(i64 (add_like_non_imm12 pat:$rs1, (XLenVT GPR:$rs2))),
             (!cast<Instruction>("SH"#i#"ADD_UW") pat:$rs1, GPR:$rs2)>;
 }
 
-def : Pat<(i64 (add_non_imm12 (and GPR:$rs1, 0xFFFFFFFE), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and GPR:$rs1, 0xFFFFFFFE), (XLenVT GPR:$rs2))),
           (SH1ADD (XLenVT (SRLIW GPR:$rs1, 1)), GPR:$rs2)>;
-def : Pat<(i64 (add_non_imm12 (and GPR:$rs1, 0xFFFFFFFC), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and GPR:$rs1, 0xFFFFFFFC), (XLenVT GPR:$rs2))),
           (SH2ADD (XLenVT (SRLIW GPR:$rs1, 2)), GPR:$rs2)>;
-def : Pat<(i64 (add_non_imm12 (and GPR:$rs1, 0xFFFFFFF8), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and GPR:$rs1, 0xFFFFFFF8), (XLenVT GPR:$rs2))),
           (SH3ADD (XLenVT (SRLIW GPR:$rs1, 3)), GPR:$rs2)>;
 
 // Use SRLI to clear the LSBs and SHXADD_UW to mask and shift.
-def : Pat<(i64 (add_non_imm12 (and GPR:$rs1, 0x1FFFFFFFE), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and GPR:$rs1, 0x1FFFFFFFE), (XLenVT GPR:$rs2))),
           (SH1ADD_UW (XLenVT (SRLI GPR:$rs1, 1)), GPR:$rs2)>;
-def : Pat<(i64 (add_non_imm12 (and GPR:$rs1, 0x3FFFFFFFC), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and GPR:$rs1, 0x3FFFFFFFC), (XLenVT GPR:$rs2))),
           (SH2ADD_UW (XLenVT (SRLI GPR:$rs1, 2)), GPR:$rs2)>;
-def : Pat<(i64 (add_non_imm12 (and GPR:$rs1, 0x7FFFFFFF8), (XLenVT GPR:$rs2))),
+def : Pat<(i64 (add_like_non_imm12 (and GPR:$rs1, 0x7FFFFFFF8), (XLenVT GPR:$rs2))),
           (SH3ADD_UW (XLenVT (SRLI GPR:$rs1, 3)), GPR:$rs2)>;
 
 def : Pat<(i64 (mul (and_oneuse GPR:$r, 0xFFFFFFFF), C3LeftShiftUW:$i)),
@@ -879,7 +879,7 @@ def : Pat<(zext GPR:$src), (ADD_UW GPR:$src, (XLenVT X0))>;
 
 foreach i = {1,2,3} in {
   defvar shxadd = !cast<Instruction>("SH"#i#"ADD");
-  def : Pat<(i32 (add_non_imm12 (shl GPR:$rs1, (i64 i)), GPR:$rs2)),
+  def : Pat<(i32 (add_like_non_imm12 (shl GPR:$rs1, (i64 i)), GPR:$rs2)),
             (shxadd GPR:$rs1, GPR:$rs2)>;
 }
 }
diff --git a/llvm/test/CodeGen/RISCV/rv64zba.ll b/llvm/test/CodeGen/RISCV/rv64zba.ll
index d9d83633a853..c81c6aeaab89 100644
--- a/llvm/test/CodeGen/RISCV/rv64zba.ll
+++ b/llvm/test/CodeGen/RISCV/rv64zba.ll
@@ -209,6 +209,24 @@ define i64 @sh1adduw_2(i64 %0, i64 %1) {
   ret i64 %5
 }
 
+define i64 @sh1adduw_3(i64 %0, i64 %1) {
+; RV64I-LABEL: sh1adduw_3:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    slli a0, a0, 32
+; RV64I-NEXT:    srli a0, a0, 31
+; RV64I-NEXT:    or a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: sh1adduw_3:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    sh1add.uw a0, a0, a1
+; RV64ZBA-NEXT:    ret
+  %3 = shl i64 %0, 1
+  %4 = and i64 %3, 8589934590
+  %5 = or disjoint i64 %4, %1
+  ret i64 %5
+}
+
 define signext i32 @sh2adduw(i32 signext %0, ptr %1) {
 ; RV64I-LABEL: sh2adduw:
 ; RV64I:       # %bb.0:
@@ -247,6 +265,24 @@ define i64 @sh2adduw_2(i64 %0, i64 %1) {
   ret i64 %5
 }
 
+define i64 @sh2adduw_3(i64 %0, i64 %1) {
+; RV64I-LABEL: sh2adduw_3:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    slli a0, a0, 32
+; RV64I-NEXT:    srli a0, a0, 30
+; RV64I-NEXT:    or a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: sh2adduw_3:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    sh2add.uw a0, a0, a1
+; RV64ZBA-NEXT:    ret
+  %3 = shl i64 %0, 2
+  %4 = and i64 %3, 17179869180
+  %5 = or disjoint i64 %4, %1
+  ret i64 %5
+}
+
 define i64 @sh3adduw(i32 signext %0, ptr %1) {
 ; RV64I-LABEL: sh3adduw:
 ; RV64I:       # %bb.0:
@@ -285,6 +321,24 @@ define i64 @sh3adduw_2(i64 %0, i64 %1) {
   ret i64 %5
 }
 
+define i64 @sh3adduw_3(i64 %0, i64 %1) {
+; RV64I-LABEL: sh3adduw_3:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    slli a0, a0, 32
+; RV64I-NEXT:    srli a0, a0, 29
+; RV64I-NEXT:    or a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: sh3adduw_3:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    sh3add.uw a0, a0, a1
+; RV64ZBA-NEXT:    ret
+  %3 = shl i64 %0, 3
+  %4 = and i64 %3, 34359738360
+  %5 = or disjoint i64 %4, %1
+  ret i64 %5
+}
+
 ; Type legalization inserts a sext_inreg after the first add. That add will be
 ; selected as sh2add which does not sign extend. SimplifyDemandedBits is unable
 ; to remove the sext_inreg because it has multiple uses. The ashr will use the
@@ -335,6 +389,24 @@ define i64 @addmul6(i64 %a, i64 %b) {
   ret i64 %d
 }
 
+define i64 @disjointormul6(i64 %a, i64 %b) {
+; RV64I-LABEL: disjointormul6:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    li a2, 6
+; RV64I-NEXT:    mul a0, a0, a2
+; RV64I-NEXT:    or a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: disjointormul6:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    sh1add a0, a0, a0
+; RV64ZBA-NEXT:    sh1add a0, a0, a1
+; RV64ZBA-NEXT:    ret
+  %c = mul i64 %a, 6
+  %d = or disjoint i64 %c, %b
+  ret i64 %d
+}
+
 define i64 @addmul10(i64 %a, i64 %b) {
 ; RV64I-LABEL: addmul10:
 ; RV64I:       # %bb.0:
@@ -1099,6 +1171,23 @@ define i64 @add4104(i64 %a) {
   ret i64 %c
 }
 
+define i64 @add4104_2(i64 %a) {
+; RV64I-LABEL: add4104_2:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    lui a1, 1
+; RV64I-NEXT:    addiw a1, a1, 8
+; RV64I-NEXT:    or a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: add4104_2:
+; RV64ZBA:       # %bb.0:
+; RV64ZBA-NEXT:    li a1, 1026
+; RV64ZBA-NEXT:    sh2add a0, a1, a0
+; RV64ZBA-NEXT:    ret
+  %c = or disjoint i64 %a, 4104
+  ret i64 %c
+}
+
 define i64 @add8208(i64 %a) {
 ; RV64I-LABEL: add8208:
 ; RV64I:       # %bb.0: