[RISCV] Add isel special case for (and (shl X, c2), c1) -> (slli_uw (srli x, c4-c2), c4).

[topperc]

Where c1 is a shifted mask with 32 set bits and c4 trailing zeros.

This is an alternative to #91626.

[llvmbot]

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

Author: Craig Topper (topperc)

Changes

Where c1 is a shifted mask with 32 set bits and c4 trailing zeros.

This is an alternative to #91626.

---

Full diff: https://github.com/llvm/llvm-project/pull/91638.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp (+17-2)
• (modified) llvm/test/CodeGen/RISCV/rv64zba.ll (+23-4)

diff --git a/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp b/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
index e73a3af92af6f..6fd16210aade9 100644
--- a/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
@@ -1322,11 +1322,11 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
         }
       }
 
-      // Turn (and (shl x, c2), c1) -> (srli (slli c2+c3), c3) if c1 is a mask
-      // shifted by c2 bits with c3 leading zeros.
       if (LeftShift && isShiftedMask_64(C1)) {
         unsigned Leading = XLen - llvm::bit_width(C1);
 
+        // Turn (and (shl x, c2), c1) -> (srli (slli c2+c3), c3) if c1 is a mask
+        // shifted by c2 bits with c3 leading zeros.
         if (C2 + Leading < XLen &&
             C1 == (maskTrailingOnes<uint64_t>(XLen - (C2 + Leading)) << C2)) {
           // Use slli.uw when possible.
@@ -1350,6 +1350,21 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
             return;
           }
         }
+
+        // Turn (and (shl x, c2), c1) -> (slli_uw (srli x, c4-c2), c4) where c1
+        // is shifted mask with 32 set bits and c4 trailing zeros.
+        unsigned Trailing = llvm::countr_zero(C1);
+        if (Leading + Trailing == 32 && C2 < Trailing &&
+            Subtarget->hasStdExtZba() && OneUseOrZExtW) {
+          SDNode *SRLI = CurDAG->getMachineNode(
+              RISCV::SRLI, DL, VT, X,
+              CurDAG->getTargetConstant(Trailing - C2, DL, VT));
+          SDNode *SLLI_UW = CurDAG->getMachineNode(
+              RISCV::SLLI_UW, DL, VT, SDValue(SRLI, 0),
+              CurDAG->getTargetConstant(Trailing, DL, VT));
+          ReplaceNode(Node, SLLI_UW);
+          return;
+        }
       }
 
       // Turn (and (shr x, c2), c1) -> (slli (srli x, c2+c3), c3) if c1 is a
diff --git a/llvm/test/CodeGen/RISCV/rv64zba.ll b/llvm/test/CodeGen/RISCV/rv64zba.ll
index 8fe221f2a297a..867775452e0c0 100644
--- a/llvm/test/CodeGen/RISCV/rv64zba.ll
+++ b/llvm/test/CodeGen/RISCV/rv64zba.ll
@@ -2866,8 +2866,7 @@ define ptr @gep_lshr_i32(ptr %0, i64 %1) {
 ;
 ; RV64ZBA-LABEL: gep_lshr_i32:
 ; RV64ZBA:       # %bb.0: # %entry
-; RV64ZBA-NEXT:    slli a1, a1, 2
-; RV64ZBA-NEXT:    srli a1, a1, 4
+; RV64ZBA-NEXT:    srli a1, a1, 2
 ; RV64ZBA-NEXT:    slli.uw a1, a1, 4
 ; RV64ZBA-NEXT:    sh2add a1, a1, a1
 ; RV64ZBA-NEXT:    add a0, a0, a1
@@ -2891,8 +2890,7 @@ define i64 @srli_slliw(i64 %1) {
 ;
 ; RV64ZBA-LABEL: srli_slliw:
 ; RV64ZBA:       # %bb.0: # %entry
-; RV64ZBA-NEXT:    slli a0, a0, 2
-; RV64ZBA-NEXT:    srli a0, a0, 4
+; RV64ZBA-NEXT:    srli a0, a0, 2
 ; RV64ZBA-NEXT:    slli.uw a0, a0, 4
 ; RV64ZBA-NEXT:    ret
 entry:
@@ -2902,6 +2900,27 @@ entry:
   ret i64 %4
 }
 
+define i64 @srli_slliw_canonical(i64 %0) {
+; RV64I-LABEL: srli_slliw_canonical:
+; RV64I:       # %bb.0: # %entry
+; RV64I-NEXT:    slli a0, a0, 2
+; RV64I-NEXT:    li a1, 1
+; RV64I-NEXT:    slli a1, a1, 36
+; RV64I-NEXT:    addi a1, a1, -16
+; RV64I-NEXT:    and a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBA-LABEL: srli_slliw_canonical:
+; RV64ZBA:       # %bb.0: # %entry
+; RV64ZBA-NEXT:    srli a0, a0, 2
+; RV64ZBA-NEXT:    slli.uw a0, a0, 4
+; RV64ZBA-NEXT:    ret
+entry:
+  %1 = shl i64 %0, 2
+  %2 = and i64 %1, 68719476720
+  ret i64 %2
+}
+
 define i64 @srli_slli_i16(i64 %1) {
 ; CHECK-LABEL: srli_slli_i16:
 ; CHECK:       # %bb.0: # %entry

[preames]

LGTM w/minor comment

[preames]

Do we need a !IsCANDI check here?

[dtcxzyw]

Emm, #91626 seems to perform better than this patch. BTW this patch converts some slli insts into slli.uw, which makes them less compressible.

[topperc]

Emm, #91626 seems to perform better than this patch. BTW this patch converts some slli insts into slli.uw, which makes them less compressible.

Can you get reproducers?

[dtcxzyw]

Emm, #91626 seems to perform better than this patch. BTW this patch converts some slli insts into slli.uw, which makes them less compressible.

Can you get reproducers?

godbolt: https://godbolt.org/z/sMv9fzWbY

; llc -mtriple=riscv64 -mattr=+c,+m,+zba -o -
define i64 @func0000000000000000(i64 %0) #0 {
entry:
  %1 = mul i64 %0, 100
  %2 = udiv i64 %1, 70
  %3 = shl i64 %2, 32
  ret i64 %3
}

Before:

func0000000000000000:                   # @func0000000000000000
        lui     a1, %hi(.LCPI0_0)
        ld      a1, %lo(.LCPI0_0)(a1)
        li      a2, 100
        mul     a0, a0, a2
        mulhu   a0, a0, a1
        srli    a0, a0, 6
        slli    a0, a0, 32
        ret

After:

func0000000000000000:                   # @func0000000000000000
	lui	a1, %hi(.LCPI1_0)
	ld	a1, %lo(.LCPI1_0)(a1)
	li	a2, 100
 	mul	a0, a0, a2
 	mulhu	a0, a0, a1
 	srli	a0, a0, 6
 	slli.uw	a0, a0, 32 *!!!*
 	ret

[topperc]

Need a check that Leading is non-zero or Trailing is less than 32

[dtcxzyw]

LGTM.

[topperc]

LGTM.

Are the results better now? Was that the only issue?

I'll add a test for that case before I commit.

[dtcxzyw]

LGTM.

Are the results better now? Was that the only issue?

Yeah, I confirmed that all regressions have been fixed :)
See dtcxzyw/llvm-codegen-benchmark@5c1d002