[llvm] e0289ad - [RISCV][P-ext] Use brev8 for bitreverse with Zbkb (#207338)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Jul 3 13:11:43 PDT 2026
Author: Hongyu Chen
Date: 2026-07-04T04:11:39+08:00
New Revision: e0289ad97908460c5dadbc4aefb3ad22cd8ac054
URL: https://github.com/llvm/llvm-project/commit/e0289ad97908460c5dadbc4aefb3ad22cd8ac054
DIFF: https://github.com/llvm/llvm-project/commit/e0289ad97908460c5dadbc4aefb3ad22cd8ac054.diff
LOG: [RISCV][P-ext] Use brev8 for bitreverse with Zbkb (#207338)
Resolves the FIXME.
Added:
llvm/test/CodeGen/RISCV/rvp-bitreverse-zbkb.ll
Modified:
llvm/lib/Target/RISCV/RISCVInstrInfoP.td
Removed:
################################################################################
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
index 6acdcb8379730..9114dfbf50a55 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
@@ -2213,9 +2213,13 @@ let append Predicates = [IsRV32] in {
// (PWMULSU_H GPR:$rs1, GPR:$rs2)>;
// 8/16-bit bitreverse patterns
- // FIXME: Use BREV8 with Zbkb.
+ // With Zbkb, brev8 reverses the bits within each byte directly; otherwise
+ // reverse all bits then swap the bytes back.
+ let append Predicates = [NoStdExtZbkb] in
def : Pat<(v4i8 (bitreverse GPR:$rs)),
(REV8_RV32 (REV_RV32 GPR:$rs))>;
+ let append Predicates = [HasStdExtZbkb] in
+ def : Pat<(v4i8 (bitreverse GPR:$rs)), (BREV8 GPR:$rs)>;
def : Pat<(v2i16 (bitreverse GPR:$rs)),
(PPAIROE_H (REV_RV32 GPR:$rs), (REV_RV32 GPR:$rs))>;
@@ -2434,9 +2438,15 @@ let append Predicates = [IsRV32] in {
// 8/16-bit bitreverse patterns
// FIXME: Custom lower?
+ // With Zbkb, brev8 reverses the bits within each byte directly.
+ let append Predicates = [NoStdExtZbkb] in
def : Pat<(v8i8 (bitreverse GPRPair:$rs)),
(BuildGPRPair (REV8_RV32 (REV_RV32 (EXTRACT_SUBREG GPRPair:$rs, sub_gpr_even))),
(REV8_RV32 (REV_RV32 (EXTRACT_SUBREG GPRPair:$rs, sub_gpr_odd))))>;
+ let append Predicates = [HasStdExtZbkb] in
+ def : Pat<(v8i8 (bitreverse GPRPair:$rs)),
+ (BuildGPRPair (BREV8 (EXTRACT_SUBREG GPRPair:$rs, sub_gpr_even)),
+ (BREV8 (EXTRACT_SUBREG GPRPair:$rs, sub_gpr_odd)))>;
def : Pat<(v4i16 (bitreverse GPRPair:$rs)),
(PPAIROE_DH (BuildGPRPair (REV_RV32 (EXTRACT_SUBREG GPRPair:$rs, sub_gpr_even)),
(REV_RV32 (EXTRACT_SUBREG GPRPair:$rs, sub_gpr_odd))),
@@ -2650,8 +2660,12 @@ let append Predicates = [IsRV64] in {
(PPAIROE_W (REV8_RV64 GPR:$rs), (REV8_RV64 GPR:$rs))>;
// 8/16/32-bit bitreverse patterns
+ // With Zbkb, brev8 reverses the bits within each byte directly.
+ let append Predicates = [NoStdExtZbkb] in
def : Pat<(v8i8 (bitreverse GPR:$rs)),
(REV8_RV64 (REV_RV64 GPR:$rs))>;
+ let append Predicates = [HasStdExtZbkb] in
+ def : Pat<(v8i8 (bitreverse GPR:$rs)), (BREV8 GPR:$rs)>;
def : Pat<(v4i16 (bitreverse GPR:$rs)),
(REV16_RV64 (REV_RV64 GPR:$rs))>;
def : Pat<(v2i32 (bitreverse GPR:$rs)),
diff --git a/llvm/test/CodeGen/RISCV/rvp-bitreverse-zbkb.ll b/llvm/test/CodeGen/RISCV/rvp-bitreverse-zbkb.ll
new file mode 100644
index 0000000000000..06c519c2818fd
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/rvp-bitreverse-zbkb.ll
@@ -0,0 +1,83 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -mattr=+experimental-p,+m,+zbb,+zbkb \
+; RUN: -verify-machineinstrs < %s | \
+; RUN: FileCheck %s --check-prefixes=CHECK,RV32
+; RUN: llc -mtriple=riscv64 -mattr=+experimental-p,+m,+zbb,+zbkb \
+; RUN: -verify-machineinstrs < %s | \
+; RUN: FileCheck %s --check-prefixes=CHECK,RV64
+
+; With Zbkb, a per-byte-lane bitreverse (v4i8/v8i8) selects a single brev8
+; instead of rev+rev8. Halfword/word lanes still need the rev-based sequence.
+
+define <4 x i8> @test_bitreverse_v4i8(<4 x i8> %a) {
+; CHECK-LABEL: test_bitreverse_v4i8:
+; CHECK: # %bb.0:
+; CHECK-NEXT: brev8 a0, a0
+; CHECK-NEXT: ret
+ %res = call <4 x i8> @llvm.bitreverse.v4i8(<4 x i8> %a)
+ ret <4 x i8> %res
+}
+
+define <2 x i16> @test_bitreverse_v2i16(<2 x i16> %a) {
+; RV32-LABEL: test_bitreverse_v2i16:
+; RV32: # %bb.0:
+; RV32-NEXT: rev a0, a0
+; RV32-NEXT: ppairoe.h a0, a0, a0
+; RV32-NEXT: ret
+;
+; RV64-LABEL: test_bitreverse_v2i16:
+; RV64: # %bb.0:
+; RV64-NEXT: rev a0, a0
+; RV64-NEXT: rev16 a0, a0
+; RV64-NEXT: ret
+ %res = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> %a)
+ ret <2 x i16> %res
+}
+
+define <8 x i8> @test_bitreverse_v8i8(<8 x i8> %a) {
+; RV32-LABEL: test_bitreverse_v8i8:
+; RV32: # %bb.0:
+; RV32-NEXT: brev8 a1, a1
+; RV32-NEXT: brev8 a0, a0
+; RV32-NEXT: ret
+;
+; RV64-LABEL: test_bitreverse_v8i8:
+; RV64: # %bb.0:
+; RV64-NEXT: brev8 a0, a0
+; RV64-NEXT: ret
+ %res = call <8 x i8> @llvm.bitreverse.v8i8(<8 x i8> %a)
+ ret <8 x i8> %res
+}
+
+define <4 x i16> @test_bitreverse_v4i16(<4 x i16> %a) {
+; RV32-LABEL: test_bitreverse_v4i16:
+; RV32: # %bb.0:
+; RV32-NEXT: rev a1, a1
+; RV32-NEXT: rev a0, a0
+; RV32-NEXT: ppairoe.dh a0, a0, a0
+; RV32-NEXT: ret
+;
+; RV64-LABEL: test_bitreverse_v4i16:
+; RV64: # %bb.0:
+; RV64-NEXT: rev a0, a0
+; RV64-NEXT: rev16 a0, a0
+; RV64-NEXT: ret
+ %res = call <4 x i16> @llvm.bitreverse.v4i16(<4 x i16> %a)
+ ret <4 x i16> %res
+}
+
+define <2 x i32> @test_bitreverse_v2i32(<2 x i32> %a) {
+; RV32-LABEL: test_bitreverse_v2i32:
+; RV32: # %bb.0:
+; RV32-NEXT: rev a1, a1
+; RV32-NEXT: rev a0, a0
+; RV32-NEXT: ret
+;
+; RV64-LABEL: test_bitreverse_v2i32:
+; RV64: # %bb.0:
+; RV64-NEXT: rev a0, a0
+; RV64-NEXT: ppairoe.w a0, a0, a0
+; RV64-NEXT: ret
+ %res = call <2 x i32> @llvm.bitreverse.v2i32(<2 x i32> %a)
+ ret <2 x i32> %res
+}
More information about the llvm-commits
mailing list