https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113696
--- Comment #1 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Pan Li <pa...@gcc.gnu.org>: https://gcc.gnu.org/g:6082024891e421bfd81d8f708a7411f4019d0d97 commit r14-9060-g6082024891e421bfd81d8f708a7411f4019d0d97 Author: Juzhe-Zhong <juzhe.zh...@rivai.ai> Date: Thu Feb 1 17:02:52 2024 +0800 RISC-V: Suppress the vsetvl fusion for conflict successors Update in v2: Add dump information. This patch fixes the following ineffective vsetvl insertion: void f (int32_t * restrict in, int32_t * restrict out, size_t n, size_t cond, size_t cond2) { for (size_t i = 0; i < n; i++) { if (i == cond) { vint8mf8_t v = *(vint8mf8_t*)(in + i + 100); *(vint8mf8_t*)(out + i + 100) = v; } else if (i == cond2) { vfloat32mf2_t v = *(vfloat32mf2_t*)(in + i + 200); *(vfloat32mf2_t*)(out + i + 200) = v; } else if (i == (cond2 - 1)) { vuint16mf2_t v = *(vuint16mf2_t*)(in + i + 300); *(vuint16mf2_t*)(out + i + 300) = v; } else { vint8mf4_t v = *(vint8mf4_t*)(in + i + 400); *(vint8mf4_t*)(out + i + 400) = v; } } } Before this patch: f: .LFB0: .cfi_startproc beq a2,zero,.L12 addi a7,a0,400 addi a6,a1,400 addi a0,a0,1600 addi a1,a1,1600 li a5,0 addi t6,a4,-1 vsetvli t3,zero,e8,mf8,ta,ma ---> ineffective uplift .L7: beq a3,a5,.L15 beq a4,a5,.L16 beq t6,a5,.L17 vsetvli t1,zero,e8,mf4,ta,ma vle8.v v1,0(a0) vse8.v v1,0(a1) vsetvli t3,zero,e8,mf8,ta,ma .L4: addi a5,a5,1 addi a7,a7,4 addi a6,a6,4 addi a0,a0,4 addi a1,a1,4 bne a2,a5,.L7 .L12: ret .L15: vle8.v v1,0(a7) vse8.v v1,0(a6) j .L4 .L17: vsetvli t1,zero,e8,mf4,ta,ma addi t5,a0,-400 addi t4,a1,-400 vle16.v v1,0(t5) vse16.v v1,0(t4) vsetvli t3,zero,e8,mf8,ta,ma j .L4 .L16: addi t5,a0,-800 addi t4,a1,-800 vle32.v v1,0(t5) vse32.v v1,0(t4) j .L4 It's obvious that we are hoisting the e8mf8 vsetvl to the top. It's ineffective since e8mf8 comes from low probability block which is if (i == cond). For this case, we disable such fusion. After this patch: f: beq a2,zero,.L12 addi a7,a0,400 addi a6,a1,400 addi a0,a0,1600 addi a1,a1,1600 li a5,0 addi t6,a4,-1 .L7: beq a3,a5,.L15 beq a4,a5,.L16 beq t6,a5,.L17 vsetvli t1,zero,e8,mf4,ta,ma vle8.v v1,0(a0) vse8.v v1,0(a1) .L4: addi a5,a5,1 addi a7,a7,4 addi a6,a6,4 addi a0,a0,4 addi a1,a1,4 bne a2,a5,.L7 .L12: ret .L15: vsetvli t3,zero,e8,mf8,ta,ma vle8.v v1,0(a7) vse8.v v1,0(a6) j .L4 .L17: addi t5,a0,-400 addi t4,a1,-400 vsetvli t1,zero,e8,mf4,ta,ma vle16.v v1,0(t5) vse16.v v1,0(t4) j .L4 .L16: addi t5,a0,-800 addi t4,a1,-800 vsetvli t3,zero,e32,mf2,ta,ma vle32.v v1,0(t5) vse32.v v1,0(t4) j .L4 Tested on both RV32/RV64 no regression. Ok for trunk ? PR target/113696 gcc/ChangeLog: * config/riscv/riscv-vsetvl.cc (pre_vsetvl::earliest_fuse_vsetvl_info): Suppress vsetvl fusion. gcc/testsuite/ChangeLog: * gcc.target/riscv/rvv/vsetvl/pr113696.c: New test.
