| Message ID | 20211020031709.359469-1-richard.henderson@linaro.org |
|---|---|
| Headers | show |
| Series | target/riscv: Rationalize XLEN and operand length | expand |
On Wed, Oct 20, 2021 at 1:26 PM Richard Henderson <richard.henderson@linaro.org> wrote: > > This is a partial patch set attempting to set things in the > right direction for both the UXL and RV128 patch sets. > > > r~ > > > Changes for v6: > * Rebase on riscv-to-apply.next. > > Changes for v5: > * Fix cpu_dump, which asserted for -accel qtest. > Instead of filtering CSRs explicitly in cpu_dump, > let the riscv_csr_operations predicate do the job. > This means we won't dump S-mode registers when RVS > is not enabled, much like we currently filter on RVH. > > Changes for v4: > * Use riscv_csrrw_debug for cpu_dump. > This fixes the issue that Alistair pointed out wrt the > MSTATUS.SD bit not being correct in the dump; note that > gdbstub already uses riscv_csrrw_debug, and so did not > have a problem. > * Align the registers in cpu_dump. > > Changes for v3: > * Fix CONFIG_ typo. > * Fix ctzw typo. > * Mark get_xlen unused (clang werror) > * Compute MSTATUS_SD on demand. > > Changes for v2: > * Set mxl/sxl/uxl at reset. > * Set sxl/uxl in write_mstatus. > > > Richard Henderson (15): > target/riscv: Move cpu_get_tb_cpu_state out of line > target/riscv: Create RISCVMXL enumeration > target/riscv: Split misa.mxl and misa.ext > target/riscv: Replace riscv_cpu_is_32bit with riscv_cpu_mxl > target/riscv: Add MXL/SXL/UXL to TB_FLAGS > target/riscv: Use REQUIRE_64BIT in amo_check64 > target/riscv: Properly check SEW in amo_op > target/riscv: Replace is_32bit with get_xl/get_xlen > target/riscv: Replace DisasContext.w with DisasContext.ol > target/riscv: Use gen_arith_per_ol for RVM > target/riscv: Adjust trans_rev8_32 for riscv64 > target/riscv: Use gen_unary_per_ol for RVB > target/riscv: Use gen_shift*_per_ol for RVB, RVI > target/riscv: Use riscv_csrrw_debug for cpu_dump > target/riscv: Compute mstatus.sd on demand Thanks! Applied to riscv-to-apply.next Alistair > > target/riscv/cpu.h | 73 +++------ > target/riscv/cpu_bits.h | 8 +- > hw/riscv/boot.c | 2 +- > linux-user/elfload.c | 2 +- > linux-user/riscv/cpu_loop.c | 2 +- > semihosting/arm-compat-semi.c | 2 +- > target/riscv/cpu.c | 195 +++++++++++++----------- > target/riscv/cpu_helper.c | 92 ++++++++++- > target/riscv/csr.c | 104 ++++++++----- > target/riscv/gdbstub.c | 10 +- > target/riscv/machine.c | 10 +- > target/riscv/monitor.c | 4 +- > target/riscv/translate.c | 174 +++++++++++++++------ > target/riscv/insn_trans/trans_rvb.c.inc | 140 +++++++++-------- > target/riscv/insn_trans/trans_rvi.c.inc | 44 +++--- > target/riscv/insn_trans/trans_rvm.c.inc | 36 ++++- > target/riscv/insn_trans/trans_rvv.c.inc | 29 ++-- > 17 files changed, 576 insertions(+), 351 deletions(-) > > -- > 2.25.1 > >
Hi Richard, I am implementing the UXL based on this patch set and have a question about how to process the PC register. As the specification said, "PC bits above XLEN are ignored, and when the PC is written, it is sign-extended to fill the widest supported XLEN." We still need special process of PC for exceptions or jump instructions. I have two methods to implement PC register access, but not sure which is the right way. First, normally process the PC register as the specification points. That means expanding the PC when setting the global TCGv variable cpu_pc, and truncating the pc_first and pc_next fields in DisasContextBase before decoding instructions. I am not sure whether the sign-extended pc is compatible with QEMU common code. Second, ignore ignore the PC special process for jump instructions. Just expand or truncate the PC register when exception processing, gdb read, or cpu dump registers. It is not a stright way, but I think it is still right. Hope to get you advice. Thanks very much. Best Regards, Zhiwei On 2021/10/20 上午11:16, Richard Henderson wrote: > This is a partial patch set attempting to set things in the > right direction for both the UXL and RV128 patch sets. > > > r~ > > > Changes for v6: > * Rebase on riscv-to-apply.next. > > Changes for v5: > * Fix cpu_dump, which asserted for -accel qtest. > Instead of filtering CSRs explicitly in cpu_dump, > let the riscv_csr_operations predicate do the job. > This means we won't dump S-mode registers when RVS > is not enabled, much like we currently filter on RVH. > > Changes for v4: > * Use riscv_csrrw_debug for cpu_dump. > This fixes the issue that Alistair pointed out wrt the > MSTATUS.SD bit not being correct in the dump; note that > gdbstub already uses riscv_csrrw_debug, and so did not > have a problem. > * Align the registers in cpu_dump. > > Changes for v3: > * Fix CONFIG_ typo. > * Fix ctzw typo. > * Mark get_xlen unused (clang werror) > * Compute MSTATUS_SD on demand. > > Changes for v2: > * Set mxl/sxl/uxl at reset. > * Set sxl/uxl in write_mstatus. > > > Richard Henderson (15): > target/riscv: Move cpu_get_tb_cpu_state out of line > target/riscv: Create RISCVMXL enumeration > target/riscv: Split misa.mxl and misa.ext > target/riscv: Replace riscv_cpu_is_32bit with riscv_cpu_mxl > target/riscv: Add MXL/SXL/UXL to TB_FLAGS > target/riscv: Use REQUIRE_64BIT in amo_check64 > target/riscv: Properly check SEW in amo_op > target/riscv: Replace is_32bit with get_xl/get_xlen > target/riscv: Replace DisasContext.w with DisasContext.ol > target/riscv: Use gen_arith_per_ol for RVM > target/riscv: Adjust trans_rev8_32 for riscv64 > target/riscv: Use gen_unary_per_ol for RVB > target/riscv: Use gen_shift*_per_ol for RVB, RVI > target/riscv: Use riscv_csrrw_debug for cpu_dump > target/riscv: Compute mstatus.sd on demand > > target/riscv/cpu.h | 73 +++------ > target/riscv/cpu_bits.h | 8 +- > hw/riscv/boot.c | 2 +- > linux-user/elfload.c | 2 +- > linux-user/riscv/cpu_loop.c | 2 +- > semihosting/arm-compat-semi.c | 2 +- > target/riscv/cpu.c | 195 +++++++++++++----------- > target/riscv/cpu_helper.c | 92 ++++++++++- > target/riscv/csr.c | 104 ++++++++----- > target/riscv/gdbstub.c | 10 +- > target/riscv/machine.c | 10 +- > target/riscv/monitor.c | 4 +- > target/riscv/translate.c | 174 +++++++++++++++------ > target/riscv/insn_trans/trans_rvb.c.inc | 140 +++++++++-------- > target/riscv/insn_trans/trans_rvi.c.inc | 44 +++--- > target/riscv/insn_trans/trans_rvm.c.inc | 36 ++++- > target/riscv/insn_trans/trans_rvv.c.inc | 29 ++-- > 17 files changed, 576 insertions(+), 351 deletions(-) >
On 10/22/21 1:26 AM, LIU Zhiwei wrote: > As the specification said, "PC bits above XLEN are ignored, and when > the PC is written, it is sign-extended to fill the widest supported XLEN." > We still need special process of PC for exceptions or jump instructions. > > I have two methods to implement PC register access, > but not sure which is the right way. > > First, normally process the PC register as the specification points. > That means expanding the PC when setting the global TCGv variable > cpu_pc, and truncating the pc_first and pc_next fields in > DisasContextBase before decoding instructions. I am not sure > whether the sign-extended pc is compatible with QEMU common code. I think extending on write is the correct thing. Jumps, exception entry and return, gdb write. Note that the read from PC for translation is in cpu_get_tb_cpu_state, before translation. You should not need to change anything wrt pc_first/pc_next/etc, because it will already have been done. > Second, ignore ignore the PC special process for jump instructions. > Just expand or truncate the PC register when exception processing, > gdb read, or cpu dump registers. It is not a stright way, but I think it is still right. I think you could make it work that way, but I don't know that it's less difficult, or that you'd have fewer changes. r~
On 2021/10/22 下午11:50, Richard Henderson wrote: > On 10/22/21 1:26 AM, LIU Zhiwei wrote: >> As the specification said, "PC bits above XLEN are ignored, and when >> the PC is written, it is sign-extended to fill the widest supported >> XLEN." >> We still need special process of PC for exceptions or jump instructions. >> >> I have two methods to implement PC register access, >> but not sure which is the right way. >> >> First, normally process the PC register as the specification points. >> That means expanding the PC when setting the global TCGv variable >> cpu_pc, and truncating the pc_first and pc_next fields in >> DisasContextBase before decoding instructions. I am not sure >> whether the sign-extended pc is compatible with QEMU common code. > > I think extending on write is the correct thing. Thanks. I will pick this way. > Jumps, exception entry and return, gdb write. If we carefully process jumps and gdb write, I think we can omit exception entry and return. Is it right? > > Note that the read from PC for translation is in cpu_get_tb_cpu_state, > before translation. You should not need to change anything wrt > pc_first/pc_next/etc, because it will already have been done. Good! Thanks. Best Regards, Zhiwei > >> Second, ignore ignore the PC special process for jump instructions. >> Just expand or truncate the PC register when exception processing, >> gdb read, or cpu dump registers. It is not a stright way, but I >> think it is still right. > > I think you could make it work that way, but I don't know that it's > less difficult, or that you'd have fewer changes. > > > r~
On 10/25/21 2:24 AM, LIU Zhiwei wrote: >> I think extending on write is the correct thing. > > Thanks. I will pick this way. > >> Jumps, exception entry and return, gdb write. > > If we carefully process jumps and gdb write, I think we can omit exception entry and > return. Is it right? No, though you'd probably have to create a special test case to see it. On exception return, a 64-bit OS can write a 64-bit value into SEPC. But when SEPC is copied to PC during SRET, UXL should be examined and the assignment should extend for a 32-bit user program. Exception entry would seem to be limited on its face by SXLEN; STVEC will always contain the same number of bits as PC, so no (additional) extension would be required. But we do have to be careful to interpret the target_ulong value properly for the current SXLEN. r~
This is a partial patch set attempting to set things in the right direction for both the UXL and RV128 patch sets. r~ Changes for v6: * Rebase on riscv-to-apply.next. Changes for v5: * Fix cpu_dump, which asserted for -accel qtest. Instead of filtering CSRs explicitly in cpu_dump, let the riscv_csr_operations predicate do the job. This means we won't dump S-mode registers when RVS is not enabled, much like we currently filter on RVH. Changes for v4: * Use riscv_csrrw_debug for cpu_dump. This fixes the issue that Alistair pointed out wrt the MSTATUS.SD bit not being correct in the dump; note that gdbstub already uses riscv_csrrw_debug, and so did not have a problem. * Align the registers in cpu_dump. Changes for v3: * Fix CONFIG_ typo. * Fix ctzw typo. * Mark get_xlen unused (clang werror) * Compute MSTATUS_SD on demand. Changes for v2: * Set mxl/sxl/uxl at reset. * Set sxl/uxl in write_mstatus. Richard Henderson (15): target/riscv: Move cpu_get_tb_cpu_state out of line target/riscv: Create RISCVMXL enumeration target/riscv: Split misa.mxl and misa.ext target/riscv: Replace riscv_cpu_is_32bit with riscv_cpu_mxl target/riscv: Add MXL/SXL/UXL to TB_FLAGS target/riscv: Use REQUIRE_64BIT in amo_check64 target/riscv: Properly check SEW in amo_op target/riscv: Replace is_32bit with get_xl/get_xlen target/riscv: Replace DisasContext.w with DisasContext.ol target/riscv: Use gen_arith_per_ol for RVM target/riscv: Adjust trans_rev8_32 for riscv64 target/riscv: Use gen_unary_per_ol for RVB target/riscv: Use gen_shift*_per_ol for RVB, RVI target/riscv: Use riscv_csrrw_debug for cpu_dump target/riscv: Compute mstatus.sd on demand target/riscv/cpu.h | 73 +++------ target/riscv/cpu_bits.h | 8 +- hw/riscv/boot.c | 2 +- linux-user/elfload.c | 2 +- linux-user/riscv/cpu_loop.c | 2 +- semihosting/arm-compat-semi.c | 2 +- target/riscv/cpu.c | 195 +++++++++++++----------- target/riscv/cpu_helper.c | 92 ++++++++++- target/riscv/csr.c | 104 ++++++++----- target/riscv/gdbstub.c | 10 +- target/riscv/machine.c | 10 +- target/riscv/monitor.c | 4 +- target/riscv/translate.c | 174 +++++++++++++++------ target/riscv/insn_trans/trans_rvb.c.inc | 140 +++++++++-------- target/riscv/insn_trans/trans_rvi.c.inc | 44 +++--- target/riscv/insn_trans/trans_rvm.c.inc | 36 ++++- target/riscv/insn_trans/trans_rvv.c.inc | 29 ++-- 17 files changed, 576 insertions(+), 351 deletions(-) -- 2.25.1