Message ID | 37a03341f16da30b83bec1f4ef51dce4e6f25264.1693915537.git.research_trasio@irq.a4lg.com |
---|---|
State | New |
Headers | show |
Series | [v2] RISC-V: Fix Zicond ICE on large constants | expand |
On 9/5/23 06:08, Tsukasa OI wrote: > From: Tsukasa OI <research_trasio@irq.a4lg.com> > > Large constant cons and/or alt will trigger ICEs building GCC target > libraries (libgomp and libatomic) when the 'Zicond' extension is enabled. > > For instance, zicond-ice-2.c (new test case in this commit) will cause > an ICE when SOME_NUMBER is 0x1000 or larger. While opposite numbers > corresponding cons/alt (two temp2 variables) are checked, cons/alt > themselves are not checked and causing 2 ICEs building > GCC target libraries as of this writing: > > 1. gcc/libatomic/config/posix/lock.c > 2. gcc/libgomp/fortran.c > > Coercing a large value into a register will fix the issue. > > It also coerce a large cons into a register on "imm, imm" case (the author > could not reproduce but possible to cause an ICE). > > gcc/ChangeLog: > > * config/riscv/riscv.cc (riscv_expand_conditional_move): Force > large constant cons/alt into a register. > > gcc/testsuite/ChangeLog: > > * gcc.target/riscv/zicond-ice-2.c: New test. This is based on > an ICE at libat_lock_n func on gcc/libatomic/config/posix/lock.c > but heavily minimized. "New test." is sufficient. No need to change it, just a note going forward. OK. Thanks for taking care of this. jeff
On 2023/09/06 10:22, Jeff Law wrote: > > > On 9/5/23 06:08, Tsukasa OI wrote: >> From: Tsukasa OI <research_trasio@irq.a4lg.com> >> >> Large constant cons and/or alt will trigger ICEs building GCC target >> libraries (libgomp and libatomic) when the 'Zicond' extension is enabled. >> >> For instance, zicond-ice-2.c (new test case in this commit) will cause >> an ICE when SOME_NUMBER is 0x1000 or larger. While opposite numbers >> corresponding cons/alt (two temp2 variables) are checked, cons/alt >> themselves are not checked and causing 2 ICEs building >> GCC target libraries as of this writing: >> >> 1. gcc/libatomic/config/posix/lock.c >> 2. gcc/libgomp/fortran.c >> >> Coercing a large value into a register will fix the issue. >> >> It also coerce a large cons into a register on "imm, imm" case (the >> author >> could not reproduce but possible to cause an ICE). >> >> gcc/ChangeLog: >> >> * config/riscv/riscv.cc (riscv_expand_conditional_move): Force >> large constant cons/alt into a register. >> >> gcc/testsuite/ChangeLog: >> >> * gcc.target/riscv/zicond-ice-2.c: New test. This is based on >> an ICE at libat_lock_n func on gcc/libatomic/config/posix/lock.c >> but heavily minimized. > "New test." is sufficient. No need to change it, just a note going > forward. > > OK. Thanks for taking care of this. > > jeff > Thanks for reviewing. I'll remember that for the next time (I am adding brief description in Binutils but not *that* long compared to this). Committing. Tsukasa
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc index 8d8f7b4f16ed..e306d57814be 100644 --- a/gcc/config/riscv/riscv.cc +++ b/gcc/config/riscv/riscv.cc @@ -3917,6 +3917,11 @@ riscv_expand_conditional_move (rtx dest, rtx op, rtx cons, rtx alt) gen_rtx_IF_THEN_ELSE (mode, cond, CONST0_RTX (mode), alt))); + /* CONS might not fit into a signed 12 bit immediate suitable + for an addi instruction. If that's the case, force it + into a register. */ + if (!SMALL_OPERAND (INTVAL (cons))) + cons = force_reg (mode, cons); riscv_emit_binary (PLUS, dest, dest, cons); return true; } @@ -3940,11 +3945,13 @@ riscv_expand_conditional_move (rtx dest, rtx op, rtx cons, rtx alt) rtx temp1 = gen_reg_rtx (mode); rtx temp2 = gen_int_mode (-1 * INTVAL (cons), mode); - /* TEMP2 might not fit into a signed 12 bit immediate suitable - for an addi instruction. If that's the case, force it into - a register. */ + /* TEMP2 and/or CONS might not fit into a signed 12 bit immediate + suitable for an addi instruction. If that's the case, force it + into a register. */ if (!SMALL_OPERAND (INTVAL (temp2))) temp2 = force_reg (mode, temp2); + if (!SMALL_OPERAND (INTVAL (cons))) + cons = force_reg (mode, cons); riscv_emit_binary (PLUS, temp1, alt, temp2); emit_insn (gen_rtx_SET (dest, @@ -3986,11 +3993,13 @@ riscv_expand_conditional_move (rtx dest, rtx op, rtx cons, rtx alt) rtx temp1 = gen_reg_rtx (mode); rtx temp2 = gen_int_mode (-1 * INTVAL (alt), mode); - /* TEMP2 might not fit into a signed 12 bit immediate suitable - for an addi instruction. If that's the case, force it into - a register. */ + /* TEMP2 and/or ALT might not fit into a signed 12 bit immediate + suitable for an addi instruction. If that's the case, force it + into a register. */ if (!SMALL_OPERAND (INTVAL (temp2))) temp2 = force_reg (mode, temp2); + if (!SMALL_OPERAND (INTVAL (alt))) + alt = force_reg (mode, alt); riscv_emit_binary (PLUS, temp1, cons, temp2); emit_insn (gen_rtx_SET (dest, diff --git a/gcc/testsuite/gcc.target/riscv/zicond-ice-2.c b/gcc/testsuite/gcc.target/riscv/zicond-ice-2.c new file mode 100644 index 000000000000..ffd8dcb5814e --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/zicond-ice-2.c @@ -0,0 +1,11 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gc_zicond -mabi=lp64d" { target { rv64 } } } */ +/* { dg-options "-march=rv32gc_zicond -mabi=ilp32d" { target { rv32 } } } */ + +#define SOME_NUMBER 0x1000 + +unsigned long +d (unsigned long n) +{ + return n > SOME_NUMBER ? SOME_NUMBER : n; +}
From: Tsukasa OI <research_trasio@irq.a4lg.com> Large constant cons and/or alt will trigger ICEs building GCC target libraries (libgomp and libatomic) when the 'Zicond' extension is enabled. For instance, zicond-ice-2.c (new test case in this commit) will cause an ICE when SOME_NUMBER is 0x1000 or larger. While opposite numbers corresponding cons/alt (two temp2 variables) are checked, cons/alt themselves are not checked and causing 2 ICEs building GCC target libraries as of this writing: 1. gcc/libatomic/config/posix/lock.c 2. gcc/libgomp/fortran.c Coercing a large value into a register will fix the issue. It also coerce a large cons into a register on "imm, imm" case (the author could not reproduce but possible to cause an ICE). gcc/ChangeLog: * config/riscv/riscv.cc (riscv_expand_conditional_move): Force large constant cons/alt into a register. gcc/testsuite/ChangeLog: * gcc.target/riscv/zicond-ice-2.c: New test. This is based on an ICE at libat_lock_n func on gcc/libatomic/config/posix/lock.c but heavily minimized. --- gcc/config/riscv/riscv.cc | 21 +++++++++++++------ gcc/testsuite/gcc.target/riscv/zicond-ice-2.c | 11 ++++++++++ 2 files changed, 26 insertions(+), 6 deletions(-) create mode 100644 gcc/testsuite/gcc.target/riscv/zicond-ice-2.c base-commit: 72b639760a891c406725854bfb08132c83f0761a