https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83707
Will Schmidt <willschm at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |willschm at gcc dot gnu.org
--- Comment #2 from Will Schmidt <willschm at gcc dot gnu.org> ---
The troublesome snippet of code reads as so
( from gcc/testsuite/g++.dg/eh/simd-3.C )
"
int main(void)
{
v v1 = vt;
try {
thrower();
} catch (int x) {
}
v2 = v1;
if (memcmp(&v2,&vt,...)) abort();
"
where thrower() is a function meant to scramble contents of the VSRs.
And the problem (that we abort during the test run) occurs when any non-zero
optimization level is specified, due to vt != v2.
high-level debug suggests that we are not storing the v1=vt result to memory,
(only to a reg), so the call to thrower clobbers the contents, and we
ultimately fail.
When built with -O0, on the "v1 = vt" statement, i can see (.expand dump)
;; Generating RTL for gimple basic block 2
;; v1_5 ={v} vt;
(insn # # # (set (reg:SI 129)
(high:SI (symbol_ref:SI ("vt") [flags 0x2] <var_decl # vt>)))
"/home/willschm/gcc/gcc-mainline-regtest_patches/gcc/testsuite/g++.dg/eh/simd-3.C":56#
(nil))
(insn # # # (set (reg/f:SI 128)
(lo_sum:SI (reg:SI 129)
(symbol_ref:SI ("vt") [flags 0x2] <var_decl # vt>)))
"/home/willschm/gcc/gcc-mainline-regtest_patches/gcc/testsuite/g++.dg/eh/simd-3.C":56#
(expr_list:REG_EQUAL (symbol_ref:SI ("vt") [flags 0x2] <var_decl # vt>)
(nil)))
(insn # # # (set (reg:V4SI 130)
(mem/v/c:V4SI (reg/f:SI 128) [1 vt+0 S16 A128]))
"/home/willschm/gcc/gcc-mainline-regtest_patches/gcc/testsuite/g++.dg/eh/simd-3.C":56#
(nil))
(insn # # 0 (set (mem/c:V4SI (plus:SI (reg/f:SI 116 virtual-stack-vars)
(const_int 8 [0x8])) [1 v1+0 S16 A128])
(reg:V4SI 130))
"/home/willschm/gcc/gcc-mainline-regtest_patches/gcc/testsuite/g++.dg/eh/simd-3.C":56#
(nil))
;; thrower ();
Noting the "set (mem/c..." statement. This code works, in that the v1 value
survives the call to thrower.
When built with -O1, that chunk of the dump now reads:
;; v1_4 ={v} vt;
(insn # # # (set (reg:SI 128)
(high:SI (symbol_ref:SI ("*.LANCHOR1") [flags 0x182])))
"/home/willschm/gcc/gcc-mainline-regtest_patches/gcc/testsuite/g++.dg/eh/simd-3.C":56#
(nil))
(insn # # # (set (reg/f:SI 127)
(lo_sum:SI (reg:SI 128)
(symbol_ref:SI ("*.LANCHOR1") [flags 0x182])))
"/home/willschm/gcc/gcc-mainline-regtest_patches/gcc/testsuite/g++.dg/eh/simd-3.C":56#
(expr_list:REG_EQUAL (symbol_ref:SI ("*.LANCHOR1") [flags 0x182])
(nil)))
(insn # # 0 (set (reg/v:V4SI 123 [ v1 ])
(mem/v/c:V4SI (reg/f:SI 127) [1 vt+0 S16 A128]))
"/home/willschm/gcc/gcc-mainline-regtest_patches/gcc/testsuite/g++.dg/eh/simd-3.C":56#
(nil))
;; thrower ();
And here, note that this is a "set (reg". There is no "set (mem" until after
the "v2 = v1" statement, after the call to thrower().
This seems to fit with what I see via objdump :
(working code)
xx <main>:
xx: 94 21 ff c0 stwu r1,-64(r1)
xx: 7c 08 02 a6 mflr r0
xx: 90 01 00 44 stw r0,68(r1)
xx: 93 e1 00 3c stw r31,60(r1)
xx: 7c 3f 0b 78 mr r31,r1
xx: 3d 20 10 02 lis r9,4098
xx: 39 29 00 40 addi r9,r9,64
xx: 7c 00 4e 18 lxvw4x vs0,0,r9 <- load.
xx: 39 20 00 10 li r9,16
xx: 7c 1f 4f 18 stxvw4x vs0,r31,r9 <- store.
xx: 4b ff ff 61 bl 1000074c <_Z7throwerv>
versus (failing code):
xx <main>:
xx: 94 21 ff d0 stwu r1,-48(r1)
xx: 7c 08 02 a6 mflr r0
xx: 90 01 00 34 stw r0,52(r1)
xx: 38 00 00 10 li r0,16
xx: 7f e1 01 ce stvx v31,r1,r0
xx: 3d 20 10 02 lis r9,4098
xx: 39 29 00 40 addi r9,r9,64
xx: 7f e0 4e 19 lxvw4x vs63,0,r9 <- load, no store.
xx: 4b ff ff 89 bl 100006f0 <_Z7throwerv>
most previous tree dump before expand (232t.optimized) shows bb2 having:
<bb 2> :
v1_5 ={v} vt;
thrower ();
versus
<bb 2> [local count: 1073741825]:
v1_4 ={v} vt;
thrower ();
