| Message ID | 20230321141206.751719-3-quic_mliebel@quicinc.com |
|---|---|
| State | New |
| Headers | show |
| Series | Update hexagon toolchain | expand |
> -----Original Message----- > From: Marco Liebel (QUIC) <quic_mliebel@quicinc.com> > Sent: Tuesday, March 21, 2023 8:12 AM > To: qemu-devel@nongnu.org > Cc: Taylor Simpson <tsimpson@quicinc.com>; Matheus Bernardino (QUIC) > <quic_mathbern@quicinc.com>; Marco Liebel (QUIC) > <quic_mliebel@quicinc.com> > Subject: [PATCH 2/2] Add test for storing .new vector > > Hexagon toolchain version 16.0.0 fixes a bug where the ecoding of storing a > .new vector was incorrect. This resulted in an incorrect valued being stored. > The test checks that the correct value is used. > > Signed-off-by: Marco Liebel <quic_mliebel@quicinc.com> > --- > tests/tcg/hexagon/hvx_misc.c | 29 +++++++++++++++++++++++++++++ > 1 file changed, 29 insertions(+) > > diff --git a/tests/tcg/hexagon/hvx_misc.c b/tests/tcg/hexagon/hvx_misc.c > index 53d5c9b44f..657e556dd4 100644 > --- a/tests/tcg/hexagon/hvx_misc.c > +++ b/tests/tcg/hexagon/hvx_misc.c > @@ -211,6 +211,34 @@ static void test_store_unaligned(void) > check_output_w(__LINE__, 2); > } > > +static void test_store_new(void) > +{ > + asm volatile( > + "r0 = #0x00000003\n\t" > + "v0 = vsplat(r0)\n\t" > + "r0 = #expect\n\t" > + "vmem(r0+#0) = v0\n\t" The idiom used in this file is the inline asm stores into the output buffer and there is vanilla C that writes to the expect buffer. So, move the above to something like this (after the inline asm). for (int j = 0; I < MAX_VEC_SIZE_BYTES / 4; j++) { expect[0].w[j] = 3; } Thanks, Taylor
On Tue, 21 Mar 2023 at 14:13, Marco Liebel <quic_mliebel@quicinc.com> wrote: > > Hexagon toolchain version 16.0.0 fixes a bug where the ecoding of > storing a .new vector was incorrect. This resulted in an incorrect > valued being stored. The test checks that the correct value is used. So is this a compiler/assembler bug? Do we need to have tests relating to those in QEMU's test suite ? thanks -- PMM
> -----Original Message----- > From: Peter Maydell <peter.maydell@linaro.org> > Sent: Dienstag, 21. März 2023 18:20 > To: Marco Liebel (QUIC) <quic_mliebel@quicinc.com> > Cc: qemu-devel@nongnu.org; Taylor Simpson <tsimpson@quicinc.com>; > Matheus Bernardino (QUIC) <quic_mathbern@quicinc.com> > Subject: Re: [PATCH 2/2] Add test for storing .new vector > > WARNING: This email originated from outside of Qualcomm. Please be wary > of any links or attachments, and do not enable macros. > > On Tue, 21 Mar 2023 at 14:13, Marco Liebel <quic_mliebel@quicinc.com> > wrote: > > > > Hexagon toolchain version 16.0.0 fixes a bug where the ecoding of > > storing a .new vector was incorrect. This resulted in an incorrect > > valued being stored. The test checks that the correct value is used. > > So is this a compiler/assembler bug? Do we need to have tests > relating to those in QEMU's test suite ? > > thanks > -- PMM The bug was in the assembler. For the instruction that does the store of the .new vector (vmem(r0+#0) = v3.new) it created the wrong output. So there should be no need to have more tests, other than the one provided by this patch. Marco
diff --git a/tests/tcg/hexagon/hvx_misc.c b/tests/tcg/hexagon/hvx_misc.c index 53d5c9b44f..657e556dd4 100644 --- a/tests/tcg/hexagon/hvx_misc.c +++ b/tests/tcg/hexagon/hvx_misc.c @@ -211,6 +211,34 @@ static void test_store_unaligned(void) check_output_w(__LINE__, 2); } +static void test_store_new(void) +{ + asm volatile( + "r0 = #0x00000003\n\t" + "v0 = vsplat(r0)\n\t" + "r0 = #expect\n\t" + "vmem(r0+#0) = v0\n\t" + + "r0 = #output\n\t" + "r1 = #0x00000001\n\t" + "r2 = #0x00000002\n\t" + "r3 = #0x00000004\n\t" + + "v1 = vsplat(r1)\n\t" + "v2 = vsplat(r2)\n\t" + "v3 = vsplat(r3)\n\t" + + "{" + " v3.w,q0 = vadd(v1.w, v2.w):carry\n\t" + " vmem(r0+#0) = v3.new\n\t" + "}" + + ::: "r0", "r1", "r2", "r3", "v0", "v1", "v2", "v3", "q0", "memory" + ); + + check_output_w(__LINE__, 1); +} + static void test_masked_store(bool invert) { void *p0 = buffer0; @@ -620,6 +648,7 @@ int main() test_load_unaligned(); test_store_aligned(); test_store_unaligned(); + test_store_new(); test_masked_store(false); test_masked_store(true); test_new_value_store();
Hexagon toolchain version 16.0.0 fixes a bug where the ecoding of storing a .new vector was incorrect. This resulted in an incorrect valued being stored. The test checks that the correct value is used. Signed-off-by: Marco Liebel <quic_mliebel@quicinc.com> --- tests/tcg/hexagon/hvx_misc.c | 29 +++++++++++++++++++++++++++++ 1 file changed, 29 insertions(+)