[x86 PATCH] PR target/94680: Clear upper bits of V2DF using movq (like V2DI).

["Roger Sayle" <roger@nextmovesoftware.com>]

[-- Attachment #1: Type: text/plain, Size: 1587 bytes --]


This simple i386 patch unblocks a more significant change.  The testcase
gcc.target/i386/sse2-pr94680.c isn't quite testing what's intended, and
alas the fix for PR target/94680 doesn't (yet) handle V2DF mode.

For the first test from sse2-pr94680.c, below

v2df foo_v2df (v2df x) {
  return __builtin_shuffle (x, (v2df) { 0, 0 }, (v2di) { 0, 2 });
}

GCC on x86_64-pc-linux-gnu with -O2 currently generates:

        movhpd  .LC0(%rip), %xmm0
        ret
.LC0:
        .long   0
        .long   0

which passes the test as it contains a mov insn and no xor.
Alas reading a zero from the constant pool isn't quite the
desired implementation.  With this patch we now generate:

        movq    %xmm0, %xmm0
        ret

The same code as we generate for V2DI, and add a stricter
test case.  My first attempt tried using VI8F_128 to generalize
the existing sse2_movq128 define_insn to both V2DI and V2DF.
Alas, CODE_FOR_sse2_movq128 is exposed as a builtin in
i386-builtin.def, requiring some internal name changes, that
ultimately the testsuite was unhappy with.  The simpler solution
(that works) is to clone/specialize a new V2DF *sse2_movq128_2.

This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check with no new failures.  Ok for mainline?


2022-03-15  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
	PR target/94680
	* config/i386/sse.md (*sse2_movq128_2): A version of sse2_movq128
	for V2DF mode.

gcc/testsuite/ChangeLog
	PR target/94680
	* gcc.target/i386/sse2-pr94680-2.c: New stricter V2DF test case.


Thanks in advance,
Roger
--


[-- Attachment #2: patchmq2.txt --]
[-- Type: text/plain, Size: 1512 bytes --]

diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index e9292e6..d017fb8 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -1599,6 +1599,19 @@
    (set_attr "prefix" "maybe_vex")
    (set_attr "mode" "TI")])
 
+(define_insn "*sse2_movq128_2"
+  [(set (match_operand:V2DF 0 "register_operand" "=v")
+	(vec_concat:V2DF
+	  (vec_select:DF
+	    (match_operand:V2DF 1 "nonimmediate_operand" "vm")
+	    (parallel [(const_int 0)]))
+	  (match_operand:DF 2 "const0_operand")))]
+  "TARGET_SSE2"
+  "%vmovq\t{%1, %0|%0, %q1}"
+  [(set_attr "type" "ssemov")
+   (set_attr "prefix" "maybe_vex")
+   (set_attr "mode" "TI")])
+
 ;; Move a DI from a 32-bit register pair (e.g. %edx:%eax) to an xmm.
 ;; We'd rather avoid this entirely; if the 32-bit reg pair was loaded
 ;; from memory, we'd prefer to load the memory directly into the %xmm
diff --git a/gcc/testsuite/gcc.target/i386/sse2-pr94680-2.c b/gcc/testsuite/gcc.target/i386/sse2-pr94680-2.c
new file mode 100644
index 0000000..abd260a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/sse2-pr94680-2.c
@@ -0,0 +1,13 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -msse2" } */
+typedef double v2df __attribute__ ((vector_size (16)));
+typedef long long v2di __attribute__((vector_size(16)));
+
+v2df foo_v2df (v2df x)
+{
+  return __builtin_shuffle (x, (v2df) { 0, 0 }, (v2di) { 0, 2 });
+}
+
+/* { dg-final { scan-assembler "movq" } } */
+/* { dg-final { scan-assembler-not "pxor" } } */
+