[PATCH 4/5] shrink-wrap: Shrink-wrapping for separate components

[Segher Boessenkool]

On Tue, Sep 27, 2016 at 03:14:51PM -0600, Jeff Law wrote:
> With transposition issue addressed, the only blocker I see are some 
> simple testcases we can add to the suite.  They don't have to be real 
> extensive.  And one motivating example for the list archives, ideally 
> the glibc malloc case.

Here are some testcases.


2016-09-30  Segher Boessenkool  <segher@kernel.crashing.org>

gcc/testsuite/
	* gcc.target/powerpc/shrink-wrap-separate-0.c: New.
	* gcc.target/powerpc/shrink-wrap-separate-1.c: New.
	* gcc.target/powerpc/shrink-wrap-separate-2.c: New.


diff --git a/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-0.c b/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-0.c
new file mode 100644
index 0000000..dea0611
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-0.c
@@ -0,0 +1,22 @@
+/* { dg-do compile { target powerpc*-*-* } } */
+/* { dg-options "-O2" } */
+/* { dg-final { scan-assembler {#before\M.*\mmflr\M} } } */
+
+/* This tests if shrink-wrapping for separate components works.
+
+   r20 (a callee-saved register) is forced live at the start, so that we
+   get it saved in a prologue at the start of the function.
+   The link register only needs to be saved if x is non-zero; without
+   separate shrink-wrapping it would however be saved in the one prologue.
+   The test tests if the mflr insn ends up behind the prologue.  */
+
+void g(void);
+
+void f(int x)
+{
+	register int r20 asm("20") = x;
+	asm("#before" : : "r"(r20));
+	if (x)
+		g();
+	asm(""); // no tailcall of g
+}
diff --git a/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-1.c b/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-1.c
new file mode 100644
index 0000000..735b606
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-1.c
@@ -0,0 +1,18 @@
+/* { dg-do compile { target powerpc*-*-* } } */
+/* { dg-options "-O2" } */
+/* { dg-final { scan-assembler {\mmflr\M.*\mbl\M.*\mmflr\M.*\mbl\M} } } */
+
+/* This tests if shrink-wrapping for separate components creates more
+   than one prologue when that is useful.  In this case, it saves the
+   link register before both the call to g and the call to h.  */
+
+void g(void) __attribute__((noreturn));
+void h(void) __attribute__((noreturn));
+
+void f(int x)
+{
+	if (x == 42)
+		g();
+	if (x == 31)
+		h();
+}
diff --git a/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-2.c b/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-2.c
new file mode 100644
index 0000000..b22564a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/shrink-wrap-separate-2.c
@@ -0,0 +1,26 @@
+/* { dg-do compile { target powerpc*-*-* } } */
+/* { dg-options "-O2" } */
+/* { dg-final { scan-assembler {\mmflr\M.*\mbl\M.*\mmflr\M.*\mbl\M} } } */
+
+/* This tests if shrink-wrapping for separate components puts a prologue
+   inside a loop when that is useful.  In this case, it saves the link
+   register before each call: both calls happen with probability .10,
+   so saving the link register happens with .80 per execution of f on
+   average, which is smaller than 1 which you would get if you saved
+   it outside the loop.  */
+
+int *a;
+void g(void);
+
+void f(int x)
+{
+	int j;
+	for (j = 0; j < 4; j++) {
+		if (__builtin_expect(a[j], 0))
+			g();
+		asm("#" : : : "memory");
+		if (__builtin_expect(a[j], 0))
+			g();
+		a[j]++;
+	}
+}