Hello!
While eyeballing following testcase:
float a[256], b[256], c[256];
void foo(void)
{
int i;
for (i=0; i<256; ++i)
c[i] = a[i] / b[i];
}
-O2 -ftree-vectorize -ffast-math
I noticed that for some reason CSE doesn't eliminate memory read, resulting in:
.L2:
vrcpps b(%rax), %ymm0
vmulps b(%rax), %ymm0, %ymm1
vmulps %ymm1, %ymm0, %ymm1
vaddps %ymm0, %ymm0, %ymm0
vsubps %ymm1, %ymm0, %ymm1
vmulps a(%rax), %ymm1, %ymm1
vmovaps %ymm1, c(%rax)
addq $32, %rax
cmpq $1024, %rax
jne .L2
Attached patch forces memory operand into register, producing:
.L2:
vmovaps b(%rax), %ymm1
vrcpps %ymm1, %ymm0
vmulps %ymm1, %ymm0, %ymm1
vmulps %ymm1, %ymm0, %ymm1
vaddps %ymm0, %ymm0, %ymm0
vsubps %ymm1, %ymm0, %ymm1
vmulps a(%rax), %ymm1, %ymm1
vmovaps %ymm1, c(%rax)
addq $32, %rax
cmpq $1024, %rax
jne .L2
The same cure could be applied for rsqrt sequences.
2011-10-21 Uros Bizjak <ubiz...@gmail.com>
* config/i386/i386.c (ix86_emit_swdivsf): Force b into register.
(ix86_emit_swsqrtsf): Force a into register.
Patch was tested on x86_64-pc-linux-gnu, committed to mainline SVN.
Uros.
Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c (revision 180255)
+++ config/i386/i386.c (working copy)
@@ -33682,6 +33682,8 @@ void ix86_emit_swdivsf (rtx res, rtx a, rtx b, enu
/* a / b = a * ((rcp(b) + rcp(b)) - (b * rcp(b) * rcp (b))) */
+ b = force_reg (mode, b);
+
/* x0 = rcp(b) estimate */
emit_insn (gen_rtx_SET (VOIDmode, x0,
gen_rtx_UNSPEC (mode, gen_rtvec (1, b),
@@ -33737,6 +33739,8 @@ void ix86_emit_swsqrtsf (rtx res, rtx a, enum mach
/* sqrt(a) = -0.5 * a * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) - 3.0)
rsqrt(a) = -0.5 * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) - 3.0) */
+ a = force_reg (mode, a);
+
/* x0 = rsqrt(a) estimate */
emit_insn (gen_rtx_SET (VOIDmode, x0,
gen_rtx_UNSPEC (mode, gen_rtvec (1, a),
