| Message ID | 56A1FBDF.6020402@foss.arm.com |
|---|---|
| State | New |
| Headers | show |
On Fri, Jan 22, 2016 at 9:52 AM, Kyrill Tkachov <kyrylo.tkachov@foss.arm.com> wrote: > Hi all, > > As part of investigating the codegen effects of a fix for PR 65932 I found > we assign > too high a cost for the sign-extending multiply instruction SMULBB. > This is because we add the cost of a multiply-extend but then also recurse > into the > SIGN_EXTEND sub-expressions rather than the registers (or subregs) being > sign-extended. > > This patch is a simple fix. The fix is right by itself, but in combination > with patch 3 > fix the gcc.target/arm/wmul-2.c testcase. > > Bootstrapped and tested on arm-none-linux-gnueabihf. > > Ok for trunk? > OK. Thanks, Ramana > Thanks, > Kyrill > > 2016-01-22 Kyrylo Tkachov <kyrylo.tkachov@arm.com> > > * config/arm/arm.c (arm_new_rtx_costs, MULT case): Properly extract > the operands of the SIGN_EXTENDs from a SMUL[TB][TB] rtx.
On 04/02/16 09:00, Ramana Radhakrishnan wrote: > On Fri, Jan 22, 2016 at 9:52 AM, Kyrill Tkachov > <kyrylo.tkachov@foss.arm.com> wrote: >> Hi all, >> >> As part of investigating the codegen effects of a fix for PR 65932 I found >> we assign >> too high a cost for the sign-extending multiply instruction SMULBB. >> This is because we add the cost of a multiply-extend but then also recurse >> into the >> SIGN_EXTEND sub-expressions rather than the registers (or subregs) being >> sign-extended. >> >> This patch is a simple fix. The fix is right by itself, but in combination >> with patch 3 >> fix the gcc.target/arm/wmul-2.c testcase. >> >> Bootstrapped and tested on arm-none-linux-gnueabihf. >> >> Ok for trunk? >> > OK. Is it ok to backport this to the GCC 5 branch? It fixes a testcase with cortex-a5 tuning and was tested by Christophe: https://gcc.gnu.org/ml/gcc-patches/2016-02/msg01238.html Thanks, Kyrill > Thanks, > Ramana >> Thanks, >> Kyrill >> >> 2016-01-22 Kyrylo Tkachov <kyrylo.tkachov@arm.com> >> >> * config/arm/arm.c (arm_new_rtx_costs, MULT case): Properly extract >> the operands of the SIGN_EXTENDs from a SMUL[TB][TB] rtx.
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c index 17f00b5a1de21de35366b82040a7ad46d65f899e..ee3ebe4561ea3d9791fabdfcec8b16af63bd4d20 100644 --- a/gcc/config/arm/arm.c +++ b/gcc/config/arm/arm.c @@ -10321,8 +10321,10 @@ arm_new_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer_code, /* SMUL[TB][TB]. */ if (speed_p) *cost += extra_cost->mult[0].extend; - *cost += rtx_cost (XEXP (x, 0), mode, SIGN_EXTEND, 0, speed_p); - *cost += rtx_cost (XEXP (x, 1), mode, SIGN_EXTEND, 1, speed_p); + *cost += rtx_cost (XEXP (XEXP (x, 0), 0), mode, + SIGN_EXTEND, 0, speed_p); + *cost += rtx_cost (XEXP (XEXP (x, 1), 0), mode, + SIGN_EXTEND, 1, speed_p); return true; } if (speed_p)
Hi all, As part of investigating the codegen effects of a fix for PR 65932 I found we assign too high a cost for the sign-extending multiply instruction SMULBB. This is because we add the cost of a multiply-extend but then also recurse into the SIGN_EXTEND sub-expressions rather than the registers (or subregs) being sign-extended. This patch is a simple fix. The fix is right by itself, but in combination with patch 3 fix the gcc.target/arm/wmul-2.c testcase. Bootstrapped and tested on arm-none-linux-gnueabihf. Ok for trunk? Thanks, Kyrill 2016-01-22 Kyrylo Tkachov <kyrylo.tkachov@arm.com> * config/arm/arm.c (arm_new_rtx_costs, MULT case): Properly extract the operands of the SIGN_EXTENDs from a SMUL[TB][TB] rtx.