Since except for Bonnell, 01 fb add %edi,%ebx
is faster and shorter than 8d 1c 1f lea (%rdi,%rbx,1),%ebx we should use add for a = a + b and a = b + a when possible if not optimizing for Bonnell. Tested on x86-64. gcc/ PR target/92807 * config/i386/i386.c (ix86_lea_outperforms): Check !TARGET_BONNELL. (ix86_avoid_lea_for_addr): When not optimizing for Bonnell, use add for a = a + b and a = b + a. gcc/testsuite/ PR target/92807 * gcc.target/i386/pr92807-1.c: New test. -- H.J.
From ad803a967a6c18ae3bd6f8381ebc8a78c31a82ae Mon Sep 17 00:00:00 2001 From: "H.J. Lu" <hjl.to...@gmail.com> Date: Tue, 3 Dec 2019 15:27:51 -0800 Subject: [PATCH] i386: Use add for a = a + b and a = b + a when possible Since except for Bonnell, 01 fb add %edi,%ebx is faster and shorter than 8d 1c 1f lea (%rdi,%rbx,1),%ebx we should use add for a = a + b and a = b + a when possible if not optimizing for Bonnell. Tested on x86-64. gcc/ PR target/92807 * config/i386/i386.c (ix86_lea_outperforms): Check !TARGET_BONNELL. (ix86_avoid_lea_for_addr): When not optimizing for Bonnell, use add for a = a + b and a = b + a. gcc/testsuite/ PR target/92807 * gcc.target/i386/pr92807-1.c: New test. --- gcc/config/i386/i386.c | 27 +++++++++++++++-------- gcc/testsuite/gcc.target/i386/pr92807-1.c | 11 +++++++++ 2 files changed, 29 insertions(+), 9 deletions(-) create mode 100644 gcc/testsuite/gcc.target/i386/pr92807-1.c diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c index 04cbbd532c0d..65f0d44916a8 100644 --- a/gcc/config/i386/i386.c +++ b/gcc/config/i386/i386.c @@ -14393,11 +14393,10 @@ ix86_lea_outperforms (rtx_insn *insn, unsigned int regno0, unsigned int regno1, { int dist_define, dist_use; - /* For Silvermont if using a 2-source or 3-source LEA for - non-destructive destination purposes, or due to wanting - ability to use SCALE, the use of LEA is justified. */ - if (TARGET_SILVERMONT || TARGET_GOLDMONT || TARGET_GOLDMONT_PLUS - || TARGET_TREMONT || TARGET_INTEL) + /* For Atom processors newer than Bonnell, if using a 2-source or + 3-source LEA for non-destructive destination purposes, or due to + wanting ability to use SCALE, the use of LEA is justified. */ + if (!TARGET_BONNELL) { if (has_scale) return true; @@ -14532,10 +14531,6 @@ ix86_avoid_lea_for_addr (rtx_insn *insn, rtx operands[]) struct ix86_address parts; int ok; - /* Check we need to optimize. */ - if (!TARGET_AVOID_LEA_FOR_ADDR || optimize_function_for_size_p (cfun)) - return false; - /* The "at least two components" test below might not catch simple move or zero extension insns if parts.base is non-NULL and parts.disp is const0_rtx as the only components in the address, e.g. if the @@ -14572,6 +14567,20 @@ ix86_avoid_lea_for_addr (rtx_insn *insn, rtx operands[]) if (parts.index) regno2 = true_regnum (parts.index); + /* Use add for a = a + b and a = b + a since it is faster and shorter + than lea for most processors. For the processors like BONNELL, if + the destination register of LEA holds an actual address which will + be used soon, LEA is better and otherwise ADD is better. */ + if (!TARGET_BONNELL + && parts.scale == 1 + && (!parts.disp || parts.disp == const0_rtx) + && (regno0 == regno1 || regno0 == regno2)) + return true; + + /* Check we need to optimize. */ + if (!TARGET_AVOID_LEA_FOR_ADDR || optimize_function_for_size_p (cfun)) + return false; + split_cost = 0; /* Compute how many cycles we will add to execution time diff --git a/gcc/testsuite/gcc.target/i386/pr92807-1.c b/gcc/testsuite/gcc.target/i386/pr92807-1.c new file mode 100644 index 000000000000..00f92930af92 --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/pr92807-1.c @@ -0,0 +1,11 @@ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ + +unsigned int +abs2 (unsigned int a) +{ + unsigned int s = ((a>>15)&0x10001)*0xffff; + return (a+s)^s; +} + +/* { dg-final { scan-assembler-not "leal" } } */ -- 2.21.0