[PATCH] PR/34012, fwprop does not look at rtx_costs

[Paolo Bonzini]

Unlike CSE, fwprop has never looked at rtx_costs except for addressing 
mode selection.  This causes pessimizations, for example when the same 
expensive constant is used in many insns.  Then it is:

1) moved to a register by CSE

2) moved back as an immediate by fwprop

3) split, or reloaded, into a different register for every insn, causing 
worse code overall.

Example with -fno-forward-propagate:

         movabsq $578721382704613384, %rax
         movq    %rsp, %rdi
         movq    %rax, (%rsp)
         movq    %rax, 8(%rsp)
         movq    %rax, 16(%rsp)
         movq    %rax, 24(%rsp)
         movq    %rax, 32(%rsp)
         movq    %rax, 40(%rsp)
         movq    %rax, 48(%rsp)
         movq    %rax, 56(%rsp)
         movq    %rax, 64(%rsp)
         movq    %rax, 72(%rsp)

becomes with fwprop enabled:

         movabsq $578721382704613384, %rax
         movabsq $578721382704613384, %r11
         movq    %rax, (%rsp)
         movabsq $578721382704613384, %rax
         movq    %rsp, %rdi
         movq    %rax, 8(%rsp)
         movabsq $578721382704613384, %r10
         movabsq $578721382704613384, %r9
         movabsq $578721382704613384, %r8
         movabsq $578721382704613384, %rsi
         movabsq $578721382704613384, %rcx
         movabsq $578721382704613384, %rdx
         movabsq $578721382704613384, %rax
         movq    %r11, 16(%rsp)
         movq    %r10, 24(%rsp)
         movq    %r9, 32(%rsp)
         movq    %r8, 40(%rsp)
         movq    %rsi, 48(%rsp)
         movq    %rcx, 56(%rsp)
         movq    %rdx, 64(%rsp)
         movq    %rax, 72(%rsp)

The solution is to reject replacing a value if it makes the destination 
of the propagation more expensive.

Ok?

Paolo
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: pr34012.patch
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20071108/95ca787c/attachment.ksh>