On 05/02/2018 04:05 PM, Jim Wilson wrote:
> This improves the code for a switch statement on targets that sign-extend
> function arguments, such as RISC-V. Given a simple testcase
>
> extern void asdf(int);
> void foo(int x) {
> switch (x) {
> case 0: asdf(10); break;
> case 1: asdf(11); break;
> case 2: asdf(12); break;
> case 3: asdf(13); break;
> case 4: asdf(14); break;
> }
> }
>
> Compiled for a 64-bit target, we get for the tablejump
>
> li a5,4
> bgtu a0,a5,.L1
> slli a0,a0,32
> lui a5,%hi(.L4)
> addi a5,a5,%lo(.L4)
> srli a0,a0,30
> add a0,a0,a5
> lw a5,0(a0)
> jr a5
>
> There is some unnecessary shifting here. a0 (x) gets shifted left by 32 then
> shifted right by 30 to zero-extend it and multiply by 4 for the table index.
> However, after the unsigned greater than branch, we know the value is between
> 0 and 4. We also know that a 32-bit int is passed as a 64-bit sign-extended
> long for this target. Thus we get the same exact value if we sign-extend
> instead of zero-extend, and the code is one instruction shorter. We get a
> slli
> by 2 instead of the slli 32/srli 30.
>
> The following patch implements this optimization. It checks for a range that
> does not have the sign-bit set, and an index value that is already sign
> extended, and then does a sign extend instead of an zero extend.
>
> This has been tested with a riscv{32,64}-{elf,linux} builds and testsuite
> runs.
> There were no regressions. It was also tested with an x86_64-linux build and
> testsuite run.
>
> Ok?
>
> Jim
>
> gcc/
> * expr.c (do_tablejump): When converting index to Pmode, if we have a
> sign extended promoted subreg, and the range does not have the sign bit
> set, then do a sign extend.
OK.
Jeff
