Allow RISC-V ABI to be configured by mabi

[zyedidia]

This PR allows the RISC-V ABI to be configured by the user. RISC-V has the following ABIs: lp64, lp64f, lp64d (64-bit only), and ilp32, ilp32f, ilp32d (32-bit only). See here for a slightly more detailed description, and here for the GCC documentation on this.

The latest LDC release changed the default 64-bit ABI from lp64 to lp64d. I think the default should continue to be lp64 because it is the most basic, and additional features/extensions should be opt-in (this PR changes back to lp64 default). The latest release also enabled the D extension by default. This PR undoes that and leaves the D extension as opt-in (with -mattr=+d), just like every other extension.

The ABI is now configurable using -mabi=... instead of hard-coding to lp64d. The mabi option is also no longer hidden. The default ABI on riscv64 is now lp64 and on riscv32 is ilp32. Also, the backend linker now gets passed the appropriate mabi flags and march flags. Instead of hard-coding to mabi=lp64d and march=rv64gc (only for 64-bit though), now the value passed via mabi is used (or the default ABI), and march is properly constructed from the enabled features (passed via mattr). For example, to compile for rv64gc (shorthand for rv64imacfdc_zicsr_zifencei) with lp64d one would now pass:

-mtriple=riscv64-unknown-elf -mattr=+m,+a,+c,+f,+d,+c -mabi=lp64d

The backend linker/gcc driver would be passed -march=rv64gc -mabi=lp64d. Here are some other examples:

-mtriple=riscv32-unknown-elf -mattr=+m,+a,+c -mabi=ilp32
gcc/linker flags: -march=rv32imac_zicsr_zifencei -mabi=ilp32

-mtriple=riscv64-unknown-elf -mattr=+m,+a,+c,+f -mabi=lp64f
gcc/linker flags: -march=rv64imafc_zicsr_zifencei -mabi=lp64f

-mtriple=riscv64-unknown-elf
gcc/linker flags: -march=rv64i_zicsr_zifencei -mabi=lp64

Since Phobos probably needs to be built with rv64gc and lp64d the flags used to invoke the build should be

-mtriple=riscv64-unknown-elf -mattr=+m,+a,+c,+f,+d,+c -mabi=lp64d

I'm not sure if code in other repos needs to be changed as a result of this.

Fixes #4321.

[JohanEngelen]

LGTM

[JohanEngelen]

It's OK like this, but you could merge all lit tests into one file, and then use --check-prefix to have different CHECKs per FileCheck invocation.

[andreas-schwab]

This probably breaks building druntime.

core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs0, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs1, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs2, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs3, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs4, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs5, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs6, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs7, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs8, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs9, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs10, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs11, 0(a2)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs0, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs1, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs2, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs3, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs4, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs5, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs6, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs7, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs8, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs9, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs10, 0(a0)
^
core/thread/osthread.d(1594):1:2: error: instruction requires the following: 'D' (Double-Precision Floating-Point)
fsd fs11, 0(a0)
^
make[2]: *** [runtime/CMakeFiles/druntime-ldc-shared.dir/build.make:346: runtime/objects-shared/core/atomic.o] Error 1

[zyedidia]

I think the builder should just be updated to compile with -mattr=+m,+a,+c,+f,+d,+c -mabi=lp64d.

[andreas-schwab]

When configured for riscv64 the default should always be rv64gc.

[zyedidia]

I think that would be reasonable if there were a way to specify the exact RISC-V extensions to use using a flag like -march. This is how Clang operates: the default (if no flags are passed) is rv64imac for bare-metal targets (riscv64-unknown-elf), and rv64gc for hosted targets (riscv64-linux-elf), but -march can be used to override the default and specify rv64i, rv64gc, or something else.

Currently with LDC, if the default were changed to be rv64gc, then to use rv64im, the user would have to pass -mattr=-a,-c,-f,-d,-c, which is un-ergonomic and strange because it relies on the default and specifies the extensions to remove from the default instead of the ones to include from the base. If we make -march work like Clang, then I would be in favor of changing the default to rv64gc or rv64imac.