tree d3485a074858c1440aee09f1bcb8e34fe56c1cdf
parent afb4ebda771f27ff05ae51cdee65269431d662d7
author Antonio Maiorano <amaiorano@google.com> 1575664473 -0500
committer Antonio Maiorano <amaiorano@google.com> 1575923267 +0000

Subzero: proper fix for assert(Dest->hasReg())

Reverted the hack fix from bb222a10622, and implemented a proper fix.

The fix in TargetX86Base<TraitsType>::lowerCast
(IceTargetLoweringX86BaseImpl.h) was suggested by Jim Stichnoth. The
original problem is that we're trying movd an i32 to an i4i32, and since
the target Variable is long-lived, it may not get a register allocated
for it, and when that happens, we end up tripping the assert in
InstX86Movd::emitIAS() that expects the destination to have been
allocated a register. The solution that Jim suggested, and is
implemented here, is to create a temporary, short-lived, variable to
first movd into, which should guarantee a register target, since
short-lived Variables usually get registers. Then we 'mov' the temporary
register Variable to Dest, which *should* support moving the i4i32
register operand to an i32 memory operand.

I said *should* above, because with the above fix, we now trip another
assert in InstX86Mov::emitIAS (Mov, not Movd). The reason being that it
doesn't actually support moving an i4i32 reg -> i32 memory operand. I
added support for this as well (IceInstX86BaseImpl.h). Note that this
assert only tripped when building with Om1 optimization level, since O2
ostensibly optimized out the mov call.

Bug: b/145529686
Change-Id: I3c998a3e308838123cb415fcbf9f277113ac7d28
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/39068
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Antonio Maiorano <amaiorano@google.com>