Strength reduction should not remove some casts #3226
Conversation
Signed-off-by: Mihai Budiu <mbudiu@vmware.com>
|
Fixed #3225 |
frontends/p4/strengthReduction.cpp
Outdated
| if (expr->getH() < size_t(cast->expr->type->width_bits())) { | ||
| expr->e0 = cast->expr; | ||
| if (auto tb = cast->expr->type->to<IR::Type_Bits>()) { | ||
| auto type = expr->type->to<IR::Type_Bits>(); |
There was a problem hiding this comment.
Why not use checkedTo here?
frontends/p4/strengthReduction.cpp
Outdated
| if (auto tb = cast->expr->type->to<IR::Type_Bits>()) { | ||
| auto type = expr->type->to<IR::Type_Bits>(); | ||
| CHECK_NULL(type); | ||
| if (tb->isSigned != type->isSigned) |
There was a problem hiding this comment.
I always assumed that slices of signed bit strings will produce an unsigned bit string. This should probably ensure that a cast to int<W> is converted into a bit<W>. A cast from int<W> to bit<W> should be acceptable though?
This is what the spec says for slicing int<W>:
The result is an unsigned bit-string of width m - l + 1, including the bits numbered from l (which becomes the least significant bit of the result) to m (the most significant bit of the result) from the source operand.
There was a problem hiding this comment.
I added another test case which fails without this change; it's in the test file:
h.h.b = ((bit<1>)h.h.c)[0:0];
There was a problem hiding this comment.
But
((bit<1>)h.h.c)[0:0];
to
((bit<1>)h.h.c);
should be acceptable, right?
There was a problem hiding this comment.
The previous result was just h.h.c, which failed to typecheck.
|
The only thing is to to add a new member d to the header and assign c to that one. Otherwise, the assignment is optimized away. |
Signed-off-by: Mihai Budiu <mbudiu@vmware.com>
|
Is this ready to be merged? |
Signed-off-by: Mihai Budiu mbudiu@vmware.com