The attached patch fixes a bug in Sema::ActOnGCCAsmStmt where it wasn't computing the corresponding ConstraintIdx from an operand number correctly when there were constraints with the '+' modifier.
For example, when the following inline-asm statement is compiled,
void g2(int one, int two, int three) {
asm volatile ("%0 %1 %2 %3 %4" : "+r" (one), "+r"(two) : "r"(three));
}
the existing code maps the operand number in the assembly template to
ConstraintIdx in the following way:
%0 => "+r" (one) => ConstraintIdx = 0
%1 => "+r" (one) => ConstraintIdx = 0
%2 => "+r" (two) => ConstraintIdx = 1
%3 => "+r" (two) => ConstraintIdx = 1
%4 => "+r" (three) => ConstraintIdx = 2
The correct mapping should be like this:
%0 => "+r" (one) => ConstraintIdx = 0
%1 => "+r" (two) => ConstraintIdx = 1
%3 => "+r" (three) => ConstraintIdx = 2
%4 => "+r" (one) => ConstraintIdx = 0
%5 => "+r" (two) => ConstraintIdx = 1
I couldn't find the rule documented anywhere in the link below, but judging
from the code I see in CodeGenFunction::EmitAsmStmt and the IR clang
generates, I believe this is the correct way to find the corresponding
constraint.
https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html
This is the IR generated:
%1 = tail call { i32, i32 } asm sideeffect "$0 $1 $2 $3 $4",
"=r,=r,r,0,1"(i32 %three, i32 %one, i32 %two)
inline-asm-read-write-constraint1.patch
Description: Binary data
_______________________________________________ cfe-commits mailing list [email protected] http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
