Global ISel 64 bit left-shift+add results in 32 bit adds

[tpopp]

I'm seeing instructions like

%i12 = sext i32 %i11 to i64
%i62 = shl i64 %i12, 1
%i64 = add i64 0, %i62
%i65 = add i64 %i64, 1

resulting in a sequence of 32 bit adds with global isel, while without results in a 64 bit shl+add. I assume this is worse performance. Is there a reason to prefer this or expect the same performance?

I'm using llc -O3 -march=amdgcn -mcpu=gfx942 -mtriple amdgcn-amd-hmcsa ./reduced.ll -global-isel={true,false} -o - in looking at this.

reduced.gisel.txt
reduced.sdisel.txt
reduced.txt

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Tres (tpopp)

I'm seeing instructions like

%i12 = sext i32 %i11 to i64
%i62 = shl i64 %i12, 1
%i64 = add i64 0, %i62
%i65 = add i64 %i64, 1

resulting in a sequence of 32 bit adds with global isel, while without results in a 64 bit shl+add. I assume this is worse performance. Is there a reason to prefer this or expect the same performance?

I'm using llc -O3 -march=amdgcn -mcpu=gfx942 -mtriple amdgcn-amd-hmcsa ./reduced.ll -global-isel={true,false} -o - in looking at this.

reduced.gisel.txt
reduced.sdisel.txt
reduced.txt

[tpopp]

I've added #122709 as a parent to link to a related issue, using that as a parent, as it seems more general.

[lialan]

Relevant issue #122688 : #122688 (comment)

[tpopp]

The issue I've found here is that, at least gfx942 supports these 64 bit adds, but the target is configured to only expect later generations to support 64 bit adds/subs. It does however see that the 64 bit subs might not be supported on gfx942. I haven't checked the hardware manual for supported operations.

[lialan]

According to @tpopp , this is a higher priority codegen issue.

[tpopp]

@arsenm I'm unclear on any details around these pseudo operations. Is it just this simple, or more likely, am I missing something? #124763

[arsenm]

@arsenm I'm unclear on any details around these pseudo operations. Is it just this simple, or more likely, am I missing something? #124763

i64 add shouldn't really be legal. It was useful in the DAG to pretend they were, in order to match addressing modes. We shouldn't need that in globalisel. We only do the manual selection for G_PTR_ADDs which failed to fold into an addressing mode.

For hasLshlAddB64 you can select add using v_lshl_add_u64, in which case we should make add legal